From bf20f7294ab75d69dd03383c054a079cb47fc022 Mon Sep 17 00:00:00 2001
From: zyddnys <zyddnys@outlook.com>
Date: Sun, 11 Oct 2020 17:14:39 -0400
Subject: [PATCH] first commit

---
 README.md                                     |   11 +
 RegDanbooru2019_8G.py                         |   33 +
 RegNetY-8.0GF_dds_8gpu.yaml                   |   27 +
 RegNetY_8G.py                                 |   67 +
 danbooru_labels.txt                           | 4096 +++++++++++++++++
 pycls/__init__.py                             |    0
 pycls/__pycache__/__init__.cpython-37.pyc     |  Bin 0 -> 132 bytes
 pycls/core/__init__.py                        |    0
 .../core/__pycache__/__init__.cpython-37.pyc  |  Bin 0 -> 137 bytes
 pycls/core/__pycache__/config.cpython-37.pyc  |  Bin 0 -> 4257 bytes
 pycls/core/__pycache__/losses.cpython-37.pyc  |  Bin 0 -> 755 bytes
 .../__pycache__/model_builder.cpython-37.pyc  |  Bin 0 -> 1317 bytes
 pycls/core/config.py                          |  410 ++
 pycls/core/losses.py                          |   28 +
 pycls/core/model_builder.py                   |   50 +
 pycls/core/optimizer.py                       |   79 +
 pycls/datasets/__init__.py                    |    0
 .../__pycache__/__init__.cpython-37.pyc       |  Bin 0 -> 141 bytes
 .../__pycache__/cifar10.cpython-37.pyc        |  Bin 0 -> 2902 bytes
 .../__pycache__/imagenet.cpython-37.pyc       |  Bin 0 -> 3347 bytes
 .../__pycache__/loader.cpython-37.pyc         |  Bin 0 -> 2056 bytes
 .../datasets/__pycache__/paths.cpython-37.pyc |  Bin 0 -> 815 bytes
 .../__pycache__/transforms.cpython-37.pyc     |  Bin 0 -> 3422 bytes
 pycls/datasets/cifar10.py                     |   83 +
 pycls/datasets/imagenet.py                    |  108 +
 pycls/datasets/loader.py                      |   80 +
 pycls/datasets/paths.py                       |   35 +
 pycls/datasets/transforms.py                  |  108 +
 pycls/models/__init__.py                      |    0
 .../__pycache__/__init__.cpython-37.pyc       |  Bin 0 -> 139 bytes
 .../models/__pycache__/anynet.cpython-37.pyc  |  Bin 0 -> 13481 bytes
 .../models/__pycache__/effnet.cpython-37.pyc  |  Bin 0 -> 7300 bytes
 .../models/__pycache__/regnet.cpython-37.pyc  |  Bin 0 -> 3809 bytes
 .../models/__pycache__/resnet.cpython-37.pyc  |  Bin 0 -> 9154 bytes
 pycls/models/anynet.py                        |  380 ++
 pycls/models/effnet.py                        |  235 +
 pycls/models/regnet.py                        |   86 +
 pycls/models/resnet.py                        |  275 ++
 pycls/utils/__init__.py                       |    0
 .../utils/__pycache__/__init__.cpython-37.pyc |  Bin 0 -> 138 bytes
 .../__pycache__/benchmark.cpython-37.pyc      |  Bin 0 -> 1877 bytes
 .../__pycache__/checkpoint.cpython-37.pyc     |  Bin 0 -> 2798 bytes
 .../__pycache__/distributed.cpython-37.pyc    |  Bin 0 -> 1883 bytes
 .../__pycache__/error_handler.cpython-37.pyc  |  Bin 0 -> 1945 bytes
 pycls/utils/__pycache__/io.cpython-37.pyc     |  Bin 0 -> 2286 bytes
 .../utils/__pycache__/logging.cpython-37.pyc  |  Bin 0 -> 3790 bytes
 pycls/utils/__pycache__/meters.cpython-37.pyc |  Bin 0 -> 7415 bytes
 .../utils/__pycache__/metrics.cpython-37.pyc  |  Bin 0 -> 3724 bytes
 .../multiprocessing.cpython-37.pyc            |  Bin 0 -> 1234 bytes
 pycls/utils/__pycache__/net.cpython-37.pyc    |  Bin 0 -> 3408 bytes
 pycls/utils/__pycache__/timer.cpython-37.pyc  |  Bin 0 -> 1031 bytes
 pycls/utils/benchmark.py                      |   89 +
 pycls/utils/checkpoint.py                     |   91 +
 pycls/utils/distributed.py                    |   61 +
 pycls/utils/error_handler.py                  |   59 +
 pycls/utils/io.py                             |   90 +
 pycls/utils/logging.py                        |  100 +
 pycls/utils/lr_policy.py                      |   47 +
 pycls/utils/meters.py                         |  239 +
 pycls/utils/metrics.py                        |  104 +
 pycls/utils/multiprocessing.py                |   57 +
 pycls/utils/net.py                            |   94 +
 pycls/utils/plotting.py                       |  132 +
 pycls/utils/timer.py                          |   35 +
 test.py                                       |   78 +
 65 files changed, 7467 insertions(+)
 create mode 100644 README.md
 create mode 100644 RegDanbooru2019_8G.py
 create mode 100644 RegNetY-8.0GF_dds_8gpu.yaml
 create mode 100644 RegNetY_8G.py
 create mode 100644 danbooru_labels.txt
 create mode 100644 pycls/__init__.py
 create mode 100644 pycls/__pycache__/__init__.cpython-37.pyc
 create mode 100644 pycls/core/__init__.py
 create mode 100644 pycls/core/__pycache__/__init__.cpython-37.pyc
 create mode 100644 pycls/core/__pycache__/config.cpython-37.pyc
 create mode 100644 pycls/core/__pycache__/losses.cpython-37.pyc
 create mode 100644 pycls/core/__pycache__/model_builder.cpython-37.pyc
 create mode 100644 pycls/core/config.py
 create mode 100644 pycls/core/losses.py
 create mode 100644 pycls/core/model_builder.py
 create mode 100644 pycls/core/optimizer.py
 create mode 100644 pycls/datasets/__init__.py
 create mode 100644 pycls/datasets/__pycache__/__init__.cpython-37.pyc
 create mode 100644 pycls/datasets/__pycache__/cifar10.cpython-37.pyc
 create mode 100644 pycls/datasets/__pycache__/imagenet.cpython-37.pyc
 create mode 100644 pycls/datasets/__pycache__/loader.cpython-37.pyc
 create mode 100644 pycls/datasets/__pycache__/paths.cpython-37.pyc
 create mode 100644 pycls/datasets/__pycache__/transforms.cpython-37.pyc
 create mode 100644 pycls/datasets/cifar10.py
 create mode 100644 pycls/datasets/imagenet.py
 create mode 100644 pycls/datasets/loader.py
 create mode 100644 pycls/datasets/paths.py
 create mode 100644 pycls/datasets/transforms.py
 create mode 100644 pycls/models/__init__.py
 create mode 100644 pycls/models/__pycache__/__init__.cpython-37.pyc
 create mode 100644 pycls/models/__pycache__/anynet.cpython-37.pyc
 create mode 100644 pycls/models/__pycache__/effnet.cpython-37.pyc
 create mode 100644 pycls/models/__pycache__/regnet.cpython-37.pyc
 create mode 100644 pycls/models/__pycache__/resnet.cpython-37.pyc
 create mode 100644 pycls/models/anynet.py
 create mode 100644 pycls/models/effnet.py
 create mode 100644 pycls/models/regnet.py
 create mode 100644 pycls/models/resnet.py
 create mode 100644 pycls/utils/__init__.py
 create mode 100644 pycls/utils/__pycache__/__init__.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/benchmark.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/checkpoint.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/distributed.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/error_handler.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/io.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/logging.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/meters.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/metrics.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/multiprocessing.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/net.cpython-37.pyc
 create mode 100644 pycls/utils/__pycache__/timer.cpython-37.pyc
 create mode 100644 pycls/utils/benchmark.py
 create mode 100644 pycls/utils/checkpoint.py
 create mode 100644 pycls/utils/distributed.py
 create mode 100644 pycls/utils/error_handler.py
 create mode 100644 pycls/utils/io.py
 create mode 100644 pycls/utils/logging.py
 create mode 100644 pycls/utils/lr_policy.py
 create mode 100644 pycls/utils/meters.py
 create mode 100644 pycls/utils/metrics.py
 create mode 100644 pycls/utils/multiprocessing.py
 create mode 100644 pycls/utils/net.py
 create mode 100644 pycls/utils/plotting.py
 create mode 100644 pycls/utils/timer.py
 create mode 100644 test.py

diff --git a/README.md b/README.md
new file mode 100644
index 0000000..195a08d
--- /dev/null
+++ b/README.md
@@ -0,0 +1,11 @@
+# Yet another Deep Danbooru project
+But based on [RegNetY-8G](https://arxiv.org/abs/2003.13678), relative lightweight, designed to run fast on GPU. \
+Training is done using mixed precision training on a single RTX2080Ti for 3 weeks. \
+Some code are from https://github.com/facebookresearch/pycls
+# What do I need?
+You need to download [save_4000000.ckpt]() from release and place on the same folder as `test.py`.
+# How to use?
+`python test.py --model save_4000000.ckpt --image <PATH_TO_IMAGE>`
+# What to do in the future?
+1. Quantize to 8 bit
+
diff --git a/RegDanbooru2019_8G.py b/RegDanbooru2019_8G.py
new file mode 100644
index 0000000..d73ac99
--- /dev/null
+++ b/RegDanbooru2019_8G.py
@@ -0,0 +1,33 @@
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from RegNetY_8G import build_model
+
+class RegDanbooru2019(nn.Module) :
+    def __init__(self) :
+        super(RegDanbooru2019, self).__init__()
+        self.backbone = build_model()
+        num_p = sum(p.numel() for p in self.backbone.parameters() if p.requires_grad)
+        print( 'Backbone has %d parameters' % num_p )
+        self.head_danbooru = nn.Linear(2016, 4096)
+
+    def forward_train_head(self, images) :
+        """
+        images of shape [N, 3, 512, 512]
+        """
+        with torch.no_grad() :
+            feats = self.backbone(images)
+            feats = F.adaptive_avg_pool2d(feats, 1).view(-1, 2016)
+        danbooru_logits = self.head_danbooru(feats) # [N, 4096]
+        return danbooru_logits
+
+    def forward(self, images) :
+        """
+        images of shape [N, 3, 512, 512]
+        """
+        feats = self.backbone(images)
+        feats = F.adaptive_avg_pool2d(feats, 1).view(-1, 2016)
+        danbooru_logits = self.head_danbooru(feats) # [N, 4096]
+        return danbooru_logits
diff --git a/RegNetY-8.0GF_dds_8gpu.yaml b/RegNetY-8.0GF_dds_8gpu.yaml
new file mode 100644
index 0000000..912754d
--- /dev/null
+++ b/RegNetY-8.0GF_dds_8gpu.yaml
@@ -0,0 +1,27 @@
+MODEL:
+  TYPE: regnet
+  NUM_CLASSES: 1000
+REGNET:
+  SE_ON: true
+  DEPTH: 17
+  W0: 192
+  WA: 76.82
+  WM: 2.19
+  GROUP_W: 56
+OPTIM:
+  LR_POLICY: cos
+  BASE_LR: 0.4
+  MAX_EPOCH: 100
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 5e-5
+  WARMUP_EPOCHS: 5
+TRAIN:
+  DATASET: imagenet
+  IM_SIZE: 512
+  BATCH_SIZE: 512
+TEST:
+  DATASET: imagenet
+  IM_SIZE: 512
+  BATCH_SIZE: 400
+NUM_GPUS: 1
+OUT_DIR: .
diff --git a/RegNetY_8G.py b/RegNetY_8G.py
new file mode 100644
index 0000000..bc78a23
--- /dev/null
+++ b/RegNetY_8G.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Test a trained classification model."""
+
+import argparse
+import sys
+
+import numpy as np
+import pycls.core.losses as losses
+import pycls.core.model_builder as model_builder
+import pycls.datasets.loader as loader
+import pycls.utils.benchmark as bu
+import pycls.utils.checkpoint as cu
+import pycls.utils.distributed as du
+import pycls.utils.logging as lu
+import pycls.utils.metrics as mu
+import pycls.utils.multiprocessing as mpu
+import pycls.utils.net as nu
+import torch
+from pycls.core.config import assert_and_infer_cfg, cfg
+from pycls.utils.meters import TestMeter
+
+
+logger = lu.get_logger(__name__)
+
+def log_model_info(model):
+    """Logs model info"""
+    logger.info("Model:\n{}".format(model))
+    logger.info("Params: {:,}".format(mu.params_count(model)))
+    logger.info("Flops: {:,}".format(mu.flops_count(model)))
+    logger.info("Acts: {:,}".format(mu.acts_count(model)))
+
+def build_model():
+
+    # Load config options
+    cfg.merge_from_file('RegNetY-8.0GF_dds_8gpu.yaml')
+    cfg.merge_from_list([])
+    assert_and_infer_cfg()
+    cfg.freeze()
+    # Setup logging
+    lu.setup_logging()
+    # Show the config
+    logger.info("Config:\n{}".format(cfg))
+
+    # Fix the RNG seeds (see RNG comment in core/config.py for discussion)
+    np.random.seed(cfg.RNG_SEED)
+    torch.manual_seed(cfg.RNG_SEED)
+    # Configure the CUDNN backend
+    torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK
+
+    # Build the model (before the loaders to speed up debugging)
+    model = model_builder.build_model()
+    log_model_info(model)
+
+    # Load model weights
+    #cu.load_checkpoint('RegNetY-8.0GF_dds_8gpu.pyth', model)
+    logger.info("Loaded model weights from: {}".format('RegNetY-8.0GF_dds_8gpu.pyth'))
+
+    del model.head
+
+    return model
+
diff --git a/danbooru_labels.txt b/danbooru_labels.txt
new file mode 100644
index 0000000..51ec5ff
--- /dev/null
+++ b/danbooru_labels.txt
@@ -0,0 +1,4096 @@
+0 1girl
+1 solo
+2 long_hair
+3 breasts
+4 blush
+5 smile
+6 looking_at_viewer
+7 short_hair
+8 open_mouth
+9 multiple_girls
+10 blue_eyes
+11 blonde_hair
+12 brown_hair
+13 skirt
+14 large_breasts
+15 hat
+16 thighhighs
+17 black_hair
+18 red_eyes
+19 bangs
+20 hair_ornament
+21 simple_background
+22 ribbon
+23 2girls
+24 gloves
+25 dress
+26 bow
+27 1boy
+28 brown_eyes
+29 navel
+30 white_background
+31 twintails
+32 shirt
+33 school_uniform
+34 underwear
+35 long_sleeves
+36 eyebrows_visible_through_hair
+37 cleavage
+38 medium_breasts
+39 sitting
+40 green_eyes
+41 animal_ears
+42 monochrome
+43 panties
+44 very_long_hair
+45 nipples
+46 blue_hair
+47 bare_shoulders
+48 jewelry
+49 comic
+50 purple_eyes
+51 hair_ribbon
+52 closed_eyes
+53 holding
+54 black_legwear
+55 weapon
+56 greyscale
+57 hair_between_eyes
+58 ponytail
+59 standing
+60 purple_hair
+61 yellow_eyes
+62 pink_hair
+63 ass
+64 swimsuit
+65 collarbone
+66 hair_bow
+67 tail
+68 silver_hair
+69 flower
+70 full_body
+71 braid
+72 :d
+73 ahoge
+74 pantyhose
+75 hairband
+76 closed_mouth
+77 boots
+78 nude
+79 red_hair
+80 heart
+81 jacket
+82 wings
+83 green_hair
+84 short_sleeves
+85 pleated_skirt
+86 japanese_clothes
+87 one_eye_closed
+88 serafuku
+89 glasses
+90 detached_sleeves
+91 food
+92 lying
+93 upper_body
+94 small_breasts
+95 sweat
+96 bikini
+97 male_focus
+98 multiple_boys
+99 hetero
+100 censored
+101 barefoot
+102 necktie
+103 shoes
+104 white_legwear
+105 sky
+106 white_hair
+107 sidelocks
+108 cowboy_shot
+109 frills
+110 earrings
+111 pussy
+112 elbow_gloves
+113 penis
+114 striped
+115 tears
+116 open_clothes
+117 white_shirt
+118 day
+119 sword
+120 parted_lips
+121 3girls
+122 shorts
+123 fang
+124 midriff
+125 alternate_costume
+126 looking_back
+127 solo_focus
+128 thighs
+129 hairclip
+130 outdoors
+131 pointy_ears
+132 tongue
+133 choker
+134 cloud
+135 cat_ears
+136 belt
+137 sleeveless
+138 multicolored_hair
+139 puffy_sleeves
+140 cum
+141 horns
+142 chibi
+143 shiny
+144 white_gloves
+145 hair_flower
+146 fingerless_gloves
+147 2boys
+148 uniform
+149 spread_legs
+150 kimono
+151 yuri
+152 pink_eyes
+153 white_panties
+154 scarf
+155 miniskirt
+156 cape
+157 artist_name
+158 on_back
+159 black_gloves
+160 teeth
+161 star
+162 wide_sleeves
+163 armpits
+164 sex
+165 water
+166 flat_chest
+167 socks
+168 sweatdrop
+169 :o
+170 bra
+171 side_ponytail
+172 tongue_out
+173 character_name
+174 bag
+175 dark_skin
+176 black_eyes
+177 hood
+178 armor
+179 sketch
+180 from_behind
+181 signature
+182 zettai_ryouiki
+183 huge_breasts
+184 orange_hair
+185 necklace
+186 covered_nipples
+187 twin_braids
+188 aqua_eyes
+189 apron
+190 blunt_bangs
+191 grey_hair
+192 bunny_ears
+193 indoors
+194 grin
+195 bowtie
+196 vest
+197 sailor_collar
+198 pants
+199 lips
+200 bracelet
+201 black_skirt
+202 military
+203 hair_over_one_eye
+204 feet
+205 vaginal
+206 two_side_up
+207 high_heels
+208 dated
+209 twitter_username
+210 nail_polish
+211 tree
+212 arms_up
+213 collar
+214 off_shoulder
+215 kneehighs
+216 book
+217 gun
+218 head_tilt
+219 mosaic_censoring
+220 4koma
+221 4girls
+222 white_dress
+223 collared_shirt
+224 sweater
+225 maid
+226 military_uniform
+227 looking_at_another
+228 groin
+229 arm_up
+230 legs
+231 cosplay
+232 cat_tail
+233 shiny_skin
+234 open_shirt
+235 hug
+236 torn_clothes
+237 grey_background
+238 holding_weapon
+239 dutch_angle
+240 wrist_cuffs
+241 plaid
+242 cup
+243 hand_on_hip
+244 petals
+245 one-piece_swimsuit
+246 wet
+247 mole
+248 no_bra
+249 kneeling
+250 aqua_hair
+251 sash
+252 hand_up
+253 maid_headdress
+254 stomach
+255 loli
+256 v-shaped_eyebrows
+257 black_dress
+258 puffy_short_sleeves
+259 pillow
+260 neckerchief
+261 from_side
+262 black_footwear
+263 red_neckwear
+264 no_panties
+265 english_text
+266 6+girls
+267 see-through
+268 black_panties
+269 bed
+270 fingernails
+271 blue_sky
+272 single_braid
+273 leotard
+274 red_ribbon
+275 pubic_hair
+276 blue_skirt
+277 blood
+278 saliva
+279 witch_hat
+280 fruit
+281 ascot
+282 orange_eyes
+283 blurry
+284 headgear
+285 gradient
+286 copyright_name
+287 sparkle
+288 speech_bubble
+289 pussy_juice
+290 ^_^
+291 bottomless
+292 underboob
+293 magical_girl
+294 alternate_hairstyle
+295 headphones
+296 headband
+297 red_bow
+298 siblings
+299 dress_shirt
+300 sideboob
+301 v
+302 bodysuit
+303 parody
+304 pantyshot
+305 profile
+306 sleeves_past_wrists
+307 no_humans
+308 eyebrows
+309 :3
+310 shiny_hair
+311 umbrella
+312 striped_legwear
+313 areolae
+314 double_bun
+315 bare_legs
+316 uncensored
+317 cameltoe
+318 window
+319 makeup
+320 gradient_background
+321 hair_tubes
+322 holding_hands
+323 side-tie_bikini
+324 capelet
+325 traditional_media
+326 leaning_forward
+327 bell
+328 symbol-shaped_pupils
+329 fox_ears
+330 chain
+331 thigh_boots
+332 drill_hair
+333 cover
+334 cum_in_pussy
+335 rose
+336 two-tone_hair
+337 coat
+338 fur_trim
+339 topless
+340 anus
+341 night
+342 bare_arms
+343 mask
+344 short_shorts
+345 glowing
+346 parted_bangs
+347 arm_support
+348 wariza
+349 mouth_hold
+350 hoodie
+351 open_jacket
+352 fox_tail
+353 blush_stickers
+354 lingerie
+355 school_swimsuit
+356 eyelashes
+357 strapless
+358 bird
+359 embarrassed
+360 turtleneck
+361 bat_wings
+362 skirt_lift
+363 moon
+364 frown
+365 transparent_background
+366 one_side_up
+367 shirt_lift
+368 bound
+369 looking_to_the_side
+370 bed_sheet
+371 bdsm
+372 black_ribbon
+373 eyepatch
+374 eating
+375 remodel_(kantai_collection)
+376 beret
+377 plaid_skirt
+378 tattoo
+379 depth_of_field
+380 ocean
+381 garter_straps
+382 crop_top
+383 wavy_mouth
+384 leaf
+385 grey_eyes
+386 back
+387 nose_blush
+388 animal
+389 underwear_only
+390 from_above
+391 on_side
+392 mole_under_eye
+393 heterochromia
+394 skindentation
+395 grabbing
+396 stuffed_toy
+397 hair_bobbles
+398 thigh_gap
+399 bandages
+400 5girls
+401 crossed_arms
+402 watermark
+403 obi
+404 gradient_hair
+405 ass_visible_through_thighs
+406 arms_behind_back
+407 covered_navel
+408 pokemon_(creature)
+409 detached_collar
+410 chair
+411 shadow
+412 wavy_hair
+413 blazer
+414 hat_ribbon
+415 flying_sweatdrops
+416 sleeveless_dress
+417 undressing
+418 cat
+419 blue_dress
+420 floating_hair
+421 sandals
+422 low_twintails
+423 bar_censor
+424 beach
+425 red_skirt
+426 sleeping
+427 neck_ribbon
+428 knee_boots
+429 looking_away
+430 katana
+431 bondage
+432 cum_on_body
+433 sleeveless_shirt
+434 fangs
+435 crossover
+436 chinese_clothes
+437 younger
+438 polka_dot
+439 short_dress
+440 outstretched_arms
+441 stuffed_animal
+442 ;d
+443 couple
+444 striped_panties
+445 toes
+446 short_twintails
+447 >_<
+448 expressionless
+449 formal
+450 breast_grab
+451 thigh_strap
+452 multiple_views
+453 panty_pull
+454 lipstick
+455 oral
+456 black_bow
+457 scrunchie
+458 /\/\/\
+459 mob_cap
+460 helmet
+461 trembling
+462 facial_mark
+463 from_below
+464 scar
+465 medium_hair
+466 bent_over
+467 hair_bun
+468 floral_print
+469 buttons
+470 light_smile
+471 face
+472 3boys
+473 swept_bangs
+474 suspenders
+475 sunlight
+476 cover_page
+477 blue_background
+478 crown
+479 happy
+480 leg_up
+481 cleavage_cutout
+482 antenna_hair
+483 wolf_ears
+484 facial_hair
+485 halterneck
+486 sisters
+487 hands_up
+488 table
+489 pov
+490 tan
+491 backpack
+492 phone
+493 bug
+494 skirt_set
+495 heart-shaped_pupils
+496 polearm
+497 on_bed
+498 crossed_legs
+499 casual
+500 fire
+501 soles
+502 no_shoes
+503 crying
+504 squatting
+505 tokin_hat
+506 pink_panties
+507 wind
+508 bike_shorts
+509 outstretched_arm
+510 insect
+511 pink_background
+512 pointing
+513 cherry_blossoms
+514 black_headwear
+515 staff
+516 monster_girl
+517 crescent
+518 black_jacket
+519 genderswap
+520 looking_down
+521 on_stomach
+522 otoko_no_ko
+523 red_dress
+524 no_pants
+525 surprised
+526 suit
+527 bunnysuit
+528 highleg
+529 all_fours
+530 brown_footwear
+531 cardigan
+532 side-tie_panties
+533 lace
+534 denim
+535 gauntlets
+536 third_eye
+537 multiple_tails
+538 shaded_face
+539 convenient_censoring
+540 loafers
+541 covering
+542 tiara
+543 cum_on_upper_body
+544 head_wings
+545 knife
+546 peaked_cap
+547 white_bikini
+548 black_bikini
+549 checkered
+550 cellphone
+551 blouse
+552 towel
+553 instrument
+554 grass
+555 ground_vehicle
+556 bottle
+557 breast_press
+558 fellatio
+559 ring
+560 standing_on_one_leg
+561 pink_bow
+562 child
+563 kiss
+564 feathers
+565 holding_sword
+566 multicolored
+567 cloudy_sky
+568 doujinshi
+569 straddling
+570 hair_intakes
+571 half-closed_eyes
+572 breasts_outside
+573 black_shirt
+574 looking_up
+575 shinkaisei-kan
+576 tank_top
+577 high_ponytail
+578 :<
+579 sunglasses
+580 facial
+581 fake_animal_ears
+582 light_brown_hair
+583 hat_bow
+584 nature
+585 no_headwear
+586 plant
+587 goggles
+588 girl_on_top
+589 ?
+590 black_neckwear
+591 between_breasts
+592 adapted_costume
+593 bow_panties
+594 sheath
+595 musical_note
+596 muscle
+597 slit_pupils
+598 tareme
+599 hand_on_own_chest
+600 white_headwear
+601 demon_girl
+602 pale_skin
+603 animal_ear_fluff
+604 bob_cut
+605 butterfly
+606 fan
+607 front-tie_top
+608 frilled_skirt
+609 ribbon_trim
+610 curtains
+611 foreshortening
+612 microphone
+613 juliet_sleeves
+614 letterboxed
+615 lavender_hair
+616 no_hat
+617 wristband
+618 mary_janes
+619 upskirt
+620 candy
+621 angry
+622 cross
+623 bikini_top
+624 box
+625 t-shirt
+626 twin_drills
+627 puffy_nipples
+628 hands
+629 frilled_dress
+630 anger_vein
+631 testicles
+632 wing_collar
+633 drooling
+634 building
+635 clenched_teeth
+636 wet_clothes
+637 tsurime
+638 white_skirt
+639 clenched_hand
+640 cum_on_breasts
+641 star_(sky)
+642 blue_bow
+643 anal
+644 machinery
+645 full_moon
+646 blue_ribbon
+647 brown_legwear
+648 animal_print
+649 broom
+650 horn
+651 christmas
+652 armband
+653 ...
+654 hair_flaps
+655 scenery
+656 waist_apron
+657 strap_slip
+658 streaked_hair
+659 dual_persona
+660 :p
+661 carrying
+662 glowing_eyes
+663 eyes_visible_through_hair
+664 rope
+665 jingle_bell
+666 white_footwear
+667 brooch
+668 genderswap_(mtf)
+669 breast_hold
+670 x_hair_ornament
+671 messy_hair
+672 skin_tight
+673 snow
+674 hakama
+675 elf
+676 corset
+677 bunny_tail
+678 night_sky
+679 border
+680 motor_vehicle
+681 clothes_writing
+682 strapless_dress
+683 web_address
+684 long_legs
+685 semi-rimless_eyewear
+686 personification
+687 rifle
+688 veil
+689 china_dress
+690 breath
+691 breasts_apart
+692 kemonomimi_mode
+693 green_skirt
+694 blurry_background
+695 lifted_by_self
+696 androgynous
+697 abs
+698 mecha
+699 thick_thighs
+700 santa_hat
+701 pendant
+702 single_thighhigh
+703 multiple_penises
+704 spiked_hair
+705 finger_to_mouth
+706 bridal_gauntlets
+707 handgun
+708 torn_legwear
+709 wolf_tail
+710 black_hairband
+711 pink_dress
+712 nontraditional_miko
+713 santa_costume
+714 sharp_teeth
+715 black_bra
+716 headset
+717 revealing_clothes
+718 o_o
+719 tentacles
+720 sneakers
+721 vertical_stripes
+722 steam
+723 red_flower
+724 cross-laced_footwear
+725 group_sex
+726 thick_eyebrows
+727 hair_over_shoulder
+728 hood_down
+729 virtual_youtuber
+730 tanline
+731 6+boys
+732 fishnets
+733 frilled_sleeves
+734 blue_sailor_collar
+735 close-up
+736 panties_under_pantyhose
+737 erection
+738 half_updo
+739 headwear_removed
+740 piercing
+741 couch
+742 licking
+743 white_bow
+744 gift
+745 sailor_dress
+746 gem
+747 claws
+748 red_footwear
+749 seiza
+750 dress_lift
+751 dog_ears
+752 paizuri
+753 doujin_cover
+754 holding_food
+755 blue_neckwear
+756 dual_wielding
+757 white_ribbon
+758 lolita_fashion
+759 forest
+760 side_braid
+761 masturbation
+762 red-framed_eyewear
+763 naughty_face
+764 micro_bikini
+765 ribbed_sweater
+766 sweater_vest
+767 lens_flare
+768 handjob
+769 turret
+770 teacup
+771 between_legs
+772 ball
+773 ejaculation
+774 eye_contact
+775 gym_uniform
+776 =_=
+777 motion_lines
+778 arms_behind_head
+779 garter_belt
+780 gen_1_pokemon
+781 buckle
+782 alcohol
+783 minigirl
+784 wading
+785 cuffs
+786 baseball_cap
+787 halloween
+788 pom_pom_(clothes)
+789 after_sex
+790 dark_skinned_male
+791 folded_ponytail
+792 crossdressing
+793 v_arms
+794 flying
+795 wide_hips
+796 blue_shirt
+797 side_slit
+798 mole_under_mouth
+799 buruma
+800 bespectacled
+801 interlocked_fingers
+802 desk
+803 fish
+804 new_year
+805 legs_apart
+806 fingering
+807 camisole
+808 jitome
+809 index_finger_raised
+810 spread_pussy
+811 spear
+812 low-tied_long_hair
+813 cloak
+814 bow_(weapon)
+815 skull
+816 epaulettes
+817 armlet
+818 yellow_neckwear
+819 tray
+820 black_background
+821 light_particles
+822 curvy
+823 bandaid
+824 bowl
+825 innertube
+826 hip_focus
+827 spot_color
+828 demon_tail
+829 oekaki
+830 purple_dress
+831 older
+832 starry_sky
+833 legs_up
+834 hands_together
+835 text_focus
+836 monster
+837 bunny
+838 black_wings
+839 rain
+840 spikes
+841 bubble
+842 ribbon-trimmed_sleeves
+843 card_(medium)
+844 rape
+845 dakimakura
+846 holding_gun
+847 clenched_hands
+848 4boys
+849 empty_eyes
+850 hat_removed
+851 hair_scrunchie
+852 pencil_skirt
+853 out_of_frame
+854 smoke
+855 own_hands_together
+856 zipper
+857 thong
+858 cake
+859 yellow_background
+860 shield
+861 feet_out_of_frame
+862 |_|
+863 under-rim_eyewear
+864 french_braid
+865 red_legwear
+866 partially_submerged
+867 long_fingernails
+868 hair_rings
+869 twins
+870 white_flower
+871 teddy_bear
+872 dog
+873 page_number
+874 cum_in_mouth
+875 black_shorts
+876 sunset
+877 contrapposto
+878 smartphone
+879 spoken_heart
+880 blue_legwear
+881 ice
+882 doggystyle
+883 black_choker
+884 snake
+885 restrained
+886 areola_slip
+887 sleeves_rolled_up
+888 faceless
+889 happy_birthday
+890 arm_at_side
+891 impossible_clothes
+892 upside-down
+893 pelvic_curtain
+894 hands_on_hips
+895 high_heel_boots
+896 spoken_ellipsis
+897 beard
+898 school_bag
+899 thighband_pantyhose
+900 yukata
+901 straight_hair
+902 reflection
+903 walking
+904 track_jacket
+905 blue_bikini
+906 curly_hair
+907 miko
+908 tress_ribbon
+909 popsicle
+910 shawl
+911 denim_shorts
+912 black_serafuku
+913 running
+914 ;)
+915 plate
+916 butt_crack
+917 sun_hat
+918 blue_jacket
+919 paw_pose
+920 yellow_bow
+921 striped_bikini
+922 backlighting
+923 short_ponytail
+924 white_bra
+925 heavy_breathing
+926 outstretched_hand
+927 black_pants
+928 sarashi
+929 hand_on_another's_head
+930 wince
+931 knees_up
+932 :q
+933 swimsuit_under_clothes
+934 shiny_clothes
+935 breast_squeeze
+936 short_hair_with_long_locks
+937 pink_skirt
+938 red_shirt
+939 arm_behind_back
+940 extra_ears
+941 cumdrip
+942 +_+
+943 long_skirt
+944 hooded_jacket
+945 red_gloves
+946 ghost
+947 upper_teeth
+948 o-ring
+949 drinking_glass
+950 demon_wings
+951 covering_breasts
+952 contemporary
+953 geta
+954 freckles
+955 cowgirl_position
+956 doll
+957 hairpin
+958 top_hat
+959 hug_from_behind
+960 bloomers
+961 eighth_note
+962 tabard
+963 single_glove
+964 aircraft
+965 pajamas
+966 mini_hat
+967 crying_with_eyes_open
+968 fishnet_legwear
+969 goggles_on_head
+970 military_hat
+971 jumping
+972 blue_panties
+973 serious
+974 hair_down
+975 pose
+976 realistic
+977 plump
+978 chopsticks
+979 full-face_blush
+980 colored_eyelashes
+981 facing_viewer
+982 ofuda
+983 cannon
+984 smirk
+985 injury
+986 frog_hair_ornament
+987 maid_apron
+988 portrait
+989 pantyshot_(sitting)
+990 object_insertion
+991 emphasis_lines
+992 red_nails
+993 tied_hair
+994 red_background
+995 white_jacket
+996 watch
+997 pink_flower
+998 paws
+999 legs_together
+1000 valentine
+1001 tassel
+1002 circlet
+1003 everyone
+1004 sheathed
+1005 angel_wings
+1006 spread_arms
+1007 purple_legwear
+1008 floating
+1009 pocket
+1010 round_teeth
+1011 two-tone_background
+1012 brown_gloves
+1013 ass_grab
+1014 zoom_layer
+1015 underwater
+1016 panties_around_one_leg
+1017 !
+1018 pink_nails
+1019 nose
+1020 multicolored_clothes
+1021 hand_between_legs
+1022 forehead
+1023 pointless_censoring
+1024 emblem
+1025 clitoris
+1026 string_bikini
+1027 feathered_wings
+1028 holding_book
+1029 fur_collar
+1030 futanari
+1031 wedding_dress
+1032 brother_and_sister
+1033 frilled_bikini
+1034 muneate
+1035 white_apron
+1036 snowing
+1037 gothic_lolita
+1038 cigarette
+1039 brown_background
+1040 cum_on_hair
+1041 holding_cup
+1042 lollipop
+1043 yaoi
+1044 camera
+1045 chocolate
+1046 hand_in_hair
+1047 arrow
+1048 knees_together_feet_apart
+1049 bunny_girl
+1050 jeans
+1051 anchor
+1052 male_pubic_hair
+1053 red_bikini
+1054 eyeshadow
+1055 back-to-back
+1056 arm_warmers
+1057 frog
+1058 competition_swimsuit
+1059 photo
+1060 lowleg
+1061 kita_high_school_uniform
+1062 purple_background
+1063 demon_horns
+1064 :t
+1065 naval_uniform
+1066 blue_theme
+1067 clothed_sex
+1068 eyewear_on_head
+1069 cropped_legs
+1070 antennae
+1071 red_rose
+1072 bun_cover
+1073 stairs
+1074 condom
+1075 hitodama
+1076 sailor_hat
+1077 low_ponytail
+1078 wand
+1079 strawberry
+1080 mountain
+1081 outline
+1082 black-framed_eyewear
+1083 meme_attire
+1084 ice_cream
+1085 adjusting_hair
+1086 mittens
+1087 bouquet
+1088 green_background
+1089 bouncing_breasts
+1090 white_skin
+1091 panties_aside
+1092 anklet
+1093 city
+1094 purple_skirt
+1095 pauldrons
+1096 grey_skirt
+1097 green_dress
+1098 light_rays
+1099 unzipped
+1100 oni
+1101 hime_cut
+1102 pocky
+1103 dragon
+1104 arched_back
+1105 asymmetrical_hair
+1106 pink_legwear
+1107 dog_tail
+1108 vibrator
+1109 mouse_ears
+1110 height_difference
+1111 cum_on_lower_body
+1112 oni_horns
+1113 hands_in_pockets
+1114 armored_dress
+1115 eyeball
+1116 rock
+1117 hand_in_pocket
+1118 faceless_male
+1119 beads
+1120 mirror
+1121 lantern
+1122 sign
+1123 anchor_symbol
+1124 gohei
+1125 card
+1126 pink_ribbon
+1127 poke_ball
+1128 nurse_cap
+1129 spoon
+1130 leg_lift
+1131 unbuttoned
+1132 white_swimsuit
+1133 untied
+1134 waitress
+1135 crotch_seam
+1136 age_difference
+1137 pink_bra
+1138 !?
+1139 handbag
+1140 sun
+1141 tentacle_hair
+1142 fighting_stance
+1143 adjusting_clothes
+1144 white_blouse
+1145 striped_shirt
+1146 folding_fan
+1147 name_tag
+1148 hair_bell
+1149 green_bow
+1150 sunflower
+1151 nipple_slip
+1152 crystal
+1153 robe
+1154 pink_lips
+1155 headpiece
+1156 open_book
+1157 sportswear
+1158 blue_footwear
+1159 visor_cap
+1160 mustache
+1161 blindfold
+1162 headdress
+1163 apple
+1164 knee_up
+1165 sitting_on_person
+1166 bandana
+1167 center_opening
+1168 paper
+1169 waving
+1170 pout
+1171 pouch
+1172 clock
+1173 drinking_straw
+1174 hoop_earrings
+1175 thought_bubble
+1176 halter_top
+1177 red_headwear
+1178 earmuffs
+1179 guitar
+1180 black_leotard
+1181 silhouette
+1182 peach
+1183 red_jacket
+1184 sleeves_past_fingers
+1185 red_scarf
+1186 lace-up_boots
+1187 argyle
+1188 cropped_jacket
+1189 battle
+1190 dildo
+1191 breastplate
+1192 wind_lift
+1193 tearing_up
+1194 petticoat
+1195 salute
+1196 jack-o'-lantern
+1197 shirtless
+1198 blanket
+1199 clothed_female_nude_male
+1200 >:)
+1201 hanging_breasts
+1202 triangular_headpiece
+1203 partially_visible_vulva
+1204 pistol
+1205 yokozuwari
+1206 no_nose
+1207 airplane
+1208 mature
+1209 backless_outfit
+1210 alternate_color
+1211 music
+1212 silent_comic
+1213 gag
+1214 top-down_bottom-up
+1215 shota
+1216 fork
+1217 robot
+1218 basket
+1219 shoulder_armor
+1220 military_vehicle
+1221 covering_mouth
+1222 mecha_musume
+1223 motion_blur
+1224 ninja
+1225 chestnut_mouth
+1226 halo
+1227 highleg_leotard
+1228 winter_clothes
+1229 toned
+1230 sexually_suggestive
+1231 borrowed_character
+1232 front-tie_bikini
+1233 number
+1234 fairy_wings
+1235 crescent_hair_ornament
+1236 bell_collar
+1237 grey_legwear
+1238 blue_flower
+1239 veins
+1240 blue_skin
+1241 tabi
+1242 scythe
+1243 wide-eyed
+1244 bukkake
+1245 sundress
+1246 gakuran
+1247 leash
+1248 fantasy
+1249 drink
+1250 bad_anatomy
+1251 frilled_apron
+1252 alternate_hair_length
+1253 hair_tie
+1254 labcoat
+1255 fence
+1256 frilled_shirt_collar
+1257 chin_rest
+1258 black_nails
+1259 petting
+1260 reading
+1261 onsen
+1262 creature
+1263 android
+1264 flat_cap
+1265 directional_arrow
+1266 bandeau
+1267 kneepits
+1268 forehead_mark
+1269 highleg_panties
+1270 hakama_skirt
+1271 style_parody
+1272 palm_tree
+1273 head_out_of_frame
+1274 snake_hair_ornament
+1275 door
+1276 gangbang
+1277 hand_on_headwear
+1278 microskirt
+1279 missionary
+1280 skirt_hold
+1281 adjusting_eyewear
+1282 bookshelf
+1283 silver_eyes
+1284 double_v
+1285 groping
+1286 flag
+1287 chinese_text
+1288 straw_hat
+1289 brown_skirt
+1290 bucket
+1291 floral_background
+1292 arms_at_sides
+1293 water_drop
+1294 :>
+1295 asymmetrical_wings
+1296 enmaided
+1297 talking
+1298 lactation
+1299 @_@
+1300 white_hairband
+1301 shoulder_bag
+1302 covering_crotch
+1303 taut_clothes
+1304 fox_girl
+1305 magic_circle
+1306 ahegao
+1307 food_on_face
+1308 strap_gap
+1309 glint
+1310 arm_grab
+1311 pink_shirt
+1312 pantyshot_(standing)
+1313 can
+1314 cat_girl
+1315 science_fiction
+1316 character_doll
+1317 bow_bra
+1318 raised_eyebrows
+1319 holding_umbrella
+1320 blue_swimsuit
+1321 orange_bow
+1322 railing
+1323 glass
+1324 pilot_suit
+1325 stuffed_bunny
+1326 string
+1327 key
+1328 wooden_floor
+1329 high_collar
+1330 tasuki
+1331 sack
+1332 gap
+1333 belt_buckle
+1334 animal_hood
+1335 yellow_ribbon
+1336 parasol
+1337 torpedo
+1338 nurse
+1339 smoking
+1340 hair_stick
+1341 hand_on_own_cheek
+1342 orange_background
+1343 striped_bow
+1344 garters
+1345 tatami
+1346 ear_piercing
+1347 witch
+1348 armored_boots
+1349 tate_eboshi
+1350 lolita_hairband
+1351 asymmetrical_legwear
+1352 succubus
+1353 grey_shirt
+1354 drinking
+1355 blue_headwear
+1356 ice_wings
+1357 beanie
+1358 ankle_boots
+1359 gift_box
+1360 road
+1361 leg_hug
+1362 high-waist_skirt
+1363 logo
+1364 size_difference
+1365 alternate_breast_size
+1366 triangle_mouth
+1367 magic
+1368 hands_clasped
+1369 beachball
+1370 pen
+1371 riding
+1372 bath
+1373 marker_(medium)
+1374 hands_on_own_chest
+1375 hibiscus
+1376 oriental_umbrella
+1377 sand
+1378 star_hair_ornament
+1379 horizon
+1380 striped_neckwear
+1381 shade
+1382 strapless_leotard
+1383 gagged
+1384 red_hakama
+1385 jacket_on_shoulders
+1386 bangle
+1387 lineart
+1388 pink_footwear
+1389 vambraces
+1390 pumpkin
+1391 black_sclera
+1392 licking_lips
+1393 partly_fingerless_gloves
+1394 gen_3_pokemon
+1395 domino_mask
+1396 dragon_girl
+1397 inverted_nipples
+1398 5boys
+1399 yukkuri_shiteitte_ne
+1400 blood_on_face
+1401 hair_censor
+1402 bridal_veil
+1403 saliva_trail
+1404 sarong
+1405 hand_on_own_face
+1406 nosebleed
+1407 cheerleader
+1408 polka_dot_background
+1409 rigging
+1410 black_sailor_collar
+1411 hair_up
+1412 graphite_(medium)
+1413 bare_back
+1414 holster
+1415 mug
+1416 flipped_hair
+1417 incest
+1418 hand_to_own_mouth
+1419 cabbie_hat
+1420 one-piece_tan
+1421 bandaged_arm
+1422 shorts_under_skirt
+1423 bat
+1424 snowflakes
+1425 purple_panties
+1426 watercraft
+1427 2koma
+1428 pink_neckwear
+1429 lace-trimmed_legwear
+1430 purple_ribbon
+1431 large_areolae
+1432 puffy_long_sleeves
+1433 fur
+1434 plugsuit
+1435 bathing
+1436 naked_shirt
+1437 wedding_band
+1438 bound_wrists
+1439 heart_of_string
+1440 animal_costume
+1441 underbust
+1442 lace-trimmed_panties
+1443 on_head
+1444 torii
+1445 biting
+1446 drunk
+1447 multicolored_eyes
+1448 +++
+1449 on_floor
+1450 plaid_vest
+1451 arm_behind_head
+1452 petite
+1453 facepaint
+1454 print_legwear
+1455 indian_style
+1456 fairy
+1457 white_leotard
+1458 cityscape
+1459 classroom
+1460 bracer
+1461 white_scarf
+1462 laughing
+1463 randoseru
+1464 lamp
+1465 turtleneck_sweater
+1466 vertical-striped_legwear
+1467 japanese_armor
+1468 open_coat
+1469 blue_nails
+1470 gen_4_pokemon
+1471 mini_crown
+1472 dark
+1473 pool
+1474 wagashi
+1475 furry
+1476 bench
+1477 dragon_horns
+1478 ranguage
+1479 car
+1480 outside_border
+1481 cravat
+1482 kicking
+1483 gen_5_pokemon
+1484 space
+1485 ribbon_choker
+1486 princess_carry
+1487 green_neckwear
+1488 open_fly
+1489 hood_up
+1490 pom_poms
+1491 covered_mouth
+1492 short_kimono
+1493 spoken_question_mark
+1494 brown_jacket
+1495 purple_bow
+1496 candle
+1497 sample
+1498 gen_2_pokemon
+1499 whisker_markings
+1500 scabbard
+1501 carrot
+1502 teapot
+1503 tea
+1504 pigeon-toed
+1505 watercolor_(medium)
+1506 cable
+1507 overalls
+1508 hair_over_eyes
+1509 against_wall
+1510 alternate_hair_color
+1511 white_kimono
+1512 leg_garter
+1513 star-shaped_pupils
+1514 greaves
+1515 bright_pupils
+1516 sports_bra
+1517 blue_shorts
+1518 holding_phone
+1519 cum_on_clothes
+1520 electricity
+1521 pink_bikini
+1522 hands_on_own_face
+1523 alternate_eye_color
+1524 assault_rifle
+1525 absurdly_long_hair
+1526 multiple_persona
+1527 raccoon_ears
+1528 patreon_username
+1529 wristwatch
+1530 mermaid
+1531 naked_apron
+1532 computer
+1533 hand_on_another's_shoulder
+1534 leaning_back
+1535 toeless_legwear
+1536 hatsune_miku
+1537 hakurei_reimu
+1538 kirisame_marisa
+1539 remilia_scarlet
+1540 izayoi_sakuya
+1541 flandre_scarlet
+1542 patchouli_knowledge
+1543 alice_margatroid
+1544 kochiya_sanae
+1545 cirno
+1546 yakumo_yukari
+1547 admiral_(kantai_collection)
+1548 konpaku_youmu
+1549 artoria_pendragon_(all)
+1550 shameimaru_aya
+1551 fujiwara_no_mokou
+1552 hong_meiling
+1553 akemi_homura
+1554 reisen_udongein_inaba
+1555 komeiji_koishi
+1556 kaname_madoka
+1557 saigyouji_yuyuko
+1558 komeiji_satori
+1559 inubashiri_momiji
+1560 kaga_(kantai_collection)
+1561 yakumo_ran
+1562 miki_sayaka
+1563 saber
+1564 kazami_yuuka
+1565 moriya_suwako
+1566 rumia
+1567 kamishirasawa_keine
+1568 konpaku_youmu_(ghost)
+1569 chen
+1570 reiuji_utsuho
+1571 kagamine_rin
+1572 sakura_kyouko
+1573 tomoe_mami
+1574 shimakaze_(kantai_collection)
+1575 kawashiro_nitori
+1576 hibiki_(kantai_collection)
+1577 hinanawi_tenshi
+1578 tatara_kogasa
+1579 jeanne_d'arc_(fate)_(all)
+1580 kaenbyou_rin
+1581 houraisan_kaguya
+1582 kongou_(kantai_collection)
+1583 koakuma
+1584 hijiri_byakuren
+1585 inazuma_(kantai_collection)
+1586 inaba_tewi
+1587 ibuki_suika
+1588 akagi_(kantai_collection)
+1589 mystia_lorelei
+1590 shigure_(kantai_collection)
+1591 tenryuu_(kantai_collection)
+1592 yasaka_kanako
+1593 mash_kyrielight
+1594 nazrin
+1595 yagokoro_eirin
+1596 akiyama_mio
+1597 houjuu_nue
+1598 ikazuchi_(kantai_collection)
+1599 megurine_luka
+1600 kagamine_len
+1601 souryuu_asuka_langley
+1602 mizuhashi_parsee
+1603 yuudachi_(kantai_collection)
+1604 akatsuki_(kantai_collection)
+1605 morichika_rinnosuke
+1606 zuikaku_(kantai_collection)
+1607 fate_testarossa
+1608 nakano_azusa
+1609 nagato_yuki
+1610 suzumiya_haruhi
+1611 hoshiguma_yuugi
+1612 nagato_(kantai_collection)
+1613 shanghai_doll
+1614 hirasawa_yui
+1615 fubuki_(kantai_collection)
+1616 toyosatomimi_no_miko
+1617 kagiyama_hina
+1618 kyubey
+1619 haruna_(kantai_collection)
+1620 hamakaze_(kantai_collection)
+1621 wriggle_nightbug
+1622 tamamo_(fate)_(all)
+1623 kashima_(kantai_collection)
+1624 toosaka_rin
+1625 onozuka_komachi
+1626 ryuujou_(kantai_collection)
+1627 takamachi_nanoha
+1628 mononobe_no_futo
+1629 nero_claudius_(fate)_(all)
+1630 tainaka_ritsu
+1631 shiki_eiki
+1632 daiyousei
+1633 toramaru_shou
+1634 fujimaru_ritsuka_(female)
+1635 northern_ocean_hime
+1636 suzuya_(kantai_collection)
+1637 nishikino_maki
+1638 abigail_williams_(fate/grand_order)
+1639 jeanne_d'arc_(alter)_(fate)
+1640 houshou_(kantai_collection)
+1641 misaka_mikoto
+1642 sonoda_umi
+1643 nishizumi_miho
+1644 yazawa_nico
+1645 shibuya_rin
+1646 matoi_ryuuko
+1647 himekaidou_hatate
+1648 atago_(kantai_collection)
+1649 toujou_nozomi
+1650 rem_(re:zero)
+1651 usami_renko
+1652 murakumo_(kantai_collection)
+1653 ayase_eli
+1654 ushio_(kantai_collection)
+1655 nagae_iku
+1656 link
+1657 serval_(kemono_friends)
+1658 shoukaku_(kantai_collection)
+1659 murasa_minamitsu
+1660 imaizumi_kagerou
+1661 maribel_hearn
+1662 yukikaze_(kantai_collection)
+1663 ayanami_rei
+1664 rensouhou-chan
+1665 inkling
+1666 amatsukaze_(kantai_collection)
+1667 kotobuki_tsumugi
+1668 kaku_seiga
+1669 kitakami_(kantai_collection)
+1670 kumoi_ichirin
+1671 illyasviel_von_einzbern
+1672 hoshii_miki
+1673 kurodani_yamame
+1674 hiiragi_kagami
+1675 scathach_(fate)_(all)
+1676 gilgamesh
+1677 pikachu
+1678 tatsuta_(kantai_collection)
+1679 kaito
+1680 ibaraki_kasen
+1681 hata_no_kokoro
+1682 female_admiral_(kantai_collection)
+1683 fujimaru_ritsuka_(male)
+1684 tamamo_no_mae_(fate)
+1685 kasodani_kyouko
+1686 akebono_(kantai_collection)
+1687 prinz_eugen_(kantai_collection)
+1688 kuujou_joutarou
+1689 minami_kotori
+1690 soga_no_tojiko
+1691 izumi_konata
+1692 jeanne_d'arc_(fate)
+1693 matou_sakura
+1694 okita_souji_(fate)_(all)
+1695 yamato_(kantai_collection)
+1696 joseph_joestar_(young)
+1697 mutsu_(kantai_collection)
+1698 yorha_no._2_type_b
+1699 miyako_yoshika
+1700 amami_haruka
+1701 suigintou
+1702 wo-class_aircraft_carrier
+1703 c.c.
+1704 kafuu_chino
+1705 kijin_seija
+1706 aki_minoriko
+1707 bismarck_(kantai_collection)
+1708 yoko_littner
+1709 meiko
+1710 kousaka_honoka
+1711 kaban_(kemono_friends)
+1712 ooi_(kantai_collection)
+1713 hieda_no_akyuu
+1714 ooyodo_(kantai_collection)
+1715 kisaragi_chihaya
+1716 sendai_(kantai_collection)
+1717 d.va_(overwatch)
+1718 shijou_takane
+1719 shikinami_asuka_langley
+1720 minase_iori
+1721 takao_(kantai_collection)
+1722 akashi_(kantai_collection)
+1723 shiranui_(kantai_collection)
+1724 kyon
+1725 aki_shizuha
+1726 letty_whiterock
+1727 hoshizora_rin
+1728 futatsuiwa_mamizou
+1729 ex-keine
+1730 astolfo_(fate)
+1731 darjeeling_(girls_und_panzer)
+1732 lillie_(pokemon)
+1733 hiei_(kantai_collection)
+1734 kinomoto_sakura
+1735 gumi
+1736 asahina_mikuru
+1737 archer
+1738 ganaha_hibiki
+1739 souryuu_(kantai_collection)
+1740 ro-500_(kantai_collection)
+1741 senketsu
+1742 kikuchi_makoto
+1743 kirishima_(kantai_collection)
+1744 nero_claudius_(fate)
+1745 chun-li
+1746 sukuna_shinmyoumaru
+1747 asashio_(kantai_collection)
+1748 sazanami_(kantai_collection)
+1749 kiso_(kantai_collection)
+1750 nishizumi_maho
+1751 kiryuuin_satsuki
+1752 verniy_(kantai_collection)
+1753 dio_brando
+1754 princess_zelda
+1755 itsumi_erika
+1756 scathach_(fate/grand_order)
+1757 yamashiro_(kantai_collection)
+1758 caesar_anthonio_zeppeli
+1759 zero_two_(darling_in_the_franxx)
+1760 asuna_(sao)
+1761 sekibanki
+1762 emiya_shirou
+1763 clownpiece
+1764 anchovy_(girls_und_panzer)
+1765 takatsuki_yayoi
+1766 shimamura_uzuki
+1767 super_sonico
+1768 akiyama_yukari
+1769 musashi_(kantai_collection)
+1770 okita_souji_(fate)
+1771 barnaby_brooks_jr
+1772 samidare_(kantai_collection)
+1773 black_rock_shooter_(character)
+1774 shuten_douji_(fate/grand_order)
+1775 hiryuu_(kantai_collection)
+1776 rider
+1777 hestia_(danmachi)
+1778 seaport_hime
+1779 yuubari_(kantai_collection)
+1780 iowa_(kantai_collection)
+1781 kisume
+1782 narukami_yuu
+1783 koizumi_hanayo
+1784 naka_(kantai_collection)
+1785 tifa_lockhart
+1786 kaburagi_t_kotetsu
+1787 ultimate_madoka
+1788 sanya_v_litvyak
+1789 saber_alter
+1790 takagaki_kaede
+1791 graf_zeppelin_(kantai_collection)
+1792 watanabe_you
+1793 samus_aran
+1794 eila_ilmatar_juutilainen
+1795 kasumi_(kantai_collection)
+1796 gokou_ruri
+1797 bowsette
+1798 cu_chulainn_(fate)_(all)
+1799 wakasagihime
+1800 kanzaki_ranko
+1801 i-19_(kantai_collection)
+1802 mordred_(fate)_(all)
+1803 touko_(pokemon)
+1804 kotomine_kirei
+1805 sagisawa_fumika
+1806 hiiragi_tsukasa
+1807 yagami_hayate
+1808 nitta_minami
+1809 morrigan_aensland
+1810 yuzuki_yukari
+1811 megumin
+1812 nia_teppelin
+1813 minamoto_no_raikou_(fate/grand_order)
+1814 djeeta_(granblue_fantasy)
+1815 fairy_(kantai_collection)
+1816 shirasaka_koume
+1817 i-58_(kantai_collection)
+1818 lunasa_prismriver
+1819 shirogane_naoto
+1820 hikari_(pokemon)
+1821 sheryl_nome
+1822 junko_(touhou)
+1823 kakyouin_noriaki
+1824 mercy_(overwatch)
+1825 tsukino_usagi
+1826 zuihou_(kantai_collection)
+1827 jougasaki_mika
+1828 medicine_melancholy
+1829 red_(pokemon)
+1830 maya_(kantai_collection)
+1831 furude_rika
+1832 haruka_(pokemon)
+1833 midorikawa_nao
+1834 lily_white
+1835 anastasia_(idolmaster)
+1836 mei_(pokemon)
+1837 oshino_shinobu
+1838 non-human_admiral_(kantai_collection)
+1839 kishin_sagume
+1840 hagiwara_yukiho
+1841 pyonta
+1842 kise_yayoi
+1843 miura_azusa
+1844 tsushima_yoshiko
+1845 princess_peach
+1846 beatrice
+1847 stocking_(psg)
+1848 miqo'te
+1849 ikari_shinji
+1850 kuma_(kantai_collection)
+1851 charlotte_(madoka_magica)
+1852 kirito
+1853 takebe_saori
+1854 kousaka_kirino
+1855 lelouch_lamperouge
+1856 jonathan_joestar
+1857 haruno_sakura
+1858 hecatia_lapislazuli
+1859 mordred_(fate)
+1860 akizuki_ritsuko
+1861 cammy_white
+1862 futaba_anzu
+1863 satonaka_chie
+1864 mankanshoku_mako
+1865 izumi_sagiri
+1866 taihou_(kantai_collection)
+1867 unzan
+1868 z1_leberecht_maass_(kantai_collection)
+1869 emiya_kiritsugu
+1870 fusou_(kantai_collection)
+1871 z3_max_schultz_(kantai_collection)
+1872 bb_(fate)_(all)
+1873 i-401_(kantai_collection)
+1874 atago_(azur_lane)
+1875 waver_velvet
+1876 simon
+1877 asakura_ryouko
+1878 lancer
+1879 lancer_(fate/zero)
+1880 ichinose_shiki
+1881 jougasaki_rika
+1882 ryuuguu_rena
+1883 murasame_(kantai_collection)
+1884 ahri
+1885 uzumaki_naruto
+1886 shirai_kuroko
+1887 mizuki_(pokemon)
+1888 aoki_reika
+1889 maekawa_miku
+1890 producer_(idolmaster)
+1891 erica_hartmann
+1892 hyuuga_hinata
+1893 mikasa_ackerman
+1894 holo
+1895 hyuuga_(kantai_collection)
+1896 miyafuji_yoshika
+1897 accelerator
+1898 kuujou_jolyne
+1899 kallen_stadtfeld
+1900 hishikawa_rikka
+1901 aisaka_taiga
+1902 hk416_(girls_frontline)
+1903 giorno_giovanna
+1904 kuroki_tomoko
+1905 kumano_(kantai_collection)
+1906 aoba_(kantai_collection)
+1907 unicorn_(azur_lane)
+1908 ushiromiya_battler
+1909 jintsuu_(kantai_collection)
+1910 elizabeth_bathory_(fate)_(all)
+1911 sakurauchi_riko
+1912 ruby_rose
+1913 futami_mami
+1914 hino_akane_(smile_precure!)
+1915 katyusha_(girls_und_panzer)
+1916 motoori_kosuzu
+1917 alice_(wonderland)
+1918 takarada_rikka
+1919 tokitsukaze_(kantai_collection)
+1920 tenshi_(angel_beats!)
+1921 meltryllis
+1922 kasumi_(pokemon)
+1923 luna_child
+1924 jack_the_ripper_(fate/apocrypha)
+1925 yamakaze_(kantai_collection)
+1926 gertrud_barkhorn
+1927 unryuu_(kantai_collection)
+1928 ashigara_(kantai_collection)
+1929 makise_kurisu
+1930 kiyoshimo_(kantai_collection)
+1931 warspite_(kantai_collection)
+1932 shimada_arisu
+1933 louise_francoise_le_blanc_de_la_valliere
+1934 isokaze_(kantai_collection)
+1935 belfast_(azur_lane)
+1936 higashikata_jousuke
+1937 kousaka_tamaki
+1938 corrin_(fire_emblem)
+1939 akitsu_maru_(kantai_collection)
+1940 reizei_mako
+1941 kashiwazaki_sena
+1942 kyonko
+1943 mario
+1944 urakaze_(kantai_collection)
+1945 re-class_battleship
+1946 shinku
+1947 doremy_sweet
+1948 taigei_(kantai_collection)
+1949 satsuki_(kantai_collection)
+1950 naganami_(kantai_collection)
+1951 sunny_milk
+1952 ishtar_(fate)_(all)
+1953 star_sapphire
+1954 mogami_(kantai_collection)
+1955 houjou_hibiki
+1956 nitocris_(fate/grand_order)
+1957 kamijou_touma
+1958 uzuki_(kantai_collection)
+1959 kay_(girls_und_panzer)
+1960 lyrica_prismriver
+1961 koshimizu_sachiko
+1962 lucina_(fire_emblem)
+1963 kagerou_(kantai_collection)
+1964 aida_mana
+1965 toshinou_kyouko
+1966 oboro_(kantai_collection)
+1967 satoshi_(pokemon)
+1968 merlin_prismriver
+1969 futami_ami
+1970 ranka_lee
+1971 miyamoto_musashi_(fate/grand_order)
+1972 shana
+1973 u-511_(kantai_collection)
+1974 common_raccoon_(kemono_friends)
+1975 usami_sumireko
+1976 florence_nightingale_(fate/grand_order)
+1977 ikamusume
+1978 fennec_(kemono_friends)
+1979 rider_(fate/zero)
+1980 akigumo_(kantai_collection)
+1981 robin_(fire_emblem)
+1982 weiss_schnee
+1983 kurumi_erika
+1984 yuuki_makoto
+1985 kiyohime_(fate/grand_order)
+1986 kirby
+1987 hoshizora_miyuki
+1988 battleship_hime
+1989 amagi_yukiko
+1990 suiseiseki
+1991 prinz_eugen_(azur_lane)
+1992 aori_(splatoon)
+1993 raising_heart
+1994 nagisa_kaworu
+1995 harusame_(kantai_collection)
+1996 mutsuki_(kantai_collection)
+1997 takami_chika
+1998 dizzy
+1999 nami_(one_piece)
+2000 takanashi_rikka
+2001 abukuma_(kantai_collection)
+2002 boko_(girls_und_panzer)
+2003 vita
+2004 charlotte_e_yeager
+2005 i-8_(kantai_collection)
+2006 hotaru_(splatoon)
+2007 hatsuyuki_(kantai_collection)
+2008 tone_(kantai_collection)
+2009 koizumi_itsuki
+2010 hoto_cocoa
+2011 kujikawa_rise
+2012 makinami_mari_illustrious
+2013 irisviel_von_einzbern
+2014 akaza_akari
+2015 jun'you_(kantai_collection)
+2016 perrine_h_clostermann
+2017 michishio_(kantai_collection)
+2018 blue_(pokemon)
+2019 shiranui_mai
+2020 son_gokuu
+2021 noumi_kudryavka
+2022 hayami_kanade
+2023 haguro_(kantai_collection)
+2024 akuma_homura
+2025 houjou_satoko
+2026 tachibana_arisu
+2027 shiratsuyu_(kantai_collection)
+2028 chloe_von_einzbern
+2029 ise_(kantai_collection)
+2030 kaenbyou_rin_(cat)
+2031 hanekawa_tsubasa
+2032 vivio
+2033 jakuzure_nonon
+2034 junketsu
+2035 lynette_bishop
+2036 isuzu_hana
+2037 christa_renz
+2038 katsuki_yuuri
+2039 sailor_moon
+2040 wa2000_(girls_frontline)
+2041 shinki
+2042 hiro_(darling_in_the_franxx)
+2043 i-168_(kantai_collection)
+2044 narmaya_(granblue_fantasy)
+2045 ia_(vocaloid)
+2046 corrin_(fire_emblem)_(female)
+2047 ryougi_shiki
+2048 takara_miyuki
+2049 kisaragi_(kantai_collection)
+2050 viktor_nikiforov
+2051 tippy_(gochiusa)
+2052 saratoga_(kantai_collection)
+2053 yuzuriha_inori
+2054 kirino_ranmaru
+2055 oda_nobunaga_(fate)_(all)
+2056 honda_mio
+2057 sakamoto_mio
+2058 uchiha_sasuke
+2059 yuki_miku
+2060 akizuki_(kantai_collection)
+2061 senjougahara_hitagi
+2062 panty_(psg)
+2063 yang_xiao_long
+2064 ereshkigal_(fate/grand_order)
+2065 chitanda_eru
+2066 kenzaki_makoto
+2067 nero_claudius_(swimsuit_caster)_(fate)
+2068 last_order
+2069 kanna_kamui
+2070 bardiche
+2071 sonozaki_mion
+2072 index
+2073 kurosawa_dia
+2074 ishtar_(fate/grand_order)
+2075 mika_(girls_und_panzer)
+2076 kirima_sharo
+2077 kunikida_hanamaru
+2078 elizabeth_bathory_(fate)
+2079 yowane_haku
+2080 hatsuzuki_(kantai_collection)
+2081 hanamura_yousuke
+2082 signum
+2083 johnny_joestar
+2084 matsuno_karamatsu
+2085 arcueid_brunestud
+2086 saten_ruiko
+2087 tamamo_no_mae_(swimsuit_lancer)_(fate)
+2088 matou_kariya
+2089 matsuno_osomatsu
+2090 noel_vermillion
+2091 hanasaki_tsubomi
+2092 matsuno_juushimatsu
+2093 ram_(re:zero)
+2094 homura_(xenoblade_2)
+2095 minamino_kanade
+2096 yahagi_(kantai_collection)
+2097 matsuno_ichimatsu
+2098 takao_(azur_lane)
+2099 maru-yu_(kantai_collection)
+2100 souseiseki
+2101 shoebill_(kemono_friends)
+2102 nachi_(kantai_collection)
+2103 ump45_(girls_frontline)
+2104 otonashi_kotori
+2105 emilia_(re:zero)
+2106 aegis_(persona)
+2107 sendai_hakurei_no_miko
+2108 oda_nobunaga_(fate)
+2109 gangut_(kantai_collection)
+2110 yura_(kantai_collection)
+2111 mima
+2112 katsushika_hokusai_(fate/grand_order)
+2113 female_protagonist_(persona_3)
+2114 kurosawa_ruby
+2115 blake_belladonna
+2116 tama_(kantai_collection)
+2117 kagari_atsuko
+2118 producer_(idolmaster_cinderella_girls_anime)
+2119 saber_lily
+2120 eren_yeager
+2121 nero_claudius_(bride)_(fate)
+2122 makoto_nanaya
+2123 matsuura_kanan
+2124 neptune_(neptune_series)
+2125 jeanne_d'arc_alter_santa_lily
+2126 formidable_(azur_lane)
+2127 aqua_(konosuba)
+2128 ushiromiya_ange
+2129 kamikaze_(kantai_collection)
+2130 tracer_(overwatch)
+2131 choukai_(kantai_collection)
+2132 hirasawa_ui
+2133 rensouhou-kun
+2134 katsuragi_(kantai_collection)
+2135 gamagoori_ira
+2136 nagatsuki_(kantai_collection)
+2137 miyuki_(kantai_collection)
+2138 hatsune_miku_(append)
+2139 serena_(pokemon)
+2140 matsuno_choromatsu
+2141 gold_(pokemon)
+2142 cure_peace
+2143 nonna_(girls_und_panzer)
+2144 ayanami_(azur_lane)
+2145 t-head_admiral
+2146 tamamo_cat_(fate)
+2147 tomoe_gozen_(fate/grand_order)
+2148 mei_(overwatch)
+2149 yorigami_shion
+2150 funami_yui
+2151 ujimatsu_chiya
+2152 hanyuu
+2153 mikazuki_munechika
+2154 mimura_kanako
+2155 mary_(pokemon)
+2156 artoria_pendragon_(lancer)
+2157 moroboshi_kirari
+2158 kasumi_(doa)
+2159 yukine_chris
+2160 fou_(fate/grand_order)
+2161 akagi_miria
+2162 princess_king_boo
+2163 suzukaze_(kantai_collection)
+2164 miyamoto_frederica
+2165 kamui_gakupo
+2166 fujibayashi_kyou
+2167 uraraka_ochako
+2168 yuri_lowell
+2169 francesca_lucchini
+2170 ymir_(shingeki_no_kyojin)
+2171 matsuno_todomatsu
+2172 kashuu_kiyomitsu
+2173 jeanne_d'arc_(alter_swimsuit_berserker)
+2174 lucky_beast_(kemono_friends)
+2175 ragna_the_bloodedge
+2176 yuugumo_(kantai_collection)
+2177 hikari_(xenoblade_2)
+2178 shinjou_akane
+2179 charlotte_dunois
+2180 orange_pekoe_(girls_und_panzer)
+2181 makigumo_(kantai_collection)
+2182 bb_(fate/extra_ccc)
+2183 furukawa_nagisa
+2184 gambier_bay_(kantai_collection)
+2185 yayoi_(kantai_collection)
+2186 tatsumi_kanji
+2187 shirayuki_(kantai_collection)
+2188 han_juri
+2189 tedeza_rize
+2190 natsume_rin
+2191 i-class_destroyer
+2192 pepperoni_(girls_und_panzer)
+2193 japanese_crested_ibis_(kemono_friends)
+2194 levi_(shingeki_no_kyojin)
+2195 kamina
+2196 sakura_miku
+2197 yamato-no-kami_yasusada
+2198 misumi_nagisa
+2199 widowmaker_(overwatch)
+2200 kamio_misuzu
+2201 watatsuki_no_yorihime
+2202 kizuna_ai
+2203 ookido_green
+2204 miyu_edelfelt
+2205 kotone_(pokemon)
+2206 yuno
+2207 northern_white-faced_owl_(kemono_friends)
+2208 pola_(kantai_collection)
+2209 isuzu_(kantai_collection)
+2210 tsukikage_yuri
+2211 suiren_(pokemon)
+2212 enemy_aircraft_(kantai_collection)
+2213 yumemi_riamu
+2214 shikinami_(kantai_collection)
+2215 mizuno_ami
+2216 lilith_aensland
+2217 enterprise_(azur_lane)
+2218 horikawa_raiko
+2219 ibaraki_douji_(fate/grand_order)
+2220 hassan_of_serenity_(fate)
+2221 uiharu_kazari
+2222 mao_(pokemon)
+2223 dark_magician_girl
+2224 agano_(kantai_collection)
+2225 kuroshio_(kantai_collection)
+2226 shirona_(pokemon)
+2227 caster_(fate/zero)
+2228 touya_(pokemon)
+2229 byleth_(fire_emblem)
+2230 failure_penguin
+2231 phosphophyllite
+2232 sinon
+2233 higashi_setsuna
+2234 konjiki_no_yami
+2235 sakata_gintoki
+2236 berserker_(fate/zero)
+2237 myoudouin_itsuki
+2238 kagura_(gintama)
+2239 nico_robin
+2240 elsa_(frozen)
+2241 elin_(tera)
+2242 felicia
+2243 etna
+2244 gran_(granblue_fantasy)
+2245 laffey_(azur_lane)
+2246 ump9_(girls_frontline)
+2247 madotsuki
+2248 yoshikawa_chinatsu
+2249 teruzuki_(kantai_collection)
+2250 aino_minako
+2251 nanami_chiaki
+2252 akizuki_ryou
+2253 tsurumaru_kuninaga
+2254 ivan_karelin
+2255 edward_elric
+2256 leafa
+2257 kawakaze_(kantai_collection)
+2258 asashimo_(kantai_collection)
+2259 honma_meiko
+2260 saotome_ranma
+2261 kuma_(persona_4)
+2262 chibi_usa
+2263 alice_margatroid_(pc-98)
+2264 haramura_nodoka
+2265 kamiya_nao
+2266 nagi
+2267 taihou_(azur_lane)
+2268 lyn_(fire_emblem)
+2269 kaguya_luna
+2270 cloud_strife
+2271 sonozaki_shion
+2272 arashio_(kantai_collection)
+2273 akagi_(azur_lane)
+2274 tina_branford
+2275 toosaka_tokiomi
+2276 namazuo_toushirou
+2277 kishibe_rohan
+2278 octoling
+2279 eevee
+2280 sengoku_nadeko
+2281 tachibana_hibiki_(symphogear)
+2282 honebami_toushirou
+2283 kamoi_(kantai_collection)
+2284 houjou_karen
+2285 myoukou_(kantai_collection)
+2286 cure_marine
+2287 gardevoir
+2288 ohara_mari
+2289 mithra
+2290 p-head_producer
+2291 ayanami_(kantai_collection)
+2292 amamiya_ren
+2293 illustrious_(azur_lane)
+2294 noire
+2295 noshiro_(kantai_collection)
+2296 seiran_(touhou)
+2297 frederica_bernkastel
+2298 nekomusume
+2299 libeccio_(kantai_collection)
+2300 shiomi_shuuko
+2301 altera_(fate)
+2302 hayashimo_(kantai_collection)
+2303 artoria_pendragon_(swimsuit_rider_alter)
+2304 yukishiro_honoka
+2305 shokuhou_misaki
+2306 tokisaki_kurumi
+2307 little_boy_admiral_(kantai_collection)
+2308 minna-dietlinde_wilcke
+2309 kasugano_sakura
+2310 n_(pokemon)
+2311 midare_toushirou
+2312 asui_tsuyu
+2313 camilla_(fire_emblem)
+2314 huang_baoling
+2315 silica
+2316 tatsumaki
+2317 hourai_doll
+2318 tsukino_mito
+2319 midna
+2320 hoshino_fumina
+2321 kururugi_suzaku
+2322 manjuu_(azur_lane)
+2323 isonami_(kantai_collection)
+2324 tsukumo_benben
+2325 m1903_springfield_(girls_frontline)
+2326 su-san
+2327 kaga_(azur_lane)
+2328 tokiko_(touhou)
+2329 sento_isuzu
+2330 ringo_(touhou)
+2331 nakano_miku
+2332 maebara_keiichi
+2333 alisa_ilinichina_amiella
+2334 jean_pierre_polnareff
+2335 helena_blavatsky_(fate/grand_order)
+2336 purple_heart
+2337 robin_(fire_emblem)_(female)
+2338 ike_(fire_emblem)
+2339 joseph_joestar
+2340 nu-13
+2341 furutaka_(kantai_collection)
+2342 tsuruya
+2343 katori_(kantai_collection)
+2344 yui_(angel_beats!)
+2345 edelgard_von_hresvelg
+2346 reiuji_utsuho_(bird)
+2347 roll
+2348 kirijou_mitsuru
+2349 slaine_troyard
+2350 yuuri_(pokemon)
+2351 hoshi_shouko
+2352 hagikaze_(kantai_collection)
+2353 ootsuki_yui
+2354 ezo_red_fox_(kemono_friends)
+2355 maikaze_(kantai_collection)
+2356 toono_akiha
+2357 oomuro_sakurako
+2358 totoki_airi
+2359 izumi-no-kami_kanesada
+2360 fukuzawa_yumi
+2361 uryuu_ryuunosuke
+2362 araragi_koyomi
+2363 kirigaya_suguha
+2364 cagliostro_(granblue_fantasy)
+2365 ru-class_battleship
+2366 rosehip_(girls_und_panzer)
+2367 saniwa_(touken_ranbu)
+2368 sakuma_mayu
+2369 bayonetta_(character)
+2370 okita_souji_(alter)_(fate)
+2371 bel_(pokemon)
+2372 artoria_pendragon_(lancer_alter)
+2373 bb_(swimsuit_mooncancer)_(fate)
+2374 shouhou_(kantai_collection)
+2375 shidare_hotaru
+2376 kinugasa_(kantai_collection)
+2377 kirigiri_kyouko
+2378 cure_beauty
+2379 cure_sunshine
+2380 amanogawa_kirara
+2381 blanc
+2382 yin
+2383 riesz
+2384 gyro_zeppeli
+2385 crystal_(pokemon)
+2386 sirius_(azur_lane)
+2387 morikubo_nono
+2388 osakabe-hime_(fate/grand_order)
+2389 rachel_alucard
+2390 eurasian_eagle_owl_(kemono_friends)
+2391 grey_wolf_(kemono_friends)
+2392 nepgear
+2393 caster
+2394 sanageyama_uzu
+2395 ta-class_battleship
+2396 hino_rei
+2397 little_red_riding_hood_(grimm)
+2398 android_18
+2399 subaru_nakajima
+2400 lum
+2401 hachikuji_mayoi
+2402 popuko
+2403 komaki_manaka
+2404 aerith_gainsborough
+2405 umikaze_(kantai_collection)
+2406 momozono_love
+2407 kamikita_komari
+2408 mysterious_heroine_x_(alter)
+2409 enoshima_junko
+2410 robin_(fire_emblem)_(male)
+2411 rydia
+2412 cure_blossom
+2413 chikuma_(kantai_collection)
+2414 diana_cavendish
+2415 chitose_(kantai_collection)
+2416 mamiya_(kantai_collection)
+2417 silver_fox_(kemono_friends)
+2418 oikawa_shizuku
+2419 yorha_no._9_type_s
+2420 arashi_(kantai_collection)
+2421 estellise_sidos_heurassein
+2422 jeanne_d'arc_(swimsuit_archer)
+2423 kadotani_anzu
+2424 ooshio_(kantai_collection)
+2425 furutani_himawari
+2426 sona_buvelle
+2427 yuri_(angel_beats!)
+2428 kooh
+2429 emperor_penguin_(kemono_friends)
+2430 kino_makoto
+2431 fuuro_(pokemon)
+2432 pharah_(overwatch)
+2433 atalanta_(fate)
+2434 horikawa_kunihiro
+2435 nanasaki_ai
+2436 armin_arlert
+2437 cure_march
+2438 hina_ichigo
+2439 karina_lyle
+2440 andou_(girls_und_panzer)
+2441 kitashirakawa_tamako
+2442 ark_royal_(kantai_collection)
+2443 yuuki_mikan
+2444 white_mage
+2445 saotome_alto
+2446 manabe_nodoka
+2447 konno_junko
+2448 toudou_shimako
+2449 kirin_(armor)
+2450 sailor_mercury
+2451 cure_happy
+2452 sakurai_momoka
+2453 frankenstein's_monster_(fate)
+2454 abe_nana
+2455 boo
+2456 javelin_(azur_lane)
+2457 kusanagi_motoko
+2458 bruno_buccellati
+2459 reisalin_stout
+2460 yorigami_jo'on
+2461 eve_(elsword)
+2462 rosalina
+2463 medjed
+2464 katsura_hinagiku
+2465 marie_antoinette_(fate/grand_order)
+2466 kasane_teto
+2467 i-26_(kantai_collection)
+2468 mikuma_(kantai_collection)
+2469 shirayuki_hime
+2470 keith_goodman
+2471 yuri_plisetsky
+2472 shindou_takuto
+2473 lusamine_(pokemon)
+2474 tooru_(maidragon)
+2475 tateyama_ayano
+2476 kero
+2477 silver_(pokemon)
+2478 airfield_hime
+2479 byleth_(fire_emblem)_(female)
+2480 miyamizu_mitsuha
+2481 littorio_(kantai_collection)
+2482 cure_melody
+2483 anjou_naruko
+2484 takeba_yukari
+2485 sorceress_(dragon's_crown)
+2486 hatsushimo_(kantai_collection)
+2487 jaguar_(kemono_friends)
+2488 kayneth_el-melloi_archibald
+2489 vegeta
+2490 oreki_houtarou
+2491 fumizuki_(kantai_collection)
+2492 tomoe_hotaru
+2493 ninomiya_asuka
+2494 daidouji_tomoyo
+2495 oumae_kumiko
+2496 kozakura_marry
+2497 aragaki_ayase
+2498 ibuki_fuuko
+2499 iida_(splatoon)
+2500 alpaca_suri_(kemono_friends)
+2501 monkey_d_luffy
+2502 yotsuba_alice
+2503 ryuu_(street_fighter)
+2504 hatsukaze_(kantai_collection)
+2505 darkness_(konosuba)
+2506 tamura_yuri
+2507 watatsuki_no_toyohime
+2508 japan_(hetalia)
+2509 tiki_(fire_emblem)
+2510 kako_(kantai_collection)
+2511 taneshima_popura
+2512 kohaku
+2513 ushiwakamaru_(fate/grand_order)
+2514 kousaka_kyousuke
+2515 mizunashi_akari
+2516 kamitsure_(pokemon)
+2517 king_hassan_(fate/grand_order)
+2518 miss_cloud
+2519 hayasui_(kantai_collection)
+2520 winry_rockbell
+2521 ranma-chan
+2522 assassin_(fate/zero)
+2523 kos-mos
+2524 bulma
+2525 tada_riina
+2526 rurina_(pokemon)
+2527 akatsuki_kirika
+2528 rowlet
+2529 hoshimiya_ichigo
+2530 seiren_(suite_precure)
+2531 hotarumaru
+2532 nishizumi_shiho
+2533 kawashima_momo
+2534 tokitarou_(fate/grand_order)
+2535 elvaan
+2536 ib_(ib)
+2537 nyarlathotep_(nyaruko-san)
+2538 katou_megumi
+2539 marth_(fire_emblem)
+2540 you_(pokemon)
+2541 oktavia_von_seckendorff
+2542 oshida_(girls_und_panzer)
+2543 sailor_venus
+2544 okabe_rintarou
+2545 prisma_illya
+2546 kurokawa_eren
+2547 ciel
+2548 ushiromiya_jessica
+2549 midoriya_izuku
+2550 g11_(girls_frontline)
+2551 ichigo_(darling_in_the_franxx)
+2552 doujima_nanako
+2553 palutena
+2554 nibutani_shinka
+2555 charizard
+2556 shirase_sakuya
+2557 mochizuki_(kantai_collection)
+2558 maou_(maoyuu)
+2559 asagumo_(kantai_collection)
+2560 haro
+2561 america_(hetalia)
+2562 laura_bodewig
+2563 miyanaga_saki
+2564 momo_velia_deviluke
+2565 cure_sunny
+2566 passion_lip
+2567 pipimi
+2568 honolulu_(azur_lane)
+2569 ramlethal_valentine
+2570 tsukumo_yatsuhashi
+2571 amagi_(kantai_collection)
+2572 au_ra
+2573 commander_(azur_lane)
+2574 xp-tan
+2575 yuzuhara_konomi
+2576 luigi
+2577 yuffie_kisaragi
+2578 united_kingdom_(hetalia)
+2579 shimazu_yoshino
+2580 lala_satalin_deviluke
+2581 ushiromiya_maria
+2582 tieria_erde
+2583 yagen_toushirou
+2584 sakura_chiyo
+2585 araragi_karen
+2586 kamado_nezuko
+2587 inumuta_houka
+2588 nowaki_(kantai_collection)
+2589 reinforce_zwei
+2590 danua
+2591 cecilia_alcott
+2592 tashkent_(kantai_collection)
+2593 jin_kisaragi
+2594 sonya_(kill_me_baby)
+2595 furudo_erika
+2596 aisha_landar
+2597 anastasia_(fate/grand_order)
+2598 consort_yu_(fate)
+2599 assam_(girls_und_panzer)
+2600 kisaragi_shintarou
+2601 aino_megumi
+2602 scathach_(swimsuit_assassin)_(fate)
+2603 sailor_mars
+2604 kishinami_hakuno_(female)
+2605 kuchiki_rukia
+2606 yamashiro_(azur_lane)
+2607 sieg_(fate/apocrypha)
+2608 suzukaze_aoba
+2609 meltryllis_(swimsuit_lancer)_(fate)
+2610 scathach_skadi_(fate/grand_order)
+2611 len
+2612 misaka_imouto
+2613 mysterious_heroine_xx_(foreigner)
+2614 cure_black
+2615 carpaccio_(girls_und_panzer)
+2616 anna_(frozen)
+2617 oshawott
+2618 saint_martha
+2619 alastor_(shakugan_no_shana)
+2620 frisk_(undertale)
+2621 mumei_(kabaneri)
+2622 roma_(kantai_collection)
+2623 anegasaki_nene
+2624 nishi_kinuyo
+2625 natori_sana
+2626 etorofu_(kantai_collection)
+2627 suzutsuki_(kantai_collection)
+2628 minamoto_sakura
+2629 piplup
+2630 azura_(fire_emblem)
+2631 zara_(kantai_collection)
+2632 nenohi_(kantai_collection)
+2633 rias_gremory
+2634 sakura_futaba
+2635 takei_hisa
+2636 hex_maniac_(pokemon)
+2637 higuchi_kaede
+2638 momoe_nagisa
+2639 miyako
+2640 shirabe_ako
+2641 teana_lanster
+2642 akitsushima_(kantai_collection)
+2643 yuuki_(pokemon)
+2644 kino
+2645 snivy
+2646 tharja
+2647 yamagishi_fuuka
+2648 ayane_(doa)
+2649 lambdadelta
+2650 mizuno_ai
+2651 komaeda_nagito
+2652 takakura_himari
+2653 sanzen'in_nagi
+2654 clarisse_(granblue_fantasy)
+2655 aquila_(kantai_collection)
+2656 kousaka_reina
+2657 yaoyorozu_momo
+2658 cure_moonlight
+2659 takoluka
+2660 star_platinum
+2661 yumehara_nozomi
+2662 stakes_of_purgatory
+2663 kinu_(kantai_collection)
+2664 yamada_elf
+2665 enkidu_(fate/strange_fake)
+2666 okazaki_yumemi
+2667 sugiura_ayano
+2668 zidane_tribal
+2669 guido_mista
+2670 yuigahama_yui
+2671 lulu_(league_of_legends)
+2672 takamaki_anne
+2673 monokuma
+2674 hinata_(angel_beats!)
+2675 sakagami_tomoyo
+2676 ichigo_hitofuri
+2677 ri-class_heavy_cruiser
+2678 xuanzang_(fate/grand_order)
+2679 kujou_karen
+2680 saitama_(one-punch_man)
+2681 lucario
+2682 jinx_(league_of_legends)
+2683 chibi_miku
+2684 sucy_manbavaran
+2685 sailor_jupiter
+2686 nursery_rhyme_(fate/extra)
+2687 female_saniwa_(touken_ranbu)
+2688 kazanari_tsubasa
+2689 yukinoshita_yukino
+2690 hatsuharu_(kantai_collection)
+2691 kars_(jojo)
+2692 angela_balzac
+2693 ogata_chieri
+2694 female_protagonist_(pokemon_go)
+2695 hime_(splatoon)
+2696 chi-chi_(dragon_ball)
+2697 eureka
+2698 oomori_yuuko
+2699 medb_(fate)_(all)
+2700 nemoto_hina
+2701 shantae_(character)
+2702 roy_(fire_emblem)
+2703 caren_hortensia
+2704 garry_(ib)
+2705 milla_maxwell
+2706 hiyou_(kantai_collection)
+2707 rita_mordio
+2708 oribe_yasuna
+2709 naoe_riki
+2710 natsume_(pokemon)
+2711 kasumigaoka_utaha
+2712 kido_tsubomi
+2713 ene_(kagerou_project)
+2714 bianca
+2715 hammann_(azur_lane)
+2716 okazaki_tomoya
+2717 prinny
+2718 pod_(nier_automata)
+2719 ten'ou_haruka
+2720 hisui
+2721 inami_mahiru
+2722 matoi_(pso2)
+2723 gokotai
+2724 yamabuki_inori
+2725 gladio_(pokemon)
+2726 ferry_(granblue_fantasy)
+2727 lio_fotia
+2728 konpaku_youki
+2729 kariya_masaki
+2730 shima_rin
+2731 mohammed_avdol
+2732 qbz-95_(girls_frontline)
+2733 jervis_(kantai_collection)
+2734 regina_(dokidoki!_precure)
+2735 slime_(dragon_quest)
+2736 me-tan
+2737 agrias_oaks
+2738 vira_lilie
+2739 arthur_pendragon_(fate)
+2740 toono_shiki
+2741 shokudaikiri_mitsutada
+2742 alphonse_elric
+2743 godzilla
+2744 hino_akane_(idolmaster)
+2745 harime_nui
+2746 dead_master
+2747 kanbaru_suruga
+2748 2k-tan
+2749 lei_lei
+2750 rockman_(character)
+2751 tanikaze_(kantai_collection)
+2752 cure_peach
+2753 araragi_tsukihi
+2754 leonardo_da_vinci_(fate/grand_order)
+2755 kogitsunemaru
+2756 commandant_teste_(kantai_collection)
+2757 kiriya_aoi
+2758 arare_(kantai_collection)
+2759 annie_leonhardt
+2760 yamada_aoi
+2761 maka_albarn
+2762 kama_(fate/grand_order)
+2763 diego_brando
+2764 cure_rhythm
+2765 isolated_island_oni
+2766 kazama_asuka
+2767 enemy_lifebuoy_(kantai_collection)
+2768 error_musume
+2769 nekomusume_(gegege_no_kitarou_6)
+2770 lightning_farron
+2771 nitocris_(swimsuit_assassin)_(fate)
+2772 manya
+2773 cure_white
+2774 taokaka
+2775 puru-see
+2776 igarashi_kyou_(eroe)
+2777 ichinose_kotomi
+2778 kusugawa_sasara
+2779 feldt_grace
+2780 cure_heart
+2781 setsuna_f_seiei
+2782 mitake_ran
+2783 touwa_erio
+2784 sf-a2_miki
+2785 faris_scherwiz
+2786 toon_link
+2787 tepig
+2788 aozaki_aoko
+2789 platinum_the_trinity
+2790 shihouin_yoruichi
+2791 kohinata_miho
+2792 musashi_(pokemon)
+2793 rena_erindel
+2794 reinforce
+2795 katsuragi_misato
+2796 platelet_(hataraku_saibou)
+2797 belldandy
+2798 cheria_barnes
+2799 lisbeth
+2800 aragaki_shinjirou
+2801 kira_yoshikage
+2802 iori_rinko
+2803 kaiou_michiru
+2804 nishida_satono
+2805 sakawa_(kantai_collection)
+2806 lion_(kemono_friends)
+2807 yuna_(ff10)
+2808 natsuki_subaru
+2809 suomi_kp31_(girls_frontline)
+2810 moose_(kemono_friends)
+2811 chiyoda_(kantai_collection)
+2812 takumi_(fire_emblem)
+2813 lenna_charlotte_tycoon
+2814 kitazawa_shiho
+2815 kanaria
+2816 fubuki_(one-punch_man)
+2817 asbel_lhant
+2818 tsukimiya_ayu
+2819 yellow_(pokemon)
+2820 takanami_(kantai_collection)
+2821 fukuji_mihoko
+2822 kawashima_ami
+2823 yamanaka_ino
+2824 igarashi_futaba_(shiromanta)
+2825 garnet_til_alexandros_xvii
+2826 himejima_akeno
+2827 bulbasaur
+2828 ara_haan
+2829 makoto_(street_fighter)
+2830 vert
+2831 nikka_edvardine_katajainen
+2832 yamagumo_(kantai_collection)
+2833 nadia
+2834 moogle
+2835 euryale
+2836 sawamura_spencer_eriri
+2837 komano_aun
+2838 medb_(fate/grand_order)
+2839 otonashi_(angel_beats!)
+2840 inoue_orihime
+2841 sesshouin_kiara
+2842 afuro_terumi
+2843 meer_campbell
+2844 lalafell
+2845 cleveland_(azur_lane)
+2846 haruno_haruka
+2847 viera
+2848 richelieu_(kantai_collection)
+2849 marie_(girls_und_panzer)
+2850 oyashio_(kantai_collection)
+2851 boudica_(fate/grand_order)
+2852 jude_mathis
+2853 boa_hancock
+2854 nagara_(kantai_collection)
+2855 sophie_(tales)
+2856 mikazuki_(kantai_collection)
+2857 hasegawa_kobato
+2858 dekomori_sanae
+2859 mary_(ib)
+2860 shiina_mayuri
+2861 sasaki_chie
+2862 kuga_natsuki
+2863 klan_klein
+2864 mysterious_heroine_x
+2865 kawakami_mai
+2866 nakano_nino
+2867 revy_(black_lagoon)
+2868 teireida_mai
+2869 bakugou_katsuki
+2870 miia_(monster_musume)
+2871 boota
+2872 hatoba_tsugu_(character)
+2873 orihara_izaya
+2874 bowser
+2875 albedo
+2876 st_ar-15_(girls_frontline)
+2877 aki_(girls_und_panzer)
+2878 asahina_mirai
+2879 napoleon_bonaparte_(fate/grand_order)
+2880 squirtle
+2881 tsunade
+2882 fujisaki_chihiro
+2883 usada_hikaru
+2884 gengar
+2885 ichihara_nina
+2886 cheren_(pokemon)
+2887 akita_neru
+2888 kano_shuuya
+2889 jill_valentine
+2890 matara_okina
+2891 sora_(kingdom_hearts)
+2892 cure_sword
+2893 hikawa_hina
+2894 matsumoto_rangiku
+2895 shinonome_nano
+2896 suzuki_jun
+2897 enma_ai
+2898 minamoto_no_raikou_(swimsuit_lancer)_(fate)
+2899 zeta_(granblue_fantasy)
+2900 hirose_kouichi
+2901 sonohara_anri
+2902 kobayakawa_sae
+2903 matsukaze_tenma
+2904 hei
+2905 rainbow_mika
+2906 hoshimiya_kate
+2907 nunnally_lamperouge
+2908 ibuki_(street_fighter)
+2909 saitou_(pokemon)
+2910 chrom_(fire_emblem)
+2911 kaizuka_inaho
+2912 filia_(skullgirls)
+2913 sailor_saturn
+2914 sion_eltnam_atlasia
+2915 kemomimi-chan_(naga_u)
+2916 kasuga_ayumu
+2917 yuuno_scrya
+2918 pit_(kid_icarus)
+2919 m4_sopmod_ii_(girls_frontline)
+2920 princess_of_moonbrook
+2921 luxanna_crownguard
+2922 sanada_akihiko
+2923 white_rabbit
+2924 aono_miki
+2925 yumeko
+2926 morgiana
+2927 shamal
+2928 euphemia_li_britannia
+2929 okita_sougo
+2930 sawa_azusa
+2931 neo_politan
+2932 zero_(rockman)
+2933 kneesocks_(psg)
+2934 prinz_eugen_(unfading_smile)_(azur_lane)
+2935 kobayakawa_rinko
+2936 nanao_yuriko
+2937 tsukuyomi_shirabe
+2938 iris_(pokemon)
+2939 caro_ru_lushe
+2940 niijima_makoto
+2941 lysithea_von_ordelia
+2942 kiss-shot_acerola-orion_heart-under-blade
+2943 tsukioka_kogane
+2944 tsuchinoko_(kemono_friends)
+2945 selvaria_bles
+2946 tsurumaki_maki
+2947 mizutani_eri
+2948 graf_eisen
+2949 midway_hime
+2950 higashikata_jousuke_(jojolion)
+2951 small-clawed_otter_(kemono_friends)
+2952 elizabeth_bathory_(brave)_(fate)
+2953 yuuki_(sao)
+2954 natori_(kantai_collection)
+2955 takasu_ryuuji
+2956 celes_chere
+2957 hoshino_ruri
+2958 destroyer_hime
+2959 super_pochaco
+2960 lucy_heartfilia
+2961 narancia_ghirga
+2962 harukaze_(kantai_collection)
+2963 alisa_(girls_und_panzer)
+2964 jigglypuff
+2965 mikazuki_yozora
+2966 sunazuka_akira
+2967 cosmog
+2968 iroha_(samurai_spirits)
+2969 alena_(dq4)
+2970 minase_nayuki
+2971 fujimura_taiga
+2972 saigusa_haruka
+2973 kyaru_(princess_connect)
+2974 m16a1_(girls_frontline)
+2975 minazuki_karen
+2976 kotegawa_yui
+2977 hinatsuru_ai
+2978 kazagumo_(kantai_collection)
+2979 reisen
+2980 kazami_yuuka_(pc-98)
+2981 lyria_(granblue_fantasy)
+2982 inuyama_aoi
+2983 mimikyu
+2984 gasai_yuno
+2985 kairi_(kingdom_hearts)
+2986 bazett_fraga_mcremitz
+2987 m4a1_(girls_frontline)
+2988 chen_(cat)
+2989 kazemaru_ichirouta
+2990 yoshida_yuuko_(machikado_mazoku)
+2991 gotland_(kantai_collection)
+2992 dante_(devil_may_cry)
+2993 hiradaira_chisaki
+2994 lavinia_whateley_(fate/grand_order)
+2995 sage_(dq3)
+2996 brynhildr_(fate)
+2997 kagamihara_nadeshiko
+2998 raven_(tales)
+2999 black_hanekawa
+3000 kuroyukihime
+3001 violet_evergarden_(character)
+3002 koyama_yuzu
+3003 magical_ruby
+3004 ginga_nakajima
+3005 naegi_makoto
+3006 bradamante_(fate/grand_order)
+3007 kikuzuki_(kantai_collection)
+3008 nakoruru
+3009 satou_kazuma
+3010 amazon_(dragon's_crown)
+3011 akali
+3012 katou_asuka
+3013 soldier_(dq3)
+3014 uzuki_sayaka
+3015 ogasawara_sachiko
+3016 android_21
+3017 lili_(tekken)
+3018 kyouhei_(pokemon)
+3019 tsubaki_yayoi
+3020 kohinata_miku
+3021 kula_diamond
+3022 amasawa_yuuko
+3023 galko
+3024 kamio_reiji_(yua)
+3025 brown_bear_(kemono_friends)
+3026 hakuryuu_(inazuma_eleven)
+3027 shannon
+3028 judith
+3029 kumada_masaru
+3030 sayo_samonji
+3031 vampy
+3032 kondou_taeko
+3033 eva_02
+3034 yorha_type_a_no._2
+3035 zooey_(granblue_fantasy)
+3036 berserker
+3037 millia_rage
+3038 beatrix_(granblue_fantasy)
+3039 buront
+3040 natalia_(idolmaster)
+3041 ingrid
+3042 takamori_aiko
+3043 okumura_haru
+3044 sonic
+3045 chi-class_torpedo_cruiser
+3046 toga_himiko
+3047 sairenji_haruna
+3048 tenjou_utena
+3049 rabbit_yukine
+3050 yamanaka_sawako
+3051 roronoa_zoro
+3052 kisaragi_momo
+3053 seeu
+3054 mukai_takumi
+3055 cure_lovely
+3056 milfeulle_sakuraba
+3057 kochou_shinobu
+3058 paul_bunyan_(fate/grand_order)
+3059 nikaidou_saki
+3060 bridget_(guilty_gear)
+3061 kurumizawa_satanichia_mcdowell
+3062 hinata_hajime
+3063 nao_(mabinogi)
+3064 nijimura_okuyasu
+3065 karna_(fate)
+3066 tanned_cirno
+3067 kanon_(umineko)
+3068 akashi_(azur_lane)
+3069 anila_(granblue_fantasy)
+3070 asseylum_vers_allusia
+3071 maruyama_aya
+3072 pannacotta_fugo
+3073 kan'u_unchou
+3074 elphelt_valentine
+3075 leo_(fire_emblem)
+3076 anya_alstreim
+3077 himekawa_yuki
+3078 natsume_kyousuke
+3079 graf_zeppelin_(azur_lane)
+3080 ak-12_(girls_frontline)
+3081 wakaba_(kantai_collection)
+3082 princess_daisy
+3083 elesis_(elsword)
+3084 edmond_dantes_(fate/grand_order)
+3085 popplio
+3086 sheik
+3087 mikisugi_aikurou
+3088 takimoto_hifumi
+3089 yumi_(senran_kagura)
+3090 quetzalcoatl_(maidragon)
+3091 nekomiya_hinata
+3092 tanamachi_kaoru
+3093 artoria_pendragon_(swimsuit_ruler)_(fate)
+3094 samson_(skullgirls)
+3095 eas
+3096 kokkoro_(princess_connect!)
+3097 hanna-justina_marseille
+3098 ellen_baker
+3099 jeanne_d'arc_(granblue_fantasy)
+3100 baiken
+3101 lunamaria_hawke
+3102 asakaze_(kantai_collection)
+3103 hidaka_ai
+3104 maria_cadenzavna_eve
+3105 solid_snake
+3106 black_heart
+3107 monika_(doki_doki_literature_club)
+3108 flonne
+3109 huang_lingyin
+3110 muppo
+3111 izayoi_liko
+3112 kisaragi_(azur_lane)
+3113 kizuna_akari
+3114 corrin_(fire_emblem)_(male)
+3115 mikko_(girls_und_panzer)
+3116 okazaki_ushio
+3117 g41_(girls_frontline)
+3118 toudou_yurika
+3119 doraemon_(character)
+3120 kurosaki_ichigo
+3121 arle_nadja
+3122 korra
+3123 shinonono_houki
+3124 may_(guilty_gear)
+3125 hasumi_souji_(eroe)
+3126 kiki
+3127 itoshiki_nozomu
+3128 ooji_mochizou
+3129 eva_beatrice
+3130 squall_leonhart
+3131 angel_(kof)
+3132 i-13_(kantai_collection)
+3133 mukaido_manaka
+3134 tsurumaki_kokoro
+3135 seto_kousuke
+3136 cure_princess
+3137 koizumi_itsuki_(female)
+3138 micaiah_(fire_emblem)
+3139 lina_inverse
+3140 charmander
+3141 roll_caskett
+3142 kurugaya_yuiko
+3143 suou_pavlichenko
+3144 mifune_miyu
+3145 kobayakawa_yutaka
+3146 riven_(league_of_legends)
+3147 asuna_(sao-alo)
+3148 sasha_braus
+3149 reines_el-melloi_archisorte
+3150 jouga_maya
+3151 saya
+3152 endou_mamoru
+3153 semiramis_(fate)
+3154 belarus_(hetalia)
+3155 sora_ginko
+3156 ness
+3157 udagawa_tomoe
+3158 ange_(princess_principal)
+3159 gwendolyn
+3160 shirakiin_ririchiyo
+3161 murasaki_shikibu_(fate)
+3162 abyssal_admiral_(kantai_collection)
+3163 jessica_albert
+3164 alice_cartelet
+3165 tatebayashi_sakurako
+3166 bottle_miku
+3167 gawain_(fate/extra)
+3168 kushieda_minori
+3169 yuel_(granblue_fantasy)
+3170 shuu_(inazuma_eleven)
+3171 ashido_mina
+3172 yuudachi_(azur_lane)
+3173 lambda-11
+3174 sand_cat_(kemono_friends)
+3175 hazama
+3176 yumizuka_satsuki
+3177 torchic
+3178 raichu
+3179 eldridge_(azur_lane)
+3180 priest_(ragnarok_online)
+3181 kagurazaka_asuna
+3182 tachibana_taki
+3183 kusakabe_misao
+3184 tsushima_(kantai_collection)
+3185 chocola_(sayori)
+3186 morishima_haruka
+3187 senkawa_chihiro
+3188 virgilia
+3189 temari
+3190 chikorita
+3191 aoba_moca
+3192 st._louis_(azur_lane)
+3193 heiwajima_shizuo
+3194 raphtalia
+3195 uehara_himari
+3196 rosa_farrell
+3197 soldier:_76_(overwatch)
+3198 colorado_(kantai_collection)
+3199 fuura_kafuka
+3200 meta_knight
+3201 sin_sack
+3202 hikawa_sayo
+3203 evangeline_a_k_mcdowell
+3204 isshiki_akane
+3205 sasaki_chiho
+3206 iori_junpei
+3207 princess_of_the_crystal
+3208 yuri_sakazaki
+3209 iggy_(jojo)
+3210 todoroki_shouto
+3211 shiro_(dennou_shoujo_youtuber_shiro)
+3212 sanji
+3213 prussia_(hetalia)
+3214 naganohara_mio
+3215 natsu_megumi
+3216 cure_pine
+3217 shampoo_(ranma_1/2)
+3218 kokonoe_rin
+3219 heshikiri_hasebe
+3220 elise_(fire_emblem)
+3221 ayatsuji_tsukasa
+3222 humboldt_penguin_(kemono_friends)
+3223 minase_akiko
+3224 erwin_(girls_und_panzer)
+3225 golden_snub-nosed_monkey_(kemono_friends)
+3226 princess_serenity
+3227 plasma-chan_(kantai_collection)
+3228 nanachi_(made_in_abyss)
+3229 katarina_du_couteau
+3230 katagiri_sanae
+3231 yae_sakura
+3232 kongou_(aoki_hagane_no_arpeggio)
+3233 hibari_(senran_kagura)
+3234 hummy_(suite_precure)
+3235 sakurajima_mai
+3236 kiana_kaslana
+3237 sugimoto_reimi
+3238 stheno
+3239 royal_penguin_(kemono_friends)
+3240 aircraft_carrier_oni
+3241 schwertkreuz
+3242 mishaguji
+3243 high_priest
+3244 tokiha_mai
+3245 eva_01
+3246 wild_tiger
+3247 tendou_akane
+3248 isobe_noriko
+3249 shizuka_rin
+3250 shirley_fenette
+3251 sailor_chibi_moon
+3252 arima_senne
+3253 sakura_(fire_emblem)
+3254 izayoi_aki
+3255 z23_(azur_lane)
+3256 antonio_lopez
+3257 dark_sakura
+3258 hacka_doll_3
+3259 busujima_saeko
+3260 vampire_(azur_lane)
+3261 uni_(neptune_series)
+3262 takagi-san
+3263 lily_black
+3264 russia_(hetalia)
+3265 amaterasu
+3266 tsurugi_kyousuke
+3267 akimoto_komachi
+3268 vee_(granblue_fantasy)
+3269 lord_el-melloi_ii
+3270 nathan_seymour
+3271 hilda_valentine_goneril
+3272 venera-sama
+3273 noctis_lucis_caelum
+3274 poison_(final_fight)
+3275 barasuishou
+3276 vanilla_(sayori)
+3277 ushiromiya_lion
+3278 kuriyama_mirai
+3279 nakano_yotsuba
+3280 kuradoberi_jam
+3281 intrepid_(kantai_collection)
+3282 roto
+3283 heles
+3284 ima-no-tsurugi
+3285 blue_rose_(tiger_&_bunny)
+3286 trish_una
+3287 strength_(black_rock_shooter)
+3288 kouki_(pokemon)
+3289 hatake_kakashi
+3290 wang_liu_mei
+3291 alvin_(tales)
+3292 rex_(xenoblade_2)
+3293 erza_scarlet
+3294 carmilla_(fate/grand_order)
+3295 chocobo
+3296 tomori_nao
+3297 mogami_shizuka
+3298 kaidou_minami
+3299 lisa_lisa
+3300 cure_diamond
+3301 priest_(dq3)
+3302 matsudaira_touko
+3303 shizuki_hitomi
+3304 sombra_(overwatch)
+3305 genji_(overwatch)
+3306 sazaki_kaoruko
+3307 kirito_(sao-ggo)
+3308 mint_blancmanche
+3309 lily_(vocaloid)
+3310 iwasaki_minami
+3311 hayasaka_mirei
+3312 ms._fortune_(skullgirls)
+3313 fujibayashi_ryou
+3314 meowth
+3315 cure_beat
+3316 silva_(granblue_fantasy)
+3317 ushiromiya_rosa
+3318 totooria_helmold
+3319 rotom
+3320 johnston_(kantai_collection)
+3321 elly
+3322 sailor_uranus
+3323 tanya_degurechaff
+3324 atago_(midsummer_march)_(azur_lane)
+3325 naomi_(girls_und_panzer)
+3326 konoha_(kagerou_project)
+3327 misaka_worst
+3328 matsuwa_(kantai_collection)
+3329 adachi_tooru
+3330 ushiromiya_natsuhi
+3331 nana_asta_deviluke
+3332 victorica_de_blois
+3333 mordred_(swimsuit_rider)_(fate)
+3334 komori_kiri
+3335 takao_(aoki_hagane_no_arpeggio)
+3336 kuroka_(high_school_dxd)
+3337 kaburagi_kaede
+3338 ichii_yui
+3339 arin
+3340 arisa_bannings
+3341 kasugano_urara_(yes!_precure_5)
+3342 yuuki_yuuna
+3343 taihou_(forbidden_feast)_(azur_lane)
+3344 mizuho_(kantai_collection)
+3345 utsugi_yuuki
+3346 espeon
+3347 tachibana_makoto
+3348 matsumae_ohana
+3349 sun_wukong
+3350 marie_rose
+3351 oyama_mahiro
+3352 son_gohan
+3353 eirika_(fire_emblem)
+3354 kitashirakawa_chiyuri
+3355 miyamoto_musashi_(swimsuit_berserker)_(fate)
+3356 lockon_stratos
+3357 koiwai_yotsuba
+3358 rossweisse
+3359 niyah
+3360 umbreon
+3361 tron_bonne
+3362 pyrrha_nikos
+3363 matoba_risa
+3364 hiiragi_yuzu
+3365 cure_muse_(yellow)
+3366 irako_(kantai_collection)
+3367 kirakishou
+3368 kiyohime_(swimsuit_lancer)_(fate)
+3369 cure_dream
+3370 girl_holding_a_cat_(kantai_collection)
+3371 madoka_aguri
+3372 jibril_(no_game_no_life)
+3373 kotonoha_akane
+3374 merlin_(fate)
+3375 hori_yuuko
+3376 ganondorf
+3377 female_assassin_(fate/zero)
+3378 scheherazade_(fate/grand_order)
+3379 kiyama_hiroto
+3380 watarase_jun
+3381 morpeko
+3382 cure_rosetta
+3383 ping_hai_(azur_lane)
+3384 diamond_(houseki_no_kuni)
+3385 robert_eo_speedwagon
+3386 xander_(fire_emblem)
+3387 i-14_(kantai_collection)
+3388 flareon
+3389 valentine_(skullgirls)
+3390 tsumiki_mikan
+3391 litten
+3392 ouma_kokichi
+3393 hakurei_reimu_(pc-98)
+3394 viral
+3395 odamaki_sapphire
+3396 yui_(sao)
+3397 astaroth_(shinrabanshou)
+3398 celestia_ludenberck
+3399 luke_fon_fabre
+3400 soma_peries
+3401 ibara_mayaka
+3402 minami_kana
+3403 tear_grants
+3404 caesar_(girls_und_panzer)
+3405 candy_(smile_precure!)
+3406 itou_chika
+3407 chara_(undertale)
+3408 luciela_r._sourcream
+3409 einhart_stratos
+3410 tsunashi_hajime
+3411 negev_(girls_frontline)
+3412 samuel_b._roberts_(kantai_collection)
+3413 butz_klauser
+3414 natsuki_rin
+3415 sakata_kintoki_(fate/grand_order)
+3416 komasan
+3417 kanzuki_karin
+3418 frankenstein's_monster_(swimsuit_saber)_(fate)
+3419 tsunashi_kaoru
+3420 elizabeth_(persona)
+3421 sakaguchi_karina
+3422 ookurikara
+3423 cure_passion
+3424 alice_carroll
+3425 african_wild_dog_(kemono_friends)
+3426 okinami_(kantai_collection)
+3427 uchi_emiri
+3428 yukikaze_panettone
+3429 hiro
+3430 yoshi
+3431 luo_tianyi
+3432 sakurazaki_setsuna
+3433 nijou_noriko
+3434 black-tailed_prairie_dog_(kemono_friends)
+3435 aioi_yuuko
+3436 isabella_valentine
+3437 sans
+3438 mai_natsume
+3439 scanty_(psg)
+3440 todoroki_yachiyo
+3441 elize_lutus
+3442 asia_argento
+3443 prosciutto
+3444 allelujah_haptism
+3445 sailor_neptune
+3446 kasuga_maru_(kantai_collection)
+3447 pichu
+3448 kojirou_(pokemon)
+3449 hau_(pokemon)
+3450 leone_abbacchio
+3451 cure_twinkle
+3452 mewtwo
+3453 hyuuga_saki
+3454 miyazaki_nodoka
+3455 ryuuhou_(kantai_collection)
+3456 anemone_(eureka_seven)
+3457 hyur
+3458 kasugano_sora
+3459 fujino_shizuru
+3460 cure_berry
+3461 admiral_graf_spee_(azur_lane)
+3462 rebecca_miyamoto
+3463 fujiwara_chika
+3464 otokura_yuuki
+3465 northern_italy_(hetalia)
+3466 jibanyan
+3467 gokotai's_tigers
+3468 siegfried_(fate)
+3469 tanigawa_kanna
+3470 american_beaver_(kemono_friends)
+3471 narusawa_ryouka
+3472 lyn_(blade_&_soul)
+3473 amamiya_hibiya
+3474 matsuda_chiyohiko
+3475 asamiya_athena
+3476 andira_(granblue_fantasy)
+3477 phosphophyllite_(ll)
+3478 elf_(dragon's_crown)
+3479 jirou_kyouka
+3480 ooiwa_wataru
+3481 zange
+3482 yamanbagiri_kunihiro
+3483 zessica_wong
+3484 acerola_(pokemon)
+3485 chrome_dokuro
+3486 arjuna_(fate/grand_order)
+3487 yuuki_haru
+3488 fubuki_shirou
+3489 yokoyama_nao
+3490 kitsu_chiri
+3491 takano_miyo
+3492 doma_umaru
+3493 vaporeon
+3494 amane_suzuha
+3495 ana_(overwatch)
+3496 ibuki_tsubasa
+3497 gentoo_penguin_(kemono_friends)
+3498 aircraft_carrier_hime
+3499 hierophant_green
+3500 walpurgisnacht_(madoka_magica)
+3501 morgana_(persona_5)
+3502 cyndaquil
+3503 mochizuki_anna
+3504 matsuoka_miu
+3505 takozonesu
+3506 matsuoka_rin
+3507 erika_(pokemon)
+3508 futatsuiwa_mamizou_(human)
+3509 takane_manaka
+3510 caitlyn_(league_of_legends)
+3511 eruruw
+3512 minazuki_(kantai_collection)
+3513 rororina_fryxell
+3514 igarashi_kyouko
+3515 tenma_gabriel_white
+3516 narciso_anasui
+3517 kirisaki_chitoge
+3518 asuka_(senran_kagura)
+3519 minakami_mai
+3520 ken_masters
+3521 okusawa_misaki
+3522 cheshire_cat
+3523 3.1-tan
+3524 kyon_no_imouto
+3525 the_world
+3526 lucas
+3527 kotonomiya_yuki
+3528 satou_shin
+3529 wii_fit_trainer
+3530 robin_hood_(fate)
+3531 nanase_haruka_(free!)
+3532 felyne
+3533 ikezawa_hanako
+3534 mishou_mai
+3535 nishizono_mio
+3536 shinomiya_kaguya
+3537 minato_aqua
+3538 lacus_clyne
+3539 kitten_(gravity_daze)
+3540 cure_honey
+3541 tsutsukakushi_tsukiko
+3542 amae_koromo
+3543 shibuki_ran
+3544 ana_coppola
+3545 nena_trinity
+3546 erina_pendleton
+3547 xenovia_quarta
+3548 sola-ui_nuada-re_sophia-ri
+3549 akamatsu_kaede
+3550 kinoshita_hideyoshi
+3551 fukuda_(girls_und_panzer)
+3552 guts
+3553 okumura_rin
+3554 heidimarie_w_schnaufer
+3555 fran
+3556 lancelot_(fate/grand_order)
+3557 rory_mercury
+3558 shoukaku_(azur_lane)
+3559 byleth_(fire_emblem)_(male)
+3560 wu_zetian_(fate/grand_order)
+3561 ninian_(fire_emblem)
+3562 villetta_nu
+3563 maeda_toushirou
+3564 futaki_kanata
+3565 jack_frost
+3566 naoi_ayato
+3567 hange_zoe
+3568 santa_claus
+3569 cheese-kun
+3570 kanno_naoe
+3571 pixiv-tan
+3572 soraka
+3573 hinamori_amu
+3574 riza_hawkeye
+3575 elizabeth_liones
+3576 mudkip
+3577 sakamoto_(nichijou)
+3578 airi_(queen's_blade)
+3579 sakutarou
+3580 cerulean_(kemono_friends)
+3581 oozora_akari
+3582 pecorine
+3583 queen_elizabeth_(azur_lane)
+3584 banned_artist
+3585 hammer_(sunset_beach)
+3586 haruyama_kazunori
+3587 mizuki_hitoshi
+3588 kouji_(campus_life)
+3589 tani_takeshi
+3590 ebifurya
+3591 itomugi-kun
+3592 tanaka_takayuki
+3593 bkub
+3594 rebecca_(keinelove)
+3595 yohane
+3596 a1
+3597 carnelian
+3598 ichimi
+3599 blade_(galaxist)
+3600 futa_(nabezoko)
+3601 iesupa
+3602 kirisawa_juuzou
+3603 peko
+3604 warugaki_(sk-ii)
+3605 yua_(checkmate)
+3606 kanon_(kurogane_knights)
+3607 dd_(ijigendd)
+3608 masao
+3609 tsuda_nanafushi
+3610 kantoku
+3611 hamu_koutarou
+3612 matsunaga_kouyou
+3613 lolita_channel
+3614 hisahiko
+3615 shino_(ponjiyuusu)
+3616 ido_(teketeke)
+3617 minaba_hideo
+3618 tsunako
+3619 drawfag
+3620 houtengeki
+3621 zounose
+3622 sayori
+3623 ueyama_michirou
+3624 tsukudani_(coke-buta)
+3625 hews_hack
+3626 minami_(colorful_palette)
+3627 abubu
+3628 shimazaki_mujirushi
+3629 mizuki_makoto
+3630 tomose_shunsaku
+3631 ha_akabouzu
+3632 ixy
+3633 tonda
+3634 matsuryuu
+3635 kanikama
+3636 yaegashi_nan
+3637 otoufu
+3638 creayus
+3639 miyo_(ranthath)
+3640 niwatazumi
+3641 rappa_(rappaya)
+3642 itou_noiji
+3643 ilya_kuvshinov
+3644 niiko_(gonnzou)
+3645 ebi_193
+3646 katahira_masashi
+3647 mizumoto_tadashi
+3648 ruu_(tksymkw)
+3649 suzuhira_hiro
+3650 mizuhara_aki
+3651 mochi_au_lait
+3652 as109
+3653 saku_usako_(rabbit)
+3654 ishikei
+3655 dei_shirou
+3656 yuureidoushi_(yuurei6214)
+3657 takeuchi_takashi
+3658 mikage_takashi
+3659 karaagetarou
+3660 nishi_koutarou
+3661 yuuji_(and)
+3662 hisona_(suaritesumi)
+3663 frapowa
+3664 eromame
+3665 onikobe_rin
+3666 bow_(bhp)
+3667 agahari
+3668 gaoo_(frpjx283)
+3669 yokochou
+3670 naga_u
+3671 homare_(fool's_art)
+3672 makuwauri
+3673 shiseki_hirame
+3674 butcha-u
+3675 taisa_(kari)
+3676 null_(nyanpyoun)
+3677 gofu
+3678 mishima_kurone
+3679 beni_shake
+3680 happoubi_jin
+3681 kure_masahiro
+3682 mattaku_mousuke
+3683 mitsumoto_jouji
+3684 clearite
+3685 koyama_shigeru
+3686 fujima_takuya
+3687 ojipon
+3688 koto_inari
+3689 dr_rex
+3690 kawashina_(momen_silicon)
+3691 milkpanda
+3692 lolicept
+3693 unya
+3694 wa_(genryusui)
+3695 boris_(noborhys)
+3696 tima
+3697 murakami_suigun
+3698 cato_(monocatienus)
+3699 bai_lao_shu
+3700 fumio_(rsqkr)
+3701 imizu_(nitro_unknown)
+3702 kagami_hirotaka
+3703 urushihara_satoshi
+3704 zen
+3705 hoshizuki_(seigetsu)
+3706 hankuri
+3707 ginhaha
+3708 yabuki_kentarou
+3709 aoshima
+3710 maturiuta_sorato
+3711 shimada_fumikane
+3712 nori_tamago
+3713 slugbox
+3714 shirosato
+3715 asanagi
+3716 pokemoa
+3717 misaki_kurehito
+3718 usashiro_mani
+3719 engiyoshi
+3720 noai_nioshi
+3721 goma_(gomasamune)
+3722 shichimenchou
+3723 eiri_(eirri)
+3724 koruri
+3725 cis_(carcharias)
+3726 pageratta
+3727 eroe
+3728 maruki_(punchiki)
+3729 meow_(nekodenki)
+3730 satou_kibi
+3731 iizuki_tasuku
+3732 kei-suwabe
+3733 catstudioinc_(punepuni)
+3734 lasterk
+3735 6_(yuchae)
+3736 toosaka_asagi
+3737 imu_sanjo
+3738 himura_kiseki
+3739 mikeou
+3740 ootsuki_wataru
+3741 ooyari_ashito
+3742 negom
+3743 lpip
+3744 ayu_(mog)
+3745 e.o.
+3746 sekina
+3747 space_jin
+3748 seo_tatsuya
+3749 wakabayashi_toshiya
+3750 yamamoto_souichirou
+3751 shirou_masamune
+3752 yuuhagi_(amaretto-no-natsu)
+3753 awa
+3754 yamashita_shun'ya
+3755 kuromiya
+3756 otohime_(youngest_princess)
+3757 yang-do
+3758 nanao_naru
+3759 masukuza_j
+3760 kashiwamochi_yomogi
+3761 ban
+3762 nagisa_kurousagi
+3763 ichikawa_feesu
+3764 jjune
+3765 tanaka_kusao
+3766 yagisaka_seto
+3767 akinbo_(hyouka_fuyou)
+3768 kara_(color)
+3769 nagian
+3770 kouu_hiyoyo
+3771 sheya
+3772 nanase_nao
+3773 eto
+3774 yume_shokunin
+3775 non_(z-art)
+3776 nishieda
+3777 komatsu_eiji
+3778 utano
+3779 ryoji_(nomura_ryouji)
+3780 isshiki_(ffmania7)
+3781 dandon_fuga
+3782 colonel_aki
+3783 pas_(paxiti)
+3784 sakurazawa_izumi
+3785 sano_toshihide
+3786 hidefu_kitayan
+3787 satou_yuuki
+3788 jin_(mugenjin)
+3789 pote_(ptkan)
+3790 chan_co
+3791 sakiyamama
+3792 setz
+3793 hase_yu
+3794 aaaa
+3795 yopparai_oni
+3796 youkan
+3797 shinama
+3798 asamura_hiori
+3799 eu03
+3800 mitsumi_misato
+3801 torinone
+3802 oouso
+3803 aono3
+3804 yoshi_tama
+3805 kazenokaze
+3806 cutesexyrobutts
+3807 7010
+3808 ucmm
+3809 minato_hitori
+3810 ikari_manatsu
+3811 agawa_ryou
+3812 fuantei
+3813 chanta_(ayatakaoisii)
+3814 yano_toshinori
+3815 swd3e2
+3816 shibasaki_shouji
+3817 oryou
+3818 aldehyde
+3819 yamato_nadeshiko
+3820 jako_(jakoo21)
+3821 rokuwata_tomoe
+3822 kenkou_cross
+3823 manji_(tenketsu)
+3824 akkijin
+3825 haruhisky
+3826 yuuki_hagure
+3827 sakuraba_yuuki
+3828 sugimura_tomokazu
+3829 lm_(legoman)
+3830 fkey
+3831 ragho_no_erika
+3832 hayashiya_zankurou
+3833 chado
+3834 aozora_market
+3835 ino
+3836 morino_hon
+3837 hemogurobin_a1c
+3838 sumiyao_(amam)
+3839 komaku_juushoku
+3840 sakazaki_freddy
+3841 onija_tarou
+3842 ghettoyouth
+3843 y.ssanoha
+3844 tsuki_wani
+3845 ichiba_youichi
+3846 urin
+3847 moneti_(daifuku)
+3848 kurimomo
+3849 gogiga_gagagigo
+3850 shiromanta
+3851 misumi_(macaroni)
+3852 namori
+3853 bosshi
+3854 mutsuki_(moonknives)
+3855 bbb_(friskuser)
+3856 cait
+3857 caffein
+3858 hairu
+3859 piromizu
+3860 sky_(freedom)
+3861 kiku_hitomoji
+3862 miuku_(marine_sapphire)
+3863 udon_(shiratama)
+3864 meito_(maze)
+3865 porurin
+3866 kusanagi_tonbo
+3867 sakino_shingetsu
+3868 momio
+3869 sutahiro_(donta)
+3870 bubukka
+3871 ogino_atsuki
+3872 nanakusa_nazuna
+3873 kamelie
+3874 niichi_(komorebi-palette)
+3875 ryouka_(suzuya)
+3876 kasuga_yukihito
+3877 mattari_yufi
+3878 hizaka
+3879 komusou_(jinrikisha)
+3880 kobuichi
+3881 umekichi
+3882 erubo
+3883 harada_takehito
+3884 m-da_s-tarou
+3885 nanaroku_(fortress76)
+3886 tetsu_(kimuchi)
+3887 robert_porter
+3888 ameyama_denshin
+3889 tenken_(gotannda)
+3890 michael
+3891 suzuki_toto
+3892 amazon_(taitaitaira)
+3893 seki_(red_shine)
+3894 moisture_(chichi)
+3895 karukan_(monjya)
+3896 petenshi_(dr._vermilion)
+3897 watarui
+3898 kouno_hikaru
+3899 shinapuu
+3900 dairi
+3901 kawata_hisashi
+3902 annin_musou
+3903 mtu_(orewamuzituda)
+3904 nigou
+3905 jetto_komusou
+3906 mogudan
+3907 kou_mashiro
+3908 obiwan
+3909 ama_mitsuki
+3910 kishida_mel
+3911 kuro_suto_sukii
+3912 shuugetsu_karasu
+3913 mery_(apfl0515)
+3914 herada_mitsuru
+3915 9law
+3916 azasuke
+3917 ryuun_(stiil)
+3918 sousou_(sousouworks)
+3919 tk8d32
+3920 donguri_suzume
+3921 waero
+3922 kusaka_souji
+3923 takayaki
+3924 magukappu
+3925 nogi_takayoshi
+3926 saki_chisuzu
+3927 774_(nanashi)
+3928 at_classics
+3929 mosha
+3930 hiro_(hirohiro31)
+3931 amazuyu_tatsuki
+3932 murata_renji
+3933 optionaltypo
+3934 paseri
+3935 tsukumo
+3936 mosho
+3937 celebi_ryousangata
+3938 matarou_(genkai_toppa)
+3939 mudou_eichi
+3940 shirabi
+3941 ishii_hisao
+3942 yamaarashi
+3943 hera_(hara0742)
+3944 yukataro
+3945 muririn
+3946 kaga3chi
+3947 kfr
+3948 nekotoufu
+3949 swordsouls
+3950 jeno
+3951 yume_no_owari
+3952 marshmallow_mille
+3953 yuzuna99
+3954 untsue
+3955 satou_shouji
+3956 komeshiro_kasu
+3957 coffee-kizoku
+3958 saru
+3959 nardack
+3960 tasaka_shinnosuke
+3961 hirano_katsuyuki
+3962 gomennasai
+3963 jason_(kaiten_kussaku_kikou)
+3964 inu_(aerodog)
+3965 yanagi_(nurikoboshi)
+3966 gurande_(g-size)
+3967 neko_majin
+3968 nora_higuma
+3969 yamu_(reverse_noise)
+3970 jizeru_(giselebon)
+3971 sakimori_(hououbds)
+3972 cool-kyou_shinja
+3973 sakimichan
+3974 culter
+3975 gochou_(atemonai_heya)
+3976 akairiot
+3977 tsuji_santa
+3978 watanabe_akio
+3979 sako_(bosscoffee)
+3980 riichu
+3981 akagashi_hagane
+3982 deco_(geigeki_honey)
+3983 mamuru
+3984 kloah
+3985 minamura_haruki
+3986 shinoasa
+3987 osashin_(osada)
+3988 danbo_(rock_clime)
+3989 totokichi
+3990 mo_(kireinamo)
+3991 socha
+3992 sakimiya_(inschool)
+3993 juurouta
+3994 ishikkoro
+3995 eretto
+3996 kaiga
+3997 natsu_(anta_tte_hitoha)
+3998 futatsuki_hisame
+3999 bococho
+4000 umigarasu_(kitsune1963)
+4001 blew_andwhite
+4002 shangguan_feiying
+4003 rifyu
+4004 yuzu_momo
+4005 satou_daiji
+4006 mikagami_hiyori
+4007 nironiro
+4008 shinshin
+4009 caryo
+4010 body_mahattaya_ginga
+4011 wata_do_chinkuru
+4012 kotoba_noriaki
+4013 kisetsu
+4014 anti_(untea9)
+4015 joy_ride
+4016 nekobungi_sumire
+4017 kuro_chairo_no_neko
+4018 ikeuchi_tanuma
+4019 kuavera
+4020 uni_mate
+4021 kichihachi
+4022 nameo_(judgemasterkou)
+4023 riyo_(lyomsnpmp)
+4024 diesel-turbo
+4025 abe_kanari
+4026 (ysy)s
+4027 kaisen_chuui
+4028 gensoukoumuten
+4029 goma_(yoku_yatta_hou_jane)
+4030 chita_(ketchup)
+4031 n.g.
+4032 mota
+4033 hana_kazari
+4034 nishimata_aoi
+4035 zuwai_kani
+4036 otokuyou
+4037 asutora
+4038 em
+4039 ryuuichi_(f_dragon)
+4040 hayashi_custom
+4041 kamizono_(spookyhouse)
+4042 shikushiku_(amamori_weekly)
+4043 hetza_(hellshock)
+4044 arinu
+4045 souji
+4046 ririko_(zhuoyandesailaer)
+4047 s-syogo
+4048 goto_p
+4049 uni8
+4050 yuki_arare
+4051 michii_yuuki
+4052 aki99
+4053 fuuma_nagi
+4054 sazaki_ichiri
+4055 senomoto_hisashi
+4056 mokufuu
+4057 neocoill
+4058 ramchi
+4059 daiaru
+4060 aoi_(annbi)
+4061 oda_ken'ichi
+4062 lack
+4063 okina_ika
+4064 kuro_(kuronell)
+4065 ominaeshi_(takenoko)
+4066 sousouman
+4067 marumikan
+4068 masakichi_(crossroad)
+4069 konpeto
+4070 givuchoko
+4071 shin_(new)
+4072 jabara_tornado
+4073 j7w
+4074 kazuma_muramasa
+4075 kumiko_shiba
+4076 koyorin
+4077 asteroid_ill
+4078 bekkankou
+4079 mikazuki_akira!
+4080 redrop
+4081 yoshitomi_akihito
+4082 kurou_(quadruple_zero)
+4083 takana_shinno
+4084 dr.p
+4085 harigane_shinshi
+4086 yutakasan-love
+4087 otoi_rekomaru
+4088 boushi-ya
+4089 homura_subaru
+4090 tebi_(tbd11)
+4091 pekeko_(pepekekeko)
+4092 kuroboshi_kouhaku
+4093 aikawa_ryou
+4094 makuran
+4095 fuukadia_(narcolepsy)
diff --git a/pycls/__init__.py b/pycls/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/pycls/__pycache__/__init__.cpython-37.pyc b/pycls/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..237625bf7dcbb4ece8a71c0b24c0588a5f8c5203
GIT binary patch
literal 132
zcmZ?b<>g`k0^5egaUl9Jh=2h`Aj1KOi&=m~3PUi1CZpd<h9ZzKg7~GYpOK%Ns$W%^
zl9E?hte=utl31KtqF+#%oKp-X<Kr{)GE3s)^$IF)aoFVMr<CTT+JTJu48#lo9El!D

literal 0
HcmV?d00001

diff --git a/pycls/core/__init__.py b/pycls/core/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/pycls/core/__pycache__/__init__.cpython-37.pyc b/pycls/core/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..269fa59988ec4e7d2754c11d25bc784a06c58e70
GIT binary patch
literal 137
zcmZ?b<>g`k0^5egaUl9Jh=2h`Aj1KOi&=m~3PUi1CZpd<h9ZzKg7{^mpOK%Ns$W%^
zl9E?hte=utl31KtqF+#%oKp-Xlk<yG_2c6+^D;}~<Mj$EZ*kb<=BJeAq}qYZ`V7Pj
E0L2X;CIA2c

literal 0
HcmV?d00001

diff --git a/pycls/core/__pycache__/config.cpython-37.pyc b/pycls/core/__pycache__/config.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e3d41535706fe7e02e106860ab77a7d2ee8c58df
GIT binary patch
literal 4257
zcmb6cOK{uP5dc9Fq$K($+i?;ra2(lYY)N($+ew<(f<Rb;O#(CkM%HVaSBMXiAd>|9
z0NEqfnaR{S<kD$wo%WPWPszQ9-h1ewo$2(5Lnbrnp`6K~xpgu<bl(Fct8F^dLWAA6
z`*!#3?%Q2Z%TrTT0iJh0`1N~9T@b!v#Qsx&#S*;n2v$Tu0uhPS6OjnG(3AQFR1g_Y
z?#cZkDv3f+dMcoDP(}({1!Y_z#h|=X!Brx1JV8nvPZE#<{!}hkBMQ(pe1uduK1!+_
zA0rbSAJ64ZkV&9V;At|&@yVP%wFRdE|8$O@%kh~UKcC|lNR7+ANRDuPHn%%NjsiWC
z+nvED$uUl!BF8y?iR;baOm7CC<#Ej53GT-XR>=u2caBVR{4zPo@p*EJ<5%)?X1V+<
zKFR%=&HbIlnZL960{3ecpUw9(|7Y<<avJQ+;!ET?j_WyoHMe`2oB{r2e1$yE@f`Q}
zGM*<daC#x9ujcp~*T0Nk%k>wz-XeaTya@7(_<By?;C?LPn>qakIm`FoBr_bp#qBKO
z@9}$D#BY<AIRBko{uaOYMSMFyXNlLt64r<ca!Yuboa0z0FLT`B_rHXV9N*z~m+(ri
zx0>5AIc<WS^Pp#f9p;ybb@B?wjeLJ!{w|pX{#|^J+q;XKxt$iLTiD8Jo6|OK=k$G^
zulv~He%=TB7r@Sa?2?Ndd(2PVA(sHYOX>jEIOhR*71rylN-!dq0e+uc0r&$l2k?hv
z9^iZACRrd?KNJxr*T`!h3g|~<k-QGrCvTAJ<Ob|)5XlwlZ;od5Xs{V>?NUDpqk*~^
z_5$_NFnS!&fT$aL>bj=8_4zStT&pgom9F324)8AR0a4aBx2%W+sW4`(o=$aX6vsFi
zBs3cCp(7i99CmR+{Xx7L(f(Kjq45#`4^0evemKAzy{P+etOD!K765qubQ{&;us`et
zI8K5-4hLwO1~J}<lB5?5g6_jyNdn<qXL>1L$V=m+&)9#v9SX$*Ay|B~{nbam`TJjP
zZT;$(?D;M1%iwmQ8^z<_Grd3lfjwW}-ulO9-~73?v2;5vg7@KI`~mEazXXsL9{Iho
z0FjJART#5DHl6`6k{6Qx@Y{(lPv`ue-|wZRI3dw)lFEZ_x0hBRR=>9$#YuSVA}E|;
z3Jb%%ZZDoA(c?ic^2r=eZ<i&CJq)vBba>;h0VKkXxFb<D!$dp-L-L^%t4V=KPsN{#
zz?XN5r0}p5-vXZeNh#xeDjw{Xfuo$t^d5jrG2f}=dXIot;(EY2(5vhm*83P_%DG;3
zXM%ptY!U@L`5bmC`OYM%f*-#F&IGqNd0=g7r$)ax(D?!+CYjC?f&Mi)@)W`*Q&60>
zk*aUQg~w`^R25Rc8^mfjc;wU2A0+WSWf>Zs^R@%1TpA~8kE&2ML$$vPM4b3EQT;@{
zS})KVNYY~fQaK6YWHjTkn^ogsFHF>nKY9P6I*34JcQ}k_5|Gi&Y`-54!nqAU>29lW
zI11EwJKF6*X9TJbhxC#VDpid(Rjbp&EA5UuKRWAy`TuVB2H*a_&C)W9^m>~H-7pT+
zWs60Zu+;R2yRc%-%K~^5&~P(^-i3}+Z>iT*aH=lR(>!q2-J_SA(UwZU$8eCTeMj&&
z_Rk5E0IDtQ2mjMSR{dxyVLdI{ZO?3_Mb|UhZd%jgIH1f!gVKnmMbFVp3wGO0(@RTi
zH26tcUDiB(6}u)fQe{7zsq7i9mrBc4daUCbxa}Ca3E1^C&rOx4y@DG|pu#pVoN1(G
z!_t<UMkB3dqpG{P#l}@!7g6<Kc*UOy^a`Zp+p7!PQ9oE1?GZBAix-HW_;HYAU75pf
zM1zIAt%rN*w9li#{(!)6+6*X$69TBi_!TfH6rLALVog3OO+UkYJz1!6EQ@p=<ho3F
z|6x;v_c-$@5fbR}odQ7lsl?W}z<I?(ydo|T5qd-#{kW0O4z=SnS#zifbVKIv6_rQK
zIDM6mpWfaT6+BmI@R-8gs^GhtL~7rE7*thwcA>VQnwg1QJ~ILJ1hq5IWvqIMvgu1-
zWIzRwme&k(WfiWZ*09&CrmZ!wuIZ}=on;~{szYFok(U8{#InZbn;=h%oO6&BX1@Y2
z&&nIHfN3FqA|-_#8Q$Wj()&VEdID4Isr2*X?~A}IKY^c+KNCe^O(2E)f+IWt3>lI~
z^~P?02s82^Ei7V2!EA=!QAr3*l|*!JJ{2FsXAQ<S4g9`US5hg8Q+eno+o`-0g##XO
zTDChLZkSH0=y&q^lcQlUNM!;C)umKnHOKOmiktrycd9TMRx%&6;7bshARQMcMfx_Z
zx-4w=u!1-k{4C8`_)ml<B7_cU_(XmpeG0+mMKITdkH^8B+@w(-1`tdNJ_-IKOsig`
zi(r(#&cJmB*cs`Oen7VZ%+$I2^#Uq`GKK^p9uu1y^era;>`K{frV9Ibv8aA=aHUdB
zq_<$NE_!v5o?@$RTE>_}U5xqhsK!5j^Sepdi|50L7ZsY=^SkkUR=<?x0ZF=*`B#7g
z(qhYQ7)>O5>um!S8%En(MH6gZ=uORa4L%iI7vJevsN@)~Wq7Cp1INl(RX$l9yVG{j
zku|g7t>U)f;1$L}Ro8ROhJmj=xQ0rawa#^6hP5z1ti+dVsJh&=^?NywJK&-cx8<U8
zX2M0vvhCqkrwJ==9Iv^k2$RdUkPOIS&aJQmz$!ZwNo!Z()!=QRa;}2PxfLWC7OEN#
z+St)N(}r{I8IEOu_0p;V&9;V=hMw)AsfJ^>S&4E{dV!^|0CN}>T6T-VFAo8oU|Ddr
zR=a7q`!f}tM22Hy(=t7L$F#I2UeQ`D4V`9lSMRu<-NIRy!^qXOb#&r8nDi_&UtZQg
zr|F<d)4^@KY3l2!$gNTKQwf!U4XJ^PX=x9z(YEzfq=0jV<#k%9(g4Q{%d}RIV!?0+
zuiizI|8n%yn&z}RZG1=5J=;P1B)8=9iG<1x%>yqydJayc=lQWF1Pz}TG-138;laI|
zb^{%QF{R&Y+ot91vnI5TXJaUyPRpPROrMQ0G{H<_Hrxy+E4ek>fqQUKwQX9sWwdN(
z9kOh+(4=EDJCMtEQ`3#KsCOEcl~$GwOJ8ki&OHkM4-52FZZr?Ygg7ByxEHpf>{~)_
zFmMw<syNmPb`7IJFEOPm`|#!dPc}@~!^@g}&#)RuUbS5h$!**5=vAi9zBPQb@&Us;
z@=){gHqJWu?S4ddd%-OQ6yh%d$dV#Wij#6lmP<+@|NE}h{vN|cctMZ<iHFvb@=Wuv
w{63e-78H@|AKrhqW|EoIKJPFmv%_EsY-ALZQ%VK+lR0E7k*%ud6-6ok6F4tY2LJ#7

literal 0
HcmV?d00001

diff --git a/pycls/core/__pycache__/losses.cpython-37.pyc b/pycls/core/__pycache__/losses.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e890df87f632893a3c0bb96fd954436b8cc4f8c5
GIT binary patch
literal 755
zcmY*X!EO^V5VgI|W)mgQ_6Fjx;y{(lh65+mB1#DoQb?r)E<UhYt(|1k-F0L;NE?cT
zC?ArD6TiYI_{xc2;KX>-mUg5Wk7ssg-n{XvZnup<j=%o+_=_O)+g+{&?&u|`eh!8s
zhB)RV#|2>oJKxK_f-(v|6=c!KF!OnX``;03@+J@93^<vfUTb}SD76l!t1>m2EOqn;
z3ZhXDTQ8l?ST|LWjzwvdTvn;eT?+;06jVP0W6%OGi19Xbi_U%h*fcoa_z-O_f-Szp
z->~s^-#kE<Xw%vvXinCT#=@vfd=fe|a}nmZp+xKO4z%#_^OwW0lqS@x<x(mmc(1J-
z5v=j<?Db&Cg5lX@62JK{vhCQFieV~k>A9%1y?3IuP;SJ5Qc_ueDpipf)_6ZYc{^g%
z{c35t;>k~>0w`LWzfb#fS&06+;=HVMpC=~K!t{ZPTwmE#D$#fCgpQV#?aYL^>3-n&
z@xdVP0PkVt7MeQ9PTc8%)+BboxxfZppd;8{P0o|Cm}S}srNiVdH6gFcq{z}F&#MS_
zrCgaFu@s_$(bP!we+Jcpq}^9n#cr@Ygs`r$5aqh}TDBR-T&8i%0;rJAqq4LFbiW=q
za)W{?r`c>r``)R09E084aJ#B;)I-)4Z(*nzt$_qbh2*PTyl^4y2%)5dJGAWu!9SFX
B!g>Gz

literal 0
HcmV?d00001

diff --git a/pycls/core/__pycache__/model_builder.cpython-37.pyc b/pycls/core/__pycache__/model_builder.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..32012be5eee30e1a9ff81c22790510f93ec8ede5
GIT binary patch
literal 1317
zcmZWp&5qkP5GJV~%d)NQ?lzkSZ3>~6z&+T#6)4goKfCCqyInNRp*mF1inLZ<N-{_)
zfLw17Bt>s|gC2I#Q(uYKp7IJkbx7Gok<uV%G{YIrd^6)GlSznRef{c>pFi>u`pbOm
zmIuru*zz|xC`u5;lyF=Tf-tI_#KpbDOFYmx+^+`7z{*@6RKsLwd5?$HC>dGa=i_RU
zOe{a(QMH%sVZ_{<57_=KP7c_Rjo2QW(BOABIiy1x{)&<h>4=Vj-l60fnoidDPaDd3
zoHw;pqRmy=)bXONEhA^|AXIcV#oEai1rYyn-JLOIxRb?V%l*KLe>rJ5H}SK&tC>oC
zwpdsynB39=Xl~q^MPM6Tp2C)|;ZSIamqa-??k&3V<!1`8xbb`GEq;SvW8eoWSPpMO
zxPj$YaD!O6)8D#>-k@t_Z;WoyHQu1*a07V0#UKS+>R!3xCFsY?$uioY4K6U~y|>Oa
z+91HhU++I@OHO61E?Ha|cxF48RJUUBgI}&6#C4-$*{)WNP>ilWf11@s(n@0?Sp4nz
z^O&+1WzLM7td6r6S;?~t&J1?F<{gHFr;hgFjg~VV_TsdpQcs#zty-1#qUoM6NNMl%
z`{R@EwEN>v=O=oQ+SsIauUIE_^jJzJO#M!TXoU6`ji@rEgR|$SDTMP(d#VxnrFQd{
zW_rAREN|Lc=_qf7VD;`jP3l?)t4w5^Gp_F*m(U#LMGL7OXDU12DyBqxHijew`)#a$
zBJnZA`uCUf%cf%Ubw_F4$vFi($<%z+<y>yh9FUq@2kC`DoQc`0(<96G$=w4_@(2z@
zj_}wyz&?pcNXD=O91(FJG`WGsEJGdGwwcDjY@k^-=pl?1l$ESMg+_%$n2>SS*EObH
zomFKHX?8Pca}j|RWaWaC0Z;&#O{PCS0i}+_)=3}W7a)^103u>zd<1n#X_Kd^KD4>c
zT2*p6<4sYNb)h}`sRz4n!dr1@9PjtH4JI>~#zk4!nkQQ?eXVExTnUrrdzyaIg!_LD
zL<}=0%$&Vv(@&*<e=4NIf~nL*#e@#h)F1#k_`SpeW6`!*{s$COYhjTI9zu-J$KM1X
UvPE;r0S@qZ_j5uga)P`60RCHZGXMYp

literal 0
HcmV?d00001

diff --git a/pycls/core/config.py b/pycls/core/config.py
new file mode 100644
index 0000000..2c43c96
--- /dev/null
+++ b/pycls/core/config.py
@@ -0,0 +1,410 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Configuration file (powered by YACS)."""
+
+import os
+
+from pycls.utils.io import cache_url
+from yacs.config import CfgNode as CN
+
+
+# Global config object
+_C = CN()
+
+# Example usage:
+#   from core.config import cfg
+cfg = _C
+
+
+# ---------------------------------------------------------------------------- #
+# Model options
+# ---------------------------------------------------------------------------- #
+_C.MODEL = CN()
+
+# Model type
+_C.MODEL.TYPE = ""
+
+# Number of weight layers
+_C.MODEL.DEPTH = 0
+
+# Number of classes
+_C.MODEL.NUM_CLASSES = 10
+
+# Loss function (see pycls/models/loss.py for options)
+_C.MODEL.LOSS_FUN = "cross_entropy"
+
+
+# ---------------------------------------------------------------------------- #
+# ResNet options
+# ---------------------------------------------------------------------------- #
+_C.RESNET = CN()
+
+# Transformation function (see pycls/models/resnet.py for options)
+_C.RESNET.TRANS_FUN = "basic_transform"
+
+# Number of groups to use (1 -> ResNet; > 1 -> ResNeXt)
+_C.RESNET.NUM_GROUPS = 1
+
+# Width of each group (64 -> ResNet; 4 -> ResNeXt)
+_C.RESNET.WIDTH_PER_GROUP = 64
+
+# Apply stride to 1x1 conv (True -> MSRA; False -> fb.torch)
+_C.RESNET.STRIDE_1X1 = True
+
+
+# ---------------------------------------------------------------------------- #
+# AnyNet options
+# ---------------------------------------------------------------------------- #
+_C.ANYNET = CN()
+
+# Stem type
+_C.ANYNET.STEM_TYPE = "plain_block"
+
+# Stem width
+_C.ANYNET.STEM_W = 32
+
+# Block type
+_C.ANYNET.BLOCK_TYPE = "plain_block"
+
+# Depth for each stage (number of blocks in the stage)
+_C.ANYNET.DEPTHS = []
+
+# Width for each stage (width of each block in the stage)
+_C.ANYNET.WIDTHS = []
+
+# Strides for each stage (applies to the first block of each stage)
+_C.ANYNET.STRIDES = []
+
+# Bottleneck multipliers for each stage (applies to bottleneck block)
+_C.ANYNET.BOT_MULS = []
+
+# Group widths for each stage (applies to bottleneck block)
+_C.ANYNET.GROUP_WS = []
+
+# Whether SE is enabled for res_bottleneck_block
+_C.ANYNET.SE_ON = False
+
+# SE ratio
+_C.ANYNET.SE_R = 0.25
+
+# ---------------------------------------------------------------------------- #
+# RegNet options
+# ---------------------------------------------------------------------------- #
+_C.REGNET = CN()
+
+# Stem type
+_C.REGNET.STEM_TYPE = "simple_stem_in"
+# Stem width
+_C.REGNET.STEM_W = 32
+# Block type
+_C.REGNET.BLOCK_TYPE = "res_bottleneck_block"
+# Stride of each stage
+_C.REGNET.STRIDE = 2
+# Squeeze-and-Excitation (RegNetY)
+_C.REGNET.SE_ON = False
+_C.REGNET.SE_R = 0.25
+
+# Depth
+_C.REGNET.DEPTH = 10
+# Initial width
+_C.REGNET.W0 = 32
+# Slope
+_C.REGNET.WA = 5.0
+# Quantization
+_C.REGNET.WM = 2.5
+# Group width
+_C.REGNET.GROUP_W = 16
+# Bottleneck multiplier (bm = 1 / b from the paper)
+_C.REGNET.BOT_MUL = 1.0
+
+
+# ---------------------------------------------------------------------------- #
+# EfficientNet options
+# ---------------------------------------------------------------------------- #
+_C.EN = CN()
+
+# Stem width
+_C.EN.STEM_W = 32
+
+# Depth for each stage (number of blocks in the stage)
+_C.EN.DEPTHS = []
+
+# Width for each stage (width of each block in the stage)
+_C.EN.WIDTHS = []
+
+# Expansion ratios for MBConv blocks in each stage
+_C.EN.EXP_RATIOS = []
+
+# Squeeze-and-Excitation (SE) ratio
+_C.EN.SE_R = 0.25
+
+# Strides for each stage (applies to the first block of each stage)
+_C.EN.STRIDES = []
+
+# Kernel sizes for each stage
+_C.EN.KERNELS = []
+
+# Head width
+_C.EN.HEAD_W = 1280
+
+# Drop connect ratio
+_C.EN.DC_RATIO = 0.0
+
+# Dropout ratio
+_C.EN.DROPOUT_RATIO = 0.0
+
+
+# ---------------------------------------------------------------------------- #
+# Batch norm options
+# ---------------------------------------------------------------------------- #
+_C.BN = CN()
+
+# BN epsilon
+_C.BN.EPS = 1e-5
+
+# BN momentum (BN momentum in PyTorch = 1 - BN momentum in Caffe2)
+_C.BN.MOM = 0.1
+
+# Precise BN stats
+_C.BN.USE_PRECISE_STATS = False
+_C.BN.NUM_SAMPLES_PRECISE = 1024
+
+# Initialize the gamma of the final BN of each block to zero
+_C.BN.ZERO_INIT_FINAL_GAMMA = False
+
+# Use a different weight decay for BN layers
+_C.BN.USE_CUSTOM_WEIGHT_DECAY = False
+_C.BN.CUSTOM_WEIGHT_DECAY = 0.0
+
+# ---------------------------------------------------------------------------- #
+# Optimizer options
+# ---------------------------------------------------------------------------- #
+_C.OPTIM = CN()
+
+# Base learning rate
+_C.OPTIM.BASE_LR = 0.1
+
+# Learning rate policy select from {'cos', 'exp', 'steps'}
+_C.OPTIM.LR_POLICY = "cos"
+
+# Exponential decay factor
+_C.OPTIM.GAMMA = 0.1
+
+# Steps for 'steps' policy (in epochs)
+_C.OPTIM.STEPS = []
+
+# Learning rate multiplier for 'steps' policy
+_C.OPTIM.LR_MULT = 0.1
+
+# Maximal number of epochs
+_C.OPTIM.MAX_EPOCH = 200
+
+# Momentum
+_C.OPTIM.MOMENTUM = 0.9
+
+# Momentum dampening
+_C.OPTIM.DAMPENING = 0.0
+
+# Nesterov momentum
+_C.OPTIM.NESTEROV = True
+
+# L2 regularization
+_C.OPTIM.WEIGHT_DECAY = 5e-4
+
+# Start the warm up from OPTIM.BASE_LR * OPTIM.WARMUP_FACTOR
+_C.OPTIM.WARMUP_FACTOR = 0.1
+
+# Gradually warm up the OPTIM.BASE_LR over this number of epochs
+_C.OPTIM.WARMUP_EPOCHS = 0
+
+
+# ---------------------------------------------------------------------------- #
+# Training options
+# ---------------------------------------------------------------------------- #
+_C.TRAIN = CN()
+
+# Dataset and split
+_C.TRAIN.DATASET = ""
+_C.TRAIN.SPLIT = "train"
+
+# Total mini-batch size
+_C.TRAIN.BATCH_SIZE = 128
+
+# Image size
+_C.TRAIN.IM_SIZE = 224
+
+# Evaluate model on test data every eval period epochs
+_C.TRAIN.EVAL_PERIOD = 1
+
+# Save model checkpoint every checkpoint period epochs
+_C.TRAIN.CHECKPOINT_PERIOD = 1
+
+# Resume training from the latest checkpoint in the output directory
+_C.TRAIN.AUTO_RESUME = True
+
+# Weights to start training from
+_C.TRAIN.WEIGHTS = ""
+
+
+# ---------------------------------------------------------------------------- #
+# Testing options
+# ---------------------------------------------------------------------------- #
+_C.TEST = CN()
+
+# Dataset and split
+_C.TEST.DATASET = ""
+_C.TEST.SPLIT = "val"
+
+# Total mini-batch size
+_C.TEST.BATCH_SIZE = 200
+
+# Image size
+_C.TEST.IM_SIZE = 256
+
+# Weights to use for testing
+_C.TEST.WEIGHTS = ""
+
+
+# ---------------------------------------------------------------------------- #
+# Common train/test data loader options
+# ---------------------------------------------------------------------------- #
+_C.DATA_LOADER = CN()
+
+# Number of data loader workers per training process
+_C.DATA_LOADER.NUM_WORKERS = 4
+
+# Load data to pinned host memory
+_C.DATA_LOADER.PIN_MEMORY = True
+
+
+# ---------------------------------------------------------------------------- #
+# Memory options
+# ---------------------------------------------------------------------------- #
+_C.MEM = CN()
+
+# Perform ReLU inplace
+_C.MEM.RELU_INPLACE = True
+
+
+# ---------------------------------------------------------------------------- #
+# CUDNN options
+# ---------------------------------------------------------------------------- #
+_C.CUDNN = CN()
+
+# Perform benchmarking to select the fastest CUDNN algorithms to use
+# Note that this may increase the memory usage and will likely not result
+# in overall speedups when variable size inputs are used (e.g. COCO training)
+_C.CUDNN.BENCHMARK = True
+
+
+# ---------------------------------------------------------------------------- #
+# Precise timing options
+# ---------------------------------------------------------------------------- #
+_C.PREC_TIME = CN()
+
+# Perform precise timing at the start of training
+_C.PREC_TIME.ENABLED = False
+
+# Total mini-batch size
+_C.PREC_TIME.BATCH_SIZE = 128
+
+# Number of iterations to warm up the caches
+_C.PREC_TIME.WARMUP_ITER = 3
+
+# Number of iterations to compute avg time
+_C.PREC_TIME.NUM_ITER = 30
+
+
+# ---------------------------------------------------------------------------- #
+# Misc options
+# ---------------------------------------------------------------------------- #
+
+# Number of GPUs to use (applies to both training and testing)
+_C.NUM_GPUS = 1
+
+# Output directory
+_C.OUT_DIR = "/tmp"
+
+# Config destination (in OUT_DIR)
+_C.CFG_DEST = "config.yaml"
+
+# Note that non-determinism may still be present due to non-deterministic
+# operator implementations in GPU operator libraries
+_C.RNG_SEED = 1
+
+# Log destination ('stdout' or 'file')
+_C.LOG_DEST = "stdout"
+
+# Log period in iters
+_C.LOG_PERIOD = 10
+
+# Distributed backend
+_C.DIST_BACKEND = "nccl"
+
+# Hostname and port for initializing multi-process groups
+_C.HOST = "localhost"
+_C.PORT = 10001
+
+# Models weights referred to by URL are downloaded to this local cache
+_C.DOWNLOAD_CACHE = "/tmp/pycls-download-cache"
+
+
+def assert_and_infer_cfg(cache_urls=True):
+    """Checks config values invariants."""
+    assert (
+        not _C.OPTIM.STEPS or _C.OPTIM.STEPS[0] == 0
+    ), "The first lr step must start at 0"
+    assert _C.TRAIN.SPLIT in [
+        "train",
+        "val",
+        "test",
+    ], "Train split '{}' not supported".format(_C.TRAIN.SPLIT)
+    assert (
+        _C.TRAIN.BATCH_SIZE % _C.NUM_GPUS == 0
+    ), "Train mini-batch size should be a multiple of NUM_GPUS."
+    assert _C.TEST.SPLIT in [
+        "train",
+        "val",
+        "test",
+    ], "Test split '{}' not supported".format(_C.TEST.SPLIT)
+    assert (
+        _C.TEST.BATCH_SIZE % _C.NUM_GPUS == 0
+    ), "Test mini-batch size should be a multiple of NUM_GPUS."
+    assert (
+        not _C.BN.USE_PRECISE_STATS or _C.NUM_GPUS == 1
+    ), "Precise BN stats computation not verified for > 1 GPU"
+    assert _C.LOG_DEST in [
+        "stdout",
+        "file",
+    ], "Log destination '{}' not supported".format(_C.LOG_DEST)
+    assert (
+        not _C.PREC_TIME.ENABLED or _C.NUM_GPUS == 1
+    ), "Precise iter time computation not verified for > 1 GPU"
+    if cache_urls:
+        cache_cfg_urls()
+
+
+def cache_cfg_urls():
+    """Download URLs in the config, cache them locally, and rewrite cfg to make
+    use of the locally cached file.
+    """
+    _C.TRAIN.WEIGHTS = cache_url(_C.TRAIN.WEIGHTS, _C.DOWNLOAD_CACHE)
+    _C.TEST.WEIGHTS = cache_url(_C.TEST.WEIGHTS, _C.DOWNLOAD_CACHE)
+
+
+def dump_cfg():
+    """Dumps the config to the output directory."""
+    cfg_file = os.path.join(_C.OUT_DIR, _C.CFG_DEST)
+    with open(cfg_file, "w") as f:
+        _C.dump(stream=f)
+
+
+def load_cfg(out_dir, cfg_dest="config.yaml"):
+    """Loads config from specified output directory."""
+    cfg_file = os.path.join(out_dir, cfg_dest)
+    _C.merge_from_file(cfg_file)
diff --git a/pycls/core/losses.py b/pycls/core/losses.py
new file mode 100644
index 0000000..feea0b6
--- /dev/null
+++ b/pycls/core/losses.py
@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Loss functions."""
+
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+# Supported loss functions
+_loss_funs = {"cross_entropy": nn.CrossEntropyLoss}
+
+
+def get_loss_fun():
+    """Retrieves the loss function."""
+    assert (
+        cfg.MODEL.LOSS_FUN in _loss_funs.keys()
+    ), "Loss function '{}' not supported".format(cfg.TRAIN.LOSS)
+    return _loss_funs[cfg.MODEL.LOSS_FUN]().cuda()
+
+
+def register_loss_fun(name, ctor):
+    """Registers a loss function dynamically."""
+    _loss_funs[name] = ctor
diff --git a/pycls/core/model_builder.py b/pycls/core/model_builder.py
new file mode 100644
index 0000000..9263d3c
--- /dev/null
+++ b/pycls/core/model_builder.py
@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Model construction functions."""
+
+import pycls.utils.logging as lu
+import torch
+from pycls.core.config import cfg
+from pycls.models.anynet import AnyNet
+from pycls.models.effnet import EffNet
+from pycls.models.regnet import RegNet
+from pycls.models.resnet import ResNet
+
+
+logger = lu.get_logger(__name__)
+
+# Supported models
+_models = {"anynet": AnyNet, "effnet": EffNet, "resnet": ResNet, "regnet": RegNet}
+
+
+def build_model():
+    """Builds the model."""
+    assert cfg.MODEL.TYPE in _models.keys(), "Model type '{}' not supported".format(
+        cfg.MODEL.TYPE
+    )
+    assert (
+        cfg.NUM_GPUS <= torch.cuda.device_count()
+    ), "Cannot use more GPU devices than available"
+    # Construct the model
+    model = _models[cfg.MODEL.TYPE]()
+    # Determine the GPU used by the current process
+    cur_device = torch.cuda.current_device()
+    # Transfer the model to the current GPU device
+    model = model.cuda(device=cur_device)
+    # Use multi-process data parallel model in the multi-gpu setting
+    if cfg.NUM_GPUS > 1:
+        # Make model replica operate on the current device
+        model = torch.nn.parallel.DistributedDataParallel(
+            module=model, device_ids=[cur_device], output_device=cur_device
+        )
+    return model
+
+
+def register_model(name, ctor):
+    """Registers a model dynamically."""
+    _models[name] = ctor
diff --git a/pycls/core/optimizer.py b/pycls/core/optimizer.py
new file mode 100644
index 0000000..bbba1cc
--- /dev/null
+++ b/pycls/core/optimizer.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Optimizer."""
+
+import pycls.utils.lr_policy as lr_policy
+import torch
+from pycls.core.config import cfg
+
+
+def construct_optimizer(model):
+    """Constructs the optimizer.
+
+    Note that the momentum update in PyTorch differs from the one in Caffe2.
+    In particular,
+
+        Caffe2:
+            V := mu * V + lr * g
+            p := p - V
+
+        PyTorch:
+            V := mu * V + g
+            p := p - lr * V
+
+    where V is the velocity, mu is the momentum factor, lr is the learning rate,
+    g is the gradient and p are the parameters.
+
+    Since V is defined independently of the learning rate in PyTorch,
+    when the learning rate is changed there is no need to perform the
+    momentum correction by scaling V (unlike in the Caffe2 case).
+    """
+    # Batchnorm parameters.
+    bn_params = []
+    # Non-batchnorm parameters.
+    non_bn_parameters = []
+    for name, p in model.named_parameters():
+        if "bn" in name:
+            bn_params.append(p)
+        else:
+            non_bn_parameters.append(p)
+    # Apply different weight decay to Batchnorm and non-batchnorm parameters.
+    bn_weight_decay = (
+        cfg.BN.CUSTOM_WEIGHT_DECAY
+        if cfg.BN.USE_CUSTOM_WEIGHT_DECAY
+        else cfg.OPTIM.WEIGHT_DECAY
+    )
+    optim_params = [
+        {"params": bn_params, "weight_decay": bn_weight_decay},
+        {"params": non_bn_parameters, "weight_decay": cfg.OPTIM.WEIGHT_DECAY},
+    ]
+    # Check all parameters will be passed into optimizer.
+    assert len(list(model.parameters())) == len(non_bn_parameters) + len(
+        bn_params
+    ), "parameter size does not match: {} + {} != {}".format(
+        len(non_bn_parameters), len(bn_params), len(list(model.parameters()))
+    )
+    return torch.optim.SGD(
+        optim_params,
+        lr=cfg.OPTIM.BASE_LR,
+        momentum=cfg.OPTIM.MOMENTUM,
+        weight_decay=cfg.OPTIM.WEIGHT_DECAY,
+        dampening=cfg.OPTIM.DAMPENING,
+        nesterov=cfg.OPTIM.NESTEROV,
+    )
+
+
+def get_epoch_lr(cur_epoch):
+    """Retrieves the lr for the given epoch (as specified by the lr policy)."""
+    return lr_policy.get_epoch_lr(cur_epoch)
+
+
+def set_lr(optimizer, new_lr):
+    """Sets the optimizer lr to the specified value."""
+    for param_group in optimizer.param_groups:
+        param_group["lr"] = new_lr
diff --git a/pycls/datasets/__init__.py b/pycls/datasets/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/pycls/datasets/__pycache__/__init__.cpython-37.pyc b/pycls/datasets/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..dd79509453e1c283e20de5219024755628ef1db9
GIT binary patch
literal 141
zcmZ?b<>g`k0^5egaUl9Jh=2h`Aj1KOi&=m~3PUi1CZpd<h9ZzKg7{^opOK%Ns$W%^
zl9E?hte=utl31KtqF+#%oKp-Xp<>1Q@$s2?nI-Y@dIgoYIBatBQ%ZAE?Lg*z24V&P
DS@0p+

literal 0
HcmV?d00001

diff --git a/pycls/datasets/__pycache__/cifar10.cpython-37.pyc b/pycls/datasets/__pycache__/cifar10.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b7da02c70520c1b3b04411ca9949d6c16483e5db
GIT binary patch
literal 2902
zcmZWrTaO#J6(%{?9ciz=#OoHR5}*kLwAyuUMcs|Nj@Ky+#GBMxAjQDRbUYksM;^_L
zNvgH95)fql8WaWk(mqM(Q-S_~JoX2)KcIXnkf;3%d1=q#NV^+{0uK*4hveaRZu($#
zH6YO5|Ke|-{|C5laI$PRF#FKupMh`^5>7eGX-Lh;LS{xQw9V)YTt+Avc%esYNq?|n
zR)VmxSZRh$BfDxwD_<L|hwGH2joC)JIj7+kx4FaJo)uo<9{2eQa94jtc)%Mk2ydj#
z{!0H^|N5yFzQWlNX*b6k_a1zB_tCB07LQddQ`I>GUCCjaYAfmXdT(PpiuQZBO8)uz
ze($e;fi`-)Soz}P{ocR-1+BSc{_)9vVuTlj!MG1yB1s86p5Vhni!*M$An<rl8Aec>
zdoa4FHjG}abm;RYUxoP!Ujv2LwRbP;#^ToQc++V9(@Ux+7F>Z1M+f>n=<;_!6zS6$
zo6tFxH<iWd%tro{en}^cv!|}yg;fhyS-q;9nak~ZKBtsS=!8VxgaP#@WVSLVkC6|c
zH#h;koWt>r13<7<#p<~A=E?NUR#B=}w;UF{&9tLLoE6$usZ`_FkE%RV7eaDaRi#iV
z2XWC#@Ok4O+$#&I#4u4=(QBQ*^kmxUblOeLN~vwUO1tUvOe(3@?n;>om6gSPAxfcL
zIBO6q?dD~#mkMoXMYq&Jgl3NFosmR6U@b3WZa!!)%8F_T$9Yj6KS^`hwzMtNysK9(
zHaceC--);e$bS3o&T%<Ncg7>mi;>(}xY15EN^-fNi#54p-gaxZQ;oDA!Bv@xq937+
z)_@Rtlde&pZBUoGHaKQ&=gv|vK~KVlVYLF?Soa-Z6xfH;Q})NJQ>JL2p-s8<6`NY0
zLfHM9JR-j(6N}s1NS$rcMb~!^H4{&?^JD}@0eAP)A}KkbE%AJ#vJ1=w3}uxT+O4wW
zX`X5uF>Bjm3q*AABv#3BjVkSG$_+9#xsC#j8e${%6IhUVGNDa+gO0D9ztfn($RrfM
zL-V1dta<9!z$h|<r-Cb!z6IT$gTsP*!VKJ)iT5}>^94~3W`R?7Ku+nD77$oH7<;5g
z7?}br&-W(ufPD7O)K+fan}JLB?FpPUb<nvd&YV7GQ+EmuKMN+Vxxt!vKO<A?*+-ue
z)tEIgP-d&D(O;X>DF8lO$FNl!{muSXO}lpeWnvk)gYmVG05WOvOe-r|gLF`eQD?ky
zGOgFpDQ~w<rULOlM*xkEp?m~gBA7jR?PqIKNFy^a`<(uXJ+r}_bi8Ujv^K@bv`qyj
zW{5r3+0`_A$t2$fdpX2SQVyy+@8QvSk4wgHeAgOyWM?lglQ@@mfLU%iLj#?mWx@bs
zIx^Ru{SVq1CZF~d(SSiTk)R#|JzT5;(T<1<h)CQlJY2B*WmbsSkagn<yo_sBRNDQ$
zRB`eY)}$I$srI_yAnNV6v?o$|99OA8owf6uM|U3_YVX0r=;%RszwH~J3>!5NMLbBQ
zb{A3Y*4`X092P%{SK*MbSqs$92%EK7J$nf_u@3ULpv#*;2)oV#dY!pI1G)k68_*`(
zf=C{3eILp52woc1xb+=42Hko>6;sySnnJGaLrDa8z72WXf!MC2nZnEkDX;JMSFmb)
zg|ZZ5dHmx~L|VlFOiLZ7t!xnYQj<-KRQ;h8Pr>0Jt#yoR)uoL5_m}7!*+A3k3_Y%G
ziFaXubeE0O(61mRBG(Rd0Li<Mm``OGBxPQTsDO!fqKEhI9%?H(Ixyx9*N;n)jmtvC
zdDP9bDr~?;UJjx}l$EiTK;*<*cpWo#GIzcNF{F#Ceod4O?t`o}e=!(>yGE~GP8k<+
z=h)&~&;<>)yTAaNCO>xBHlRGmIaI>!MJj-Rw=kG7h!>&C4m*(9Hejo^ld>9#O<=_y
zkY&PivVy12YrGueEZ=`Z)Js*S(m@nGgq@3w_bNQ$QpUTGIgfV?EOZZ}K86pf--n#H
zjb)Y{3KRhFjCziV0|<3Y!!N7gMNys>QS=DDneZlbyngPK^|wPXFH|B7geZba7!683
z%#m+K(X(NkFTTK8yiB4<yn+{DsG2x1poo2(et-lE;w6IVN}w%%2;I~;pW4h{^&S7J
z@B5DV*1jP=L^+cWP*6L?Z~&-dp$NB49qlYDXy?KewZB{xKgV5HYRRF>a@oP652Z{y
zW~#mAoaICKFH(x+SZ|o2Q_JH~db2(uDMi{z%A%X~>I@TrC82=f-w1(bsZ-l+<bHuS
w+h(_S0B67-Nb{&mAQ((c74-LNbnjqQlQ)0>7*Ozix&=SrXwwQTNF>((0l{A0I{*Lx

literal 0
HcmV?d00001

diff --git a/pycls/datasets/__pycache__/imagenet.cpython-37.pyc b/pycls/datasets/__pycache__/imagenet.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..588f332b114177d7e1a7ef0f78890a8153d1e63d
GIT binary patch
literal 3347
zcma)8U2Ggz6~1?Vc4x=C_S$hAJ1Jo$`nM{qouVRW5^r4FO^uob;gD3*CZoyR+3{><
zXV!Pd!QQL|s{4YV0;&9{2#K}ILy;=r0SUna5xgN9iMJ|OLTv%Hn0F*7FPw8{e;^cz
zS>3<0=iYnn`Mz`ReX3M434G_j^4j&QlZ5;U2g5H9oeNNs-$TWSMHppDNG&Q`#ZqLe
zTB>X{OP6h~omU7YZNoBXHlxrsrDVb?4kRV3By}fctA<nUvQ?(UEB2?onL}zFV;a+0
z&R4BjmS+YlK<~Jx_ve_&CRou|eF|D9q@=_q*_5xcGMi>I&>my6>^NwspfP`uusL?(
zCBja4rIy)RXq`Mzty4_dB-Q!u^hVq9y@r=on3Fn*m)4HpprlcyLUo(I|DDyZh426K
zUfuugtzZ4%<l-OdzWLT`o%80~b^q?0@~iuMM-QJd-gvX_f5c*E-~Qa)x-V7U`d!`s
z;;&!-VdKIpbsyv}e>dWFzxu*Ae)G3S|9Hp$$JEi!f1<u$_rHAa{N8h~!xsIo{`B3W
zZ^K^$<vV`;iR-`IDE{@1|NK+Sf4TPt`~{j%+&%l=9l!R?{@>5MC+aSmJNy*rT!4~%
z8Y+(fN(2!CbQGq(L@W(`3N84R<)O`?U!l#TPoXtf5xi0o#?UJt;vZLqweiDNhq4MK
zc@C<Sv}j-H0W8VM6bRARuzx^bq&<ZxH*?7nNL7$3nKae=Ii_XfLrO`H_K2PDDNq|d
z(k~nWg%|;Y^hS4n4dAJCoOG-5$nE_{Dp8zPn(<D=s*2E4?gWugZ#iN2q0LShq+^|A
zr_+gf>M_vfmC><It;SK3@*Ot~B7Y>V+}^L%YE@GxaUwL_Oys=XAW0K3xsoIvPlGsG
z=RD>j2g|jcROG_g_dPDOAZo_Kw2?)2Hq%bf6ftGH!@caF&2Cgxp(S3}6os*_E>(X-
z_%U$f2cKHpird~|caO1XFIgNAbFs7MhRL8F$dbhX=^FuUJA1;g?H~$L+kO}kfk-BV
zo}{K)q-C0e-x$o=P}8NT1z!$|0Y!rOXXp&TBrUZ^TH2wqPbFNmpT{u(3m{UG8X2=v
zfD6H1BL@ID0BB}ROD90vJdkpb7JUMBY4-|B*T}VJ_W|ryv8ODMeRW75pog-r?bhU)
z8m@V*vaf?)sXsZ;(R%u!(o<RP`_hW-uU{qG>?)BK=KJ|{3bhZw9%&=e%os?)mQ2bD
zKrFL6J0e0-Nw>TTB2?>^Z|}=|APM&*&|>5#j4&j9g5@++J#uXhd_(#WYv>gY0dNI~
zyA5#%J1E`4_2~JfPkiRuV-Gffs|xpo4#aV{a8O$kv?)Q)s##d@gTgsPbhpEoF^Uu1
zQ6t^QhTN^)$EVu4Fm|0VS+0%djjY8G)u3$RLUT|BXz%0;fF^hz!txbmTieWNCoQO|
zd>+&U4FtV)zhUM%7PzS!w>!&caY3{!N4g*XKP($fgj45(b`~MJ8#+nSoo(#2Z+N^C
zH^KYB<^&zyJwBF7wM^WomN`<D<Gm!bRvbjUgkxkopTMe!6<%l}m&nK$`4B>w1>7=i
zxs)BSL=?Per|oek^{nZU7+ssf-$Iw6DsoPz<3&uUp%+ziybP;ari|6>1jcYZD?gZN
z<=}L<7oIY5U<VB8F}i#ThG(E84?soK89J*dbY7iPa<oWIIw@oIv@%D#^Z(1QE;=5+
z41ht<ehUOZd$?|IDtsxWEdV%GTUtxU5Pew%lG0YLN3&}c$5ADPa0VTtr6YAwqhHA8
zWaA#~o5RcZknU5MZJr{~4)=KxW=w!|A}yw+LrtzQ3Hwf^<wMw^haphxKJ<C+bsX*u
zLkvKc=@|0Xmtuai5=5Pyw8oLC{_WWxUwpUwvpS?nC1^vc47#2uz+}g6a>w=Y+=RSS
z7wR)tSAnjN!bs4L&|}U#euOUJnADE2%Cz5!FZkbk6~AB9tJ9Jy!h|~`N-&8hVj8+E
zZrivOvt1r{I4&df%U@X8Xo&pAC3|zjS{GAWF%P=H5GS;oVE}t35O|)@F0XH1789-)
zr5=~_ML{m#i4#v4Vc>7!H6sf4`o=~3%F45w!|wX(rZC+&jJX|wDL}hR>njbR+M8=t
zMbb^4?>-KR3A>Xy6j=HUSejtW64g{RT2^P&GA%1*#ZXLWCuw(f%%kzN5r@#}SY+Z~
zgAS%Lz!%uF09;xifV<LDT539El7dOQpTk>RrteHs!xRp8x7x&qmEPc(9G`|~l;*~r
zJ^oP`RrBzsy7jn#5CJ$De-LC4VoAE!@q`j}M7|luPWs8mc^$Pi=>VYz5%YF4ob}MT
z;f4S04%_$AAobd|?ZC`LK4+C#IKFwG1ICJ@9Em(ZxTOPdW{P11w-_9QLdxcNn@U@&
zG=t<nBU<Eb8`6Vq-vm9G6AA5>$3D;gY*Y()KVwF)ZFncx?U?O^*e}}l_Kp({{(;3<
z?AkUzj=N&Akj~%;GvB~!yu0k-mBBI3e+){3$<LsgVoVq(j0MA#|5P8~mrzS4PP_&9
z3RFUe1PI6==3+rUE^5PPMeX0B!x#$r5^5gHG<VV<OltV}L1yc62#IzuR>B>4mBrlM
z5@p%cGJRZ1OlK>&G52baHJgE-g_{W=+Szl8e-^Ej>HIwQmT_m@;oRBdPoe<tAEC+R
t@@v_{T^{wryd#WrnX{Mifs<6A0tZn83!Q-<wnf!cO?WHKP>r6}{sj$Yjnx1E

literal 0
HcmV?d00001

diff --git a/pycls/datasets/__pycache__/loader.cpython-37.pyc b/pycls/datasets/__pycache__/loader.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..285adf4197d482f11c40db3202805720493e1b2f
GIT binary patch
literal 2056
zcmb_dUvC>l5Wm~|>$9CWX_A(ff@y(bAvv_-u|-W2w~a`hDt3i9Ax_uZb9~PI)9#we
z<{}}&122{M06e7PiLbP;JoPK^#O&U=Y11O{z@Bb)c6MfOW_IRx_nJ*dfpO=jKfZT>
z`%5eqLj&e6y!2NfSaB7sVic<hLW-&+nyag&#BdGR(PA^PTuaLI*iLG0P4Y(UBz3ng
zc{6S#O}8m|D_%;L-DOn?YhfMRzazJWYuNckaaV91H(*}HO}qs21-y(~Ft1@WP&zB~
z#*WYYcAWV*B;A+bo6_&7Tnk2rFj?Et$R{_y0M_13{KK#xGT>KtBFaegbjm_J@RK|a
z2}qj<eu}eXAwb*;N^5RM!Zr<AV2BUEs8#Pa%=9-Pg~F6m^@Va|(#s5C^;9dA7wU=n
zv&!^SL-6z#$k7{eQ^=1E2G;K>g;A&{=-6b|#Kzi@PQGKc!aPw^g!N;KIg|RNF=-ZN
zfeMQ-Y^12r3JbKCCd-nx3LTrztz)~eAv4ze=604snA3pKHXDa+e3q#8C?isK82u2Y
za)*YjJHITF30(WekH<IKX~x=gn&%mTJk77Zu{ZW9?3ZWeJ{!MOAqoBa9g{mxeHM&8
z8qGs)(eZRNiUm`qnb*^4;yugAQAj9v@+kF^Fv-Y_*D%R)FZL<xtnik%vo+ir^oHIx
z%zOI}xPBDQC|}y5G$bs_(jFlh;U>h9_>3c*yG?M(6ES+iAGdiJva=;Y;2yX8Pab&>
z9zPinagW?!841SRn6fCQTo<<75mD~RYT|X_gtxc9wbMJ`bz%F>{=wJ1g8_FQ@Akb%
zy+`{8-*#%eAuJ(lUg{?yH)tM5jEGJmbwQSaxV2ZFajQ}%4dH}YmhPVl`3RPOf4Mo%
zl5lfA!#JJM%}RAMp9L|kXeFbYRng}&zTyRoyY@=yML3MMfhcHIwNXp8H3u!B4P>hb
z9tU0%9vhMCpc4qi#jt>pb^RGIOkwH-O*Bb$X2{tDYF$xIROoa-f#(L)n>Jq`5<g0-
zR%|~bex5^14m$?93I~x71bH(n9b10Ayc}bAu(jLgR(12-7(CwF9da#78F%h&4Ywb9
zgI%{r*2Os@SFe$4u;d$OIbgz*DigP0kBU?)sEoM<v6Bw0|CgxOh9PC=gTDV?gVu+=
z!H|3c&X7+9xgm(i`@e^|{#uyu=_<-QVDomApMx|Ir7wn<McvE0y$g)Mor%g2(?;k7
z(T|x9xMeUvr3oNcg<hqOc7n)lAwRVW*a6HI=&Q{yTpN@feg-k!`G*yRw%IHX-vK54
zj?JBjMu1L!8icYBhzSnKwj`^9y8wji;s9O~trg~3Fzz7I7V=s!C8gwC4akOj2r5+m
z$l$_T>MCj=DetUv%kyv+c%Hi^8>+i#rLKSs%I(FP8&W{l;3%>oKDkmF1Q`jtZ~>#}
zP*$zGaqgh%mq9czaxb3ST>M6eE_UCS@Y9v(BR%Qjv)__Tg?s5Ol?q48bhU(g*98&f
nOvG)Mkn^tgux}?Bp2p!<HmK1{Ky(2OYGd)}j%Hu6>vsJgY>5pV

literal 0
HcmV?d00001

diff --git a/pycls/datasets/__pycache__/paths.cpython-37.pyc b/pycls/datasets/__pycache__/paths.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e9103017520148771c74d6df4ca0d2f1a14f7abc
GIT binary patch
literal 815
zcmZWnJ&)5c7`9zENk4KU6%s>`U?@WaLSj159Rwtn6Dmv@u%g&)bC))i9ZpKbz`)O7
zxQdCt<duoPz{LBSMD&Cuzpwqs@$=!l48t*j@%+p8kIx2#{KQQ^hOjt;Y2HDRBqfqc
zR??IzddY4Eu5vXzP&qiT)Ki}1GXixuBhvfEQeXNq_)5}1j^r5nNKWJw`dG5}WFEFB
zNmge@)zLnyH)in*SjdIt63{k?^J10hCy(1vTx_$o+NpZZtlyzF=)s)gXE3mX1M&){
znL_zYYH~vl<jMm!+CEBDt@O6oDH9c|sNSdunUH^znJ9BRCuIRdo^YZpzfz5vQ#<X>
zie0u<KZp*1;P=ybQ*Bk;Hd5{y6L&}AeUq1_s~rYg=W*ZI5ZDC5j$wvs85F`azQip%
zh&wyTR)e9t#94W;RJAVDH8{bGoaks(>0?uiIo0@p(0pi4@ZU0mtqs?z{-@#|K)s3y
ztaGOb7T9lxN>Rrt{0g9iwGWqST^Pu<53qmENH&mEk!NMuEG~L2%yP`na!lhtNxHks
z1#|yM(h02R6YC2ht6T`nDx*WdEWfOZo%Lm*-L-)btD;mwq@hS&zY)p#hjWp<UAiD#
myc*qUyzht|)737O;B;HbPo;W+kilb|1^f>8=#&S;VE6~VmBZWs

literal 0
HcmV?d00001

diff --git a/pycls/datasets/__pycache__/transforms.cpython-37.pyc b/pycls/datasets/__pycache__/transforms.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b4898d3646c5ac135d3b79c8bfb5b7a6b2b14765
GIT binary patch
literal 3422
zcmaJ@&u=706|SoOIX~@nW*yrx%izE&vt;eC2o$6x&azpNmB5M;(4x_Xp)=Lvnej~b
zY;`%dhaMEgSq?}m?U56*l@k|E`AazVi6{r~jSCkp@V)AB+i{}wXsTZQ>VEaT?|oH2
ztJTU3&rd)4^tXShG4=&D&L0cLoA@U^6q2=><Wh`yo10p+g{iH!EiGw(!rG2>q>I{>
z%hH#nPk7stFUqp4!185P)-YC*b-94LELY@`d;udB*^t-KQ<dU9)>@q|KO6_UVN<Ih
zN_w#x2YL`k$>te+U^`v%i5`3YzmGb33&r6Znx$+A-&WZ%e~Z2U>dZ=6YOS)AQ@MiC
z)Vga*^a<*FY2No_@ewS1T})qlCsg#+q&W%I?QTDaqHxrV$dA!rYJ6$Fa{ui|&D^!t
zrsDW`OJolG-3>D*=?9arWoKeA&g^j*M46RnneoA2nes99zprog<8in(J(4myO15O6
zgCx{jlcVk^nV0jvWUBxn*_<3@<!(HRRfn8eq;1pdm}u|@U*U?v*cH@+hfs3xH|T7@
zkQ^xRp2*9dH4_?G0TCgEw4PbmyFGom*!wh8adQ$#If!<@yxo>C8#YDecM%yKM7qVX
z*I0(YK}3_x!qS$dC`OqNyORv2VOApRU?6qh0Ff6ZGl#y{!SD)N2`yt{iK}ZgW&!LX
z>a;Hxf3vUO!GQmK&rPbDeatl<iWBP;@!n(VG3{(ms5esvc3Mxd%kxs&PpuxJKV#AQ
z<x%R-XXT@Y5zh2>fl5DCgJ~S;VASl529q3^ST)~%bibej@L7UiZ3~^o+By3l9uT5_
zFG)g0WPYGjtQ4}1WiD7F<8kKgM+0KiIcj{8IkA$Vx(e08g%cfe;s)N5R8cTd<#q1!
z8)BWSRkRDxh$aPSe?r56W~&Az48inF%wQpPkj78NDtjzu*38yoXr0&~j+OEib}D90
z(PPhhT)=13n$*FMzsO*@_!-N9=%wDpdvo&biR6#k`Ix63Nasv93V3rUnq3vop}mq~
zhEHk%8#tq!+iAImA&N+0&{Wr9Wgfj}5NXvwtL5h`w26UcZN~c<Kgjq|#t$=ZFz(Rz
zuA*GYeKBJ#ye6T1W5VZEzGQwIe2ptgtRm$I;UeWu&`9~v2<9A7U<xF~9+shIK;aIF
z5ZXURjW%lj9z}!+HnV4rnaz~cT!YryPQ@`|L6X&Ar~H`D>}ahRvoIzujXByGA`Xa(
z-4n1qO7a;b;KbLZVfhqaLc-rDuur;>H_G+PT;4e&Jwa~Q0B5Ks@yO_3t70I}Ec7Dn
z1KrP@-YAZhsREr@-4E}8*`dk$>cgF1J$SeCi-$W8?!BA2L86Z)VNQMzs{;MQI|$@D
zK9tAzTVxTd+fKg|g$HftfU2tza>(QGN0^sTzL8!uxy{SSfOWCPm$}+RZxJ02jUqaK
zK;x^?$+4tZtS|{IbByP`h}flGa6Ps0V}|j>ePiJ`<WGg+J+j~;`V`Y1#3-3+8@zXc
zy*0gAVBZZ1UBmmyS7V@FhPN#zXZiPFtD7h?yBiKhImNxNLr;C1rZ%bg4izK<d0`Z{
zl7G|?+=yu|Axi*mnHzGyh6VrpedBLG|9tw{_L(_dj4-2{Qjh>O3x9)Lgm_%0I+52-
zod%n^&-vjk4cgL){G7kPKJ#)Bf$UhP5DU!l=dxl;2O?9_j&`+2G);Ywx|EY0bT3~W
z$wP#ZO8$mZCS551pVLZOMhcZ@6<9VB>}Fa{9aB?VsL!MqEt)lD>A};Ak+-UuCDh-)
zH0q>fGg?CstN2w+$ZKgWEt7RA<n^?EfdctJA%D)34PC+6R2|mRdM<#l>(WO$V#Ub<
z{8*%Pd~)*<)7Mh2moSSRi(|SCm8M$-O&&(wa6-q&ZKTjB{6@a)B06~fy*)SnwY~dm
z^Ze($=O5eo#f7WyVq)usoL`jh7gDO0sY#r}?Izxj<eYlR6Q$KFH1-k|q_c`tQr)8B
z`&7J2#ce8xM<y3r3weT5s<%sl3WH8h1>MXa2Zx;jGC&onpV>5=@lLy<1GO9K4z=6W
zAen?+-BBb%ZTG;ix9#?G3IAroUz;mBX_7PxehEJkl5o-v->7;Ws=kI(w9FUz4g5Cv
zI?`d0tJg8Bdrp6%enJ0_&`6Q@pjDMF40Np;f)?mA7b~2jYzXB?3B|x^$))qmYS26=
z`bAs1=*Jm*xT&EgL$|obuFr7FrY^3m4#?`RuvKUc_1tAo9?!3>{w12+8-<5Ep`MfE
zo%{Ei!+|D!`MO)HT7&y%bbz*at|C7%HUe-sNX%%iKi&l^GB>{s<6gTzCYhgvfmVh&
z=0%<fGa@{i^n(uWzHKiY>~=m3Msp+VW{w$bTWa?MMNnRqCx0~9?Q4kP2N*04UmHQP
z`7PXdamCrrohI>z?{p!+)$0eS)E%;Gnb+yaxZCL%(i(ndPP9Ls9OYXfMa%CGbXhWI
yxuI!JSPHoL#3j#+gy#3vySVJ`kHR;oE1@j0g^$~TZ~4COmMi6|U-!%9>;DI*a|6`?

literal 0
HcmV?d00001

diff --git a/pycls/datasets/cifar10.py b/pycls/datasets/cifar10.py
new file mode 100644
index 0000000..059b2e2
--- /dev/null
+++ b/pycls/datasets/cifar10.py
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""CIFAR10 dataset."""
+
+import os
+import pickle
+
+import numpy as np
+import pycls.datasets.transforms as transforms
+import pycls.utils.logging as lu
+import torch
+import torch.utils.data
+from pycls.core.config import cfg
+
+
+logger = lu.get_logger(__name__)
+
+# Per-channel mean and SD values in BGR order
+_MEAN = [125.3, 123.0, 113.9]
+_SD = [63.0, 62.1, 66.7]
+
+
+class Cifar10(torch.utils.data.Dataset):
+    """CIFAR-10 dataset."""
+
+    def __init__(self, data_path, split):
+        assert os.path.exists(data_path), "Data path '{}' not found".format(data_path)
+        assert split in ["train", "test"], "Split '{}' not supported for cifar".format(
+            split
+        )
+        logger.info("Constructing CIFAR-10 {}...".format(split))
+        self._data_path = data_path
+        self._split = split
+        # Data format:
+        #   self._inputs - (split_size, 3, im_size, im_size) ndarray
+        #   self._labels - split_size list
+        self._inputs, self._labels = self._load_data()
+
+    def _load_batch(self, batch_path):
+        with open(batch_path, "rb") as f:
+            d = pickle.load(f, encoding="bytes")
+        return d[b"data"], d[b"labels"]
+
+    def _load_data(self):
+        """Loads data in memory."""
+        logger.info("{} data path: {}".format(self._split, self._data_path))
+        # Compute data batch names
+        if self._split == "train":
+            batch_names = ["data_batch_{}".format(i) for i in range(1, 6)]
+        else:
+            batch_names = ["test_batch"]
+        # Load data batches
+        inputs, labels = [], []
+        for batch_name in batch_names:
+            batch_path = os.path.join(self._data_path, batch_name)
+            inputs_batch, labels_batch = self._load_batch(batch_path)
+            inputs.append(inputs_batch)
+            labels += labels_batch
+        # Combine and reshape the inputs
+        inputs = np.vstack(inputs).astype(np.float32)
+        inputs = inputs.reshape((-1, 3, cfg.TRAIN.IM_SIZE, cfg.TRAIN.IM_SIZE))
+        return inputs, labels
+
+    def _prepare_im(self, im):
+        """Prepares the image for network input."""
+        im = transforms.color_norm(im, _MEAN, _SD)
+        if self._split == "train":
+            im = transforms.horizontal_flip(im=im, p=0.5)
+            im = transforms.random_crop(im=im, size=cfg.TRAIN.IM_SIZE, pad_size=4)
+        return im
+
+    def __getitem__(self, index):
+        im, label = self._inputs[index, ...].copy(), self._labels[index]
+        im = self._prepare_im(im)
+        return im, label
+
+    def __len__(self):
+        return self._inputs.shape[0]
diff --git a/pycls/datasets/imagenet.py b/pycls/datasets/imagenet.py
new file mode 100644
index 0000000..b2c68fd
--- /dev/null
+++ b/pycls/datasets/imagenet.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""ImageNet dataset."""
+
+import os
+import re
+
+import cv2
+import numpy as np
+import pycls.datasets.transforms as transforms
+import pycls.utils.logging as lu
+import torch
+import torch.utils.data
+from pycls.core.config import cfg
+
+
+logger = lu.get_logger(__name__)
+
+# Per-channel mean and SD values in BGR order
+_MEAN = [0.406, 0.456, 0.485]
+_SD = [0.225, 0.224, 0.229]
+
+# Eig vals and vecs of the cov mat
+_EIG_VALS = np.array([[0.2175, 0.0188, 0.0045]])
+_EIG_VECS = np.array(
+    [[-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203]]
+)
+
+
+class ImageNet(torch.utils.data.Dataset):
+    """ImageNet dataset."""
+
+    def __init__(self, data_path, split):
+        assert os.path.exists(data_path), "Data path '{}' not found".format(data_path)
+        assert split in [
+            "train",
+            "val",
+        ], "Split '{}' not supported for ImageNet".format(split)
+        logger.info("Constructing ImageNet {}...".format(split))
+        self._data_path = data_path
+        self._split = split
+        self._construct_imdb()
+
+    def _construct_imdb(self):
+        """Constructs the imdb."""
+        # Compile the split data path
+        split_path = os.path.join(self._data_path, self._split)
+        logger.info("{} data path: {}".format(self._split, split_path))
+        # Images are stored per class in subdirs (format: n<number>)
+        self._class_ids = sorted(
+            f for f in os.listdir(split_path) if re.match(r"^n[0-9]+$", f)
+        )
+        # Map ImageNet class ids to contiguous ids
+        self._class_id_cont_id = {v: i for i, v in enumerate(self._class_ids)}
+        # Construct the image db
+        self._imdb = []
+        for class_id in self._class_ids:
+            cont_id = self._class_id_cont_id[class_id]
+            im_dir = os.path.join(split_path, class_id)
+            for im_name in os.listdir(im_dir):
+                self._imdb.append(
+                    {"im_path": os.path.join(im_dir, im_name), "class": cont_id}
+                )
+        logger.info("Number of images: {}".format(len(self._imdb)))
+        logger.info("Number of classes: {}".format(len(self._class_ids)))
+
+    def _prepare_im(self, im):
+        """Prepares the image for network input."""
+        # Train and test setups differ
+        if self._split == "train":
+            # Scale and aspect ratio
+            im = transforms.random_sized_crop(
+                im=im, size=cfg.TRAIN.IM_SIZE, area_frac=0.08
+            )
+            # Horizontal flip
+            im = transforms.horizontal_flip(im=im, p=0.5, order="HWC")
+        else:
+            # Scale and center crop
+            im = transforms.scale(cfg.TEST.IM_SIZE, im)
+            im = transforms.center_crop(cfg.TRAIN.IM_SIZE, im)
+        # HWC -> CHW
+        im = im.transpose([2, 0, 1])
+        # [0, 255] -> [0, 1]
+        im = im / 255.0
+        # PCA jitter
+        if self._split == "train":
+            im = transforms.lighting(im, 0.1, _EIG_VALS, _EIG_VECS)
+        # Color normalization
+        im = transforms.color_norm(im, _MEAN, _SD)
+        return im
+
+    def __getitem__(self, index):
+        # Load the image
+        im = cv2.imread(self._imdb[index]["im_path"])
+        im = im.astype(np.float32, copy=False)
+        # Prepare the image for training / testing
+        im = self._prepare_im(im)
+        # Retrieve the label
+        label = self._imdb[index]["class"]
+        return im, label
+
+    def __len__(self):
+        return len(self._imdb)
diff --git a/pycls/datasets/loader.py b/pycls/datasets/loader.py
new file mode 100644
index 0000000..49ea817
--- /dev/null
+++ b/pycls/datasets/loader.py
@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Data loader."""
+
+import pycls.datasets.paths as dp
+import torch
+from pycls.core.config import cfg
+from pycls.datasets.cifar10 import Cifar10
+from pycls.datasets.imagenet import ImageNet
+from torch.utils.data.distributed import DistributedSampler
+from torch.utils.data.sampler import RandomSampler
+
+
+# Supported datasets
+_DATASET_CATALOG = {"cifar10": Cifar10, "imagenet": ImageNet}
+
+
+def _construct_loader(dataset_name, split, batch_size, shuffle, drop_last):
+    """Constructs the data loader for the given dataset."""
+    assert dataset_name in _DATASET_CATALOG.keys(), "Dataset '{}' not supported".format(
+        dataset_name
+    )
+    assert dp.has_data_path(dataset_name), "Dataset '{}' has no data path".format(
+        dataset_name
+    )
+    # Retrieve the data path for the dataset
+    data_path = dp.get_data_path(dataset_name)
+    # Construct the dataset
+    dataset = _DATASET_CATALOG[dataset_name](data_path, split)
+    # Create a sampler for multi-process training
+    sampler = DistributedSampler(dataset) if cfg.NUM_GPUS > 1 else None
+    # Create a loader
+    loader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=(False if sampler else shuffle),
+        sampler=sampler,
+        num_workers=cfg.DATA_LOADER.NUM_WORKERS,
+        pin_memory=cfg.DATA_LOADER.PIN_MEMORY,
+        drop_last=drop_last,
+    )
+    return loader
+
+
+def construct_train_loader():
+    """Train loader wrapper."""
+    return _construct_loader(
+        dataset_name=cfg.TRAIN.DATASET,
+        split=cfg.TRAIN.SPLIT,
+        batch_size=int(cfg.TRAIN.BATCH_SIZE / cfg.NUM_GPUS),
+        shuffle=True,
+        drop_last=True,
+    )
+
+
+def construct_test_loader():
+    """Test loader wrapper."""
+    return _construct_loader(
+        dataset_name=cfg.TEST.DATASET,
+        split=cfg.TEST.SPLIT,
+        batch_size=int(cfg.TEST.BATCH_SIZE / cfg.NUM_GPUS),
+        shuffle=False,
+        drop_last=False,
+    )
+
+
+def shuffle(loader, cur_epoch):
+    """"Shuffles the data."""
+    assert isinstance(
+        loader.sampler, (RandomSampler, DistributedSampler)
+    ), "Sampler type '{}' not supported".format(type(loader.sampler))
+    # RandomSampler handles shuffling automatically
+    if isinstance(loader.sampler, DistributedSampler):
+        # DistributedSampler shuffles data based on epoch
+        loader.sampler.set_epoch(cur_epoch)
diff --git a/pycls/datasets/paths.py b/pycls/datasets/paths.py
new file mode 100644
index 0000000..93eef0c
--- /dev/null
+++ b/pycls/datasets/paths.py
@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Dataset paths."""
+
+import os
+
+
+# Default data directory (/path/pycls/pycls/datasets/data)
+_DEF_DATA_DIR = os.path.join(os.path.dirname(__file__), "data")
+
+# Data paths
+_paths = {
+    "cifar10": _DEF_DATA_DIR + "/cifar10",
+    "imagenet": _DEF_DATA_DIR + "/imagenet",
+}
+
+
+def has_data_path(dataset_name):
+    """Determines if the dataset has a data path."""
+    return dataset_name in _paths.keys()
+
+
+def get_data_path(dataset_name):
+    """Retrieves data path for the dataset."""
+    return _paths[dataset_name]
+
+
+def register_path(name, path):
+    """Registers a dataset path dynamically."""
+    _paths[name] = path
diff --git a/pycls/datasets/transforms.py b/pycls/datasets/transforms.py
new file mode 100644
index 0000000..f5f2092
--- /dev/null
+++ b/pycls/datasets/transforms.py
@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Image transformations."""
+
+import math
+
+import cv2
+import numpy as np
+
+
+def color_norm(im, mean, std):
+    """Performs per-channel normalization (CHW format)."""
+    for i in range(im.shape[0]):
+        im[i] = im[i] - mean[i]
+        im[i] = im[i] / std[i]
+    return im
+
+
+def zero_pad(im, pad_size):
+    """Performs zero padding (CHW format)."""
+    pad_width = ((0, 0), (pad_size, pad_size), (pad_size, pad_size))
+    return np.pad(im, pad_width, mode="constant")
+
+
+def horizontal_flip(im, p, order="CHW"):
+    """Performs horizontal flip (CHW or HWC format)."""
+    assert order in ["CHW", "HWC"]
+    if np.random.uniform() < p:
+        if order == "CHW":
+            im = im[:, :, ::-1]
+        else:
+            im = im[:, ::-1, :]
+    return im
+
+
+def random_crop(im, size, pad_size=0):
+    """Performs random crop (CHW format)."""
+    if pad_size > 0:
+        im = zero_pad(im=im, pad_size=pad_size)
+    h, w = im.shape[1:]
+    y = np.random.randint(0, h - size)
+    x = np.random.randint(0, w - size)
+    im_crop = im[:, y : (y + size), x : (x + size)]
+    assert im_crop.shape[1:] == (size, size)
+    return im_crop
+
+
+def scale(size, im):
+    """Performs scaling (HWC format)."""
+    h, w = im.shape[:2]
+    if (w <= h and w == size) or (h <= w and h == size):
+        return im
+    h_new, w_new = size, size
+    if w < h:
+        h_new = int(math.floor((float(h) / w) * size))
+    else:
+        w_new = int(math.floor((float(w) / h) * size))
+    im = cv2.resize(im, (w_new, h_new), interpolation=cv2.INTER_LINEAR)
+    return im.astype(np.float32)
+
+
+def center_crop(size, im):
+    """Performs center cropping (HWC format)."""
+    h, w = im.shape[:2]
+    y = int(math.ceil((h - size) / 2))
+    x = int(math.ceil((w - size) / 2))
+    im_crop = im[y : (y + size), x : (x + size), :]
+    assert im_crop.shape[:2] == (size, size)
+    return im_crop
+
+
+def random_sized_crop(im, size, area_frac=0.08, max_iter=10):
+    """Performs Inception-style cropping (HWC format)."""
+    h, w = im.shape[:2]
+    area = h * w
+    for _ in range(max_iter):
+        target_area = np.random.uniform(area_frac, 1.0) * area
+        aspect_ratio = np.random.uniform(3.0 / 4.0, 4.0 / 3.0)
+        w_crop = int(round(math.sqrt(float(target_area) * aspect_ratio)))
+        h_crop = int(round(math.sqrt(float(target_area) / aspect_ratio)))
+        if np.random.uniform() < 0.5:
+            w_crop, h_crop = h_crop, w_crop
+        if h_crop <= h and w_crop <= w:
+            y = 0 if h_crop == h else np.random.randint(0, h - h_crop)
+            x = 0 if w_crop == w else np.random.randint(0, w - w_crop)
+            im_crop = im[y : (y + h_crop), x : (x + w_crop), :]
+            assert im_crop.shape[:2] == (h_crop, w_crop)
+            im_crop = cv2.resize(im_crop, (size, size), interpolation=cv2.INTER_LINEAR)
+            return im_crop.astype(np.float32)
+    return center_crop(size, scale(size, im))
+
+
+def lighting(im, alpha_std, eig_val, eig_vec):
+    """Performs AlexNet-style PCA jitter (CHW format)."""
+    if alpha_std == 0:
+        return im
+    alpha = np.random.normal(0, alpha_std, size=(1, 3))
+    rgb = np.sum(
+        eig_vec * np.repeat(alpha, 3, axis=0) * np.repeat(eig_val, 3, axis=0), axis=1
+    )
+    for i in range(im.shape[0]):
+        im[i] = im[i] + rgb[2 - i]
+    return im
diff --git a/pycls/models/__init__.py b/pycls/models/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/pycls/models/__pycache__/__init__.cpython-37.pyc b/pycls/models/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..58e45f14e8cdd8eef6fde32e51ea69cc4a2c1899
GIT binary patch
literal 139
zcmZ?b<>g`k0^5egaUl9Jh=2h`Aj1KOi&=m~3PUi1CZpd<h9ZzKg7{^kpOK%Ns$W%^
zl9E?hte=utl31KtqF+#%oKp-XbMsSDft>jG%)HE!_;|g7%3B;ZK*7?SR6CGqpMjVG
E01q@F0{{R3

literal 0
HcmV?d00001

diff --git a/pycls/models/__pycache__/anynet.cpython-37.pyc b/pycls/models/__pycache__/anynet.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..bf4beebbffd2ea32a5f7a8c853aef7e1ff0def19
GIT binary patch
literal 13481
zcmd5@+i%=fdgpC8oU87Z<y)AxxlH27iPP?OvvD@IEGOQz$F3ze-Pujq(U6iy8qJIk
zDNCA+irqNCVu9wN3oOtAMIoVmXwirM1AXdCp9=J$Pkd}23Zp<#^f52(?{`SfMH<I9
zWw#Xgk^IOZ&-u>x`)=pxN}-TduwDAeU;W{qy`w1q&WmI-aPlE`|68h}xJpHFRaa}M
zo0_Jm%BEg1)QFRcIc6mtbF9rwB_q%EMs_n-$*GEG?B~70fm$iLhMW3Asf@X%o5pec
zBgM7c%omEA@iOZZ5B16<I9WFbPA=w5fs=O&;1pucG&n_f44ko;GXu`JI|0r_%sBzh
zq&o%9RLq$LXWE?sXC~&H1m}c13(joJnFHsfI|t5O%sB<lDfcuur(@1(aL%}A!8sdq
z&VX~yeF>bGV$NAtTT)7Pr+B%!TlNBbv*mgXfBq4rTP~@gURzzOL9lYuKq>A8?EXAR
zpag25x$0l&dumtds{8tZvSIqK28OHcr@G34`cVB`?c&~?g3suk-@M}mqW-}1?ckng
z`+>J<uWmJKLA}+qS9a}Yb<>+K>6N1Je1~_PT79)DD!JZuy%`q$`ewV~MfdB?&Y2~?
zCfIFz_G_Q*y=FICf$eX#+bt1zZpjSucRZhuEbvufCce5}4)aU$MZAYAW5e6^!{TM%
z_XOLsC`3zy=4wl9R)dljW@LN#nM!nuhWmf3h#chb)x`_<TASX5&aUe=cl`@)HK_Vt
zaG||hYxpsZBzvLS+--Wne0w*{uX#b#3A|~FBF%#+x~6KXrA=ZN1zaDa^k0$E_Y(21
zG+MQdk@7D*s5a}3M%9szRK_U&m8xH_MT~G-GFm~<@S0w2BjR?>UXd^SnvwYZ=(X~>
zuz-=c!k0#ulf3M`Qu0I^{TCKV1_XnUUEwDjISTPPhd~ZPL@2|sK+#5z;<gReQQ1=B
zHSE1jkI($qapEZz6~+X*UD1E0tiq(Y`WIHkh|NOkX8=OQgif1b8qoRJtGb<hjG=qD
zn6D+T<1vMuNB?aQf~KoH)b_OlSg1hXH(d1q948g%>n6aW@nc`=;G|u`#weGxN5r6*
z0a*qrjg@-UFIiy<dhQ7^4R&ZbPQ6(V94FM8&2ZwfTWtqG+T{mpw_B~oMK`pn57wME
z&O`G?z3Bl9+G?$&hlcMpRzqVOKgUmNczwokYK^MzJ5FDw;v_!8=KCgyq88P>+8K|#
zKA(KMMiaq}NA`LFyZ;7ASJ{VwI8b3I)P3#{?E%DkpwV>f>&ZQuiLzjCAN?=3k|9px
zJTxBEy=^qL-toj)FiM&@N4~mqc<j?a>~>YSzk}O8bw^RN>Lm8gSkl_~iBg(Eqs@WP
zvW^3UZo#fNVcv1>Z&e%dowVb)tr}!O35Zug!a^*dn%8I~LZTfgsp4fmL?x7?@)k~{
zz%13!@+rxyf!Et;PPzFPK?ar1%}BM=UE_-k)Gp<k@Iom!?OHgRQvK49RaT7Qhn-hr
zWZ6>b-?M-8?#{b*t<`*R-o8>kZ{P76Tc6rH7t1vwiW`qTt^5!dUBgv2EOD}%ijO9b
zu+#9KG~zV$Qfp!2RgR`XP3`IjIy@+~R4aGBJ=6-@ZNbhDSYvx-^Sr&becrA%UEB8@
zVYi^7t)^c}MIfI=J7j;Ps4*nyfggyiT2M;M;TA99i7>V8w6=oK#I3sPh1$wysI5^s
z@XBNaI2zF~aMZqsJ<<j3lxjkEJ2Qi(sIR`saPUaok?Q*=I0=+>`Y|YNS0_AO<4Y3&
z)xg)gT41cF)=f#%Zc0Me!kr8_+4Y=jeo1)e<vTRK18?ZRqwv^w6py@POo}>Pene{`
z&B3*j5#}~L(exUQL<jwHm~K~Hx87V6CvXM3se5fdv^L?Rn!(m4AfB!_+l^|?D~*de
z{=i&-vcO2F)i8IZ8r1HU;rA}Of;bMfE9FpMyuB3aH*ehx4Y-A7Xx#B`+y!ItCLHYI
zjl0hE^6eX!7ZyXa>IkS?sII_F@kw(<GJs?vRt}*cE(BY82)*Jwc-OG|MG!?Z)M<>^
zq;^t0rFLeA$E>FuJ)=XRaF?UTkL{eK**Q?IDxc=}Ao(s-c~8Bc{RE~6DyuTWv1<3~
z4Oe*0k{&hi75p$vHLAO?tWu5ox<v5b!9HjMr-xdVm=?WUE|)Cv8Xgw(C!$0`Ll~J6
z_EXvt@g~oD-%0dWS`b3}ee6<oX+er<D=z_31Gl%a%iN@Q8MGmz99JCA_AQ7N7Nr+n
z7S*QDXs7d*ytadcQ6)$pYg1~2nCY1m8|@i(bdixCG8vG+AtdNO#^Zg#j8>IIH2vWq
zw10%%mrWq7XNSA;1YmzI8~N)1wwvxM>qRGmvsD(f{t2ALyS$L#_-&p=5Ek@j;#(k)
z!?$r*j-e}lmp8sc@*W8hE51uY)e)CSJ|LkH6d#gYCb>dFq=>5^p|x6XRvW<UXyNxY
zG_tDHUkCdroX#B{gyTV~pj{XQsrU$t!wn-uH`p)}Bo0$f9BNPk(ig=iwe}DAExHQn
z;SJ!E+#!4>0P#6KgZOlFU+3f9jGKk8PfK5K!Cd9VI9||s7lCr!MSNnfFnUg0{*L{o
zeeLy~(s}!ueJQ@xgYY-(_`f3s+_xCd>H68d0V^=z^9LNp1wO<;RG<>ji@1i3LV!EN
z{{F;Jvhf7FK$J;m{Y(yDx@NO1VJrh!xaKEfIV0V##BgY|MQi=BAV!??bl41w4xGQ^
zZ`9ik-}o7Y5ut(Zv~%W=D25(BV$EqGe5y2~pAEfTMEs*}FgZaGxDFo%yDlT6C=^7H
zD8s^sd_oV5{ZwGCr`Ii+;AG-JQHrRP4YO=sl&6Us_=JQ`sIEpbk@iYjvk!TXMk_QJ
zT0Fgl`YC?n%hD0@X~unU5ho9I|EO`|H|fY^*J(OsM}I?hRM|%d4`jX;<t($xp4u#;
z({T>NYn&rGj&hwS%t(owc#rr2$@fXhB)34~L`E>`7-h?T4t~@Hbl{9uWFj*&+=axP
z$(SSa_?SbAn1=g_fd_$jKlyMRc$k(Y?az>D)A10u=M6oyrNz$6OZT@tuj9RiwBW79
zomxGR0f+ti(qbtxZbP{=ttmh2Gr$`iVjO63PSOuBZsXwr(-MbzFM-R<*Y`#T7r*3p
zdcj3*X|ZoOlAiNO4O`%&9Mu4Lh!c8b2m3k<iA<H^h~L0<j0t24s=N!8>fhbZ94H3{
z(36b~3N(ayf1fOaBVz;^8AQ=QDHnm}6rYqCb9TvtT$@3?+K57gQFux()z>y#bvHCt
z9d9Rbt`v*-JrZeF`?i(J^XJ&3YKoRed_E3LjjNpkWiuK|G(A!}EK^9JFnxs0AdIG&
zI%1eKv5dR2<xfJyU$X(ksG@eV_-vy7FY3*yBER6f2fTSskXysv+c@Cdo>%3k`bDmO
z$o+bc8T7qq9CyF7^Uir2`TThsXF~}*8_MSR40=XGp{sX|ZmMe@m`5OS7e5++#PfUG
zzhe`6C^^&X%0{n$eb^-R2SSE5J*gNdC`8TIy2iBfkV#s)o9?DgDSK9+A;DI=X(Y`F
zNSJsV;HSFj$hk<Gi4*TE?&QG9uNUIPyC~m*HzMkW^Xghi->okZfjeW5earZfw@h?R
z@sCo-YS)ZgnF%J*&dftpMx3emD@^YPTRJVjp^a$ijGKXHnQ=}?pl6>ka)@v*GY8a9
zhI59{TobLWwl8ys<<e}Z*PB6@0<Y=z1CAF?8pI-1hzO}@z89*sih&AoQ(OlJl_?tN
z!|DET^j6Wgot28&+}d>3{Ks0HzsGx|wO6!p&62f%Nm!l}vNAvml#l8I=SFJYQKLl2
zaF@ErMm|NE;OUc(e`0F~l8@7)T6J7Dix~bJesI8MofPEOu=h3zPFx+$KngMgVReyJ
z#;CfOiZc(>&AIuZ45TO~@RrUc#2;}6(ks|MCJpJ|@8u%@5jE?-zb8G~aAxvi7LAQN
z7K`>TlvU57oo2Yz>&O+(K=ch-B3d5VdZ|>uo&+L_EFW70VDR~Lp)&);f5kqSHM~4Z
z2*asEV%vG-E=9qRh$+F-rw)<BdPpn(gajfAs%2O!otX@(S<Z}V9mBap6m36^mm3xJ
zgm~*vCNZsc&W;kvQFHVEDxLVZUO)Qj#HkVeNDP!NDwO98NV#+@xdlP`&q*s}Zedu=
zEyy1-S~?iqo-@!AL}Z6~Kj*bMd26Fn_QLgRm+uVE&hbHh?AbC3{%XM33}h6tSTQ=G
z_%D8jzhq@f=5f85>Ap!xI?iLyq+G*Aza}v>f$+7H5c}~YgiOHmNsZ!}>Wzhd$PgOP
z?+1^qQA7!@Of)`3d!s_5%yJKlXB5W%2S16cHc{QbGmFt15YSO<QAFGq_|s!|<zMlw
z1Fn2}s5uElUNrBZI{YWUG~mN6sl7FDdmFpS?RnwAibiMdFdVLLR@XcL<Twz)FFZ3M
zFnu_H2rdHPfL|{^0S+v?4#5F4a!3FkHz#K%b8^FQz#9rk1~djf@`<7ZNT&RV{_Bx+
zAcV?ORPv;WL;^h=h?mCt*b$tncy{Q7nVZ#}cnaU(^!-qlE+^7tf-Xs(pJ0}bHBO1Y
zim-$nJyBl@_g4+2akggwN)!EoJ}w+l{g_D(;wU4?JysmC6P>we=}_OT94C(Wg(El$
zje67B4pW%<cD6@5m$@M*%?z$?dIFlR;fpVhCW)PNJel<=JHfEwDX>!-@G#Gbk^~fg
z#LreqMu6gMq8-V^a%Xl}NWW@Syut<zpkjKcImZPIe&9sX^>~`f6678yFhUE<YnB4I
zlg@YJg-|}P?ALu+h^%>bYc<LeeS7;tv;YumZ)|9EdkYcTy=vPl&DU5J;>IJZMD%-d
z1<}_RxKM|&4y+?Zt`+=x085!sMTt;#!%T0rm>4cmjE34g#o3GFCX$}_VV3lQ2AP_;
zz+Nzv-dPT?*5;v#B?%wzWx5$mrfA|+S8rlT$X+(Sexs{4v0Py<$3m&AW6mMtYO`2Q
ziBet;Q|Q^Q$^JexS?T6-1_?bQS~n$8*U?u#+r#3n6ejW4yfDj5En43cEv3tk5`>^W
z2$y772%%c9SSwhp$OShQEa>23u~7)SNV#y-6s@R@tK%rsbf)^!`G)WnrsU`yZqTb}
z#-&3D<BBz@q{mh^riQ(F2wgcr?y;iFj4;(HM8{~M?)=c=&_vh#f568Np+yxH)EZGq
zb=5iK<E|#DE~z1DN>WqOw4|1#8A-EqSm*p@R#lpJ_@2Q9Px9Q7l(|M~Ju7KW(!8Vv
zNsHny<#7zUuHbBZeIj}%w0ts{TAw~pWBlgJf=a|CURVRXEhc%Pxqhf^WB2_~U)cm%
z+xF!ou&{9-1xNlytdsBuVuPeXvPse;X_34`VuO@sqU9%bGPs&V#!k_qC-ZXohvmiP
zFtfD0c+*+_;cYB!lBb`9*(*10E&L$nU0uAr{P9v~escXPkLji5JJ+u+E``>WTg%SP
zyEkz5(Vbg&Z#$nXg{h@Q=N6V4<8UWT-Mn>m@kW>{-@WN9+_=27gde4<?RH}q%P_Yp
zdAXW%+pDkL3;cLaBs4d$v~kTJZIrH23mL95xis)+9{DTkN4b-zpUSn7OapogIcc=G
z(S;YVO9RHbSTw%{lZ1JN1C2x0O~Zhtn|aP*eF5Wk6(xKW@GbFn*8<lV;93DxCo1qr
z1D3**xgaIiW#zkhH`Uc-G2IMMs>i(oj9_u+`+?QXNt5}(kF~wAAE)-l_a^ox_onuy
zp*k7YL~F*}G!0~s?H0Qe-O28Fx6qyHPNO()(MWbD_KNqrpC~~N_p^xhS+bAP;Y_*n
z@<OuE5Nqb_2x0Hpu7ByXz4P`q4?{rHA%-RF{y4}QX0txE_OM#`Pqq67XS3G8#fAbW
zg^1O6+Wgz0=D3IK(^uK|8+AXZwKm%yta0d=ZkS5vmFW8eG5Y-bs5w>}{s%Zoe(@{~
zk{%&$jQ9z@B7RE3)<uMtLwpW$uip+iv(VA~&-S9!MvF@MHl8XKLbK6YTk`~?QiLh=
zcA>$xg?gvnmh0BB9uwddQ@oY&P~!JQHSokLuBCO1RxKM`EDadSxq@fLZzV5745wWc
z)lI)*Nppb}-_nOw%(n0zz<O{zq-zV+?NIdxeA^)joX3MMC2$HS3YHz7z$C-8W&t1w
zo76Mde+v^39r}9-mT$_RnX#)r)IK`2UAB+0TfDX@{>P1Db7wS#3;af}8gb5=5#*N6
zi8pWvr{sS%%x?vC{G$lf0GJ6=@-j>(*BVxUjtxH8^q?Cc;*a?bD`<(@z4>NSrpMt#
z^kl6iym`oUwZ0bF;4IhiN17HbYj}tsOYMsE%DgF;Y)e&&77l;R2M9|!EhT3{WMxs-
zQ)DtI1K&skrP=SxSe_9Iq4j<=03T2Ya(30g0umTXHLquNt6&u-$0pBA&P~2Nd1CUF
N$#au4IL=I-{67!_PAC8X

literal 0
HcmV?d00001

diff --git a/pycls/models/__pycache__/effnet.cpython-37.pyc b/pycls/models/__pycache__/effnet.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3f91695257e1a8847698223cb281017b311e83f7
GIT binary patch
literal 7300
zcmcIp%WoUU8Q&Lomk+&Y`4Kk{Ht54lYdLk$0!^KVEHSa0n08I34u~XMio232lS_Jb
zWy#E{2&w=*1t<#i(4Hg|Jro6c=&`2){TF&D=Gser@2M!z{(iGdkyhlyFx(;L+u3hs
z=kdMg<K>x|f`;<iC%^q@{y9zi2Vcg;LgP);=x+eN*3x|4H+p*CFf?83n=MmMF;<4j
zwd@Szv>Y|o?dMy0g)@7Fez8^5wZPmj1v3YFYxZr;xBT3vnx6~un{#_+YaSTecYtv+
z%mOg3p9dzNVamW1{30;L408gQl0O5?OomwmX4anrW-h~=1ZLh}0A?Y>obru^RxOVf
zYwPRXPB#eSdJtFo13&0ROAl~ftzOlo*;(J{@MT=gtP#D8MxeEHkj3!zmicpS9mMg?
zPu-T49;=m057V-Lu31{k!E&y2u-uiP?T^Y2t$Q1&r4H%M!e<VZi$xWHSc~;U+cWl!
z15iHE_RU03jKn-32Q*Bq*Q=JyMO(u_2!179&+CTW*z;t;>kPsu7F(SdW2`9XtxIbg
zy)w7$4Yp!wgdLnT?|GeGJBmE-_u9WNUbsEz2Ny>7eLuV(UGUp+I|||p!~30Hl);oz
z7lQTmFo>6i_XRh68Z}a<&}a40e0J8Q@m?M}po_6996thxwN0Hzg3u@``qpQ5qIm}J
zW@5zFW^U6~(D8Ghndouj{2mD56=H3_=-US%k5|Honax?>xnl_*U(RLk?q}?s2WEjp
z^YeZ|%~4jQE7l+A*nuE;3lu8e2}Bt5yr?@0qz%G#{Xjayw(obt4QZ`)+fmh!W-yGT
z+Xt1xc&lGE#4P6^*g#bUMAe7QsxHcSt)2j#Lus!J!h08eX@MvAWbtx4?%b{qME{~M
z42+V-<+?O$HyYBscKw=kILccCryASc=(a4V_eixDmi_iH?%oTQ?`_-|40@Q<ZQtAQ
zhG<KpR+lrYH?QBg{$A6&x!k;ZT{^2`FeJMjKWo_UcEg}8q_N(qTEfBdf;<s<0vZKT
zIs`Lp*#KWbjphL~!_pUxv$|`PjFMh9lt5FmmnQPo;h9+Y<N}J{@1p?`W(8mNjYL0C
zzg@mTun%;!^nD98=i|*c4vbwrhCKM@o*~}CD2Jm$e5t;1$KOB^th!?4ga4w4uVbZZ
zUc7*ZfNj&n8G^3@Nb6oV*oHQc$<FO|)Q)4J#*1&F8~o&UrM`272GUs{i0!uUUqW|8
z`e^zoy`Y!%Gk~MHvBYPipzu({u-y-oB6?7otsdJY&%3+T?qxlW=lO$<=P4ahWESVS
z{7j~kouJo4?;$%NQ$_R(#TBaiB5FkDxw>VPEO$;BfesLdg~M2)a{ymvxRk-NK~7sd
zkKv=LlD|s31A%MD-9cDc-wN?E2w$!2R9>n?-HraB>z~`HcBn=yd~!|Jh#K&)4I5i6
zQ_x(fRn+NmG*W(iLUm+yOq!SPp+-uk^wF%^(!{h(iyEg~;aKWpqQgW$pbqrK%u*UM
z7Y{_|wsf+6XY3Z=#_$Q1Q}U7#-{)M}x-;pzV+K27i631cpi2<nA>d{e3&lmYj%OSx
z`wD7Ay;6)b6&*TwJuD<2%ai&Wqfr}ut8sTL2u8vAcIcn4?R2_v%ENPwTD7v)8+7he
zwl7rDb9KfG@<}@K8F8x*^DkA^1gbgZfl^+67lS8~Ts0@Ag?=Xrj_2*0T<6er6dSb(
zYp1tFyR^cgs{mvyEd;DPJjK2V3lFOSF9JUd4`Ri+J$e=kR)6eOER3@cv;zy@I7$d}
z^=a#^On4KRIn1A0ZdaUQyF#FBN1+jbpkdr?_XItj;+k|CsS-$Q-3xZ6k1Z&-6Cp_D
zicIGjQX>K*YU%Sj>u82&how(__Cq$5AF0z*;c?Qj6$K5Z#9L5a)mFbuu&;9inP3;N
z>{zPZM@u!g@TH<n5qOPrPuZlRAhv^gSjLQfQsK6*U0#8Q9KC#Pu-5Gbm2P-15OLsF
z)&_Ch3&H@BJ(2HP<y<<xy40bKVd0Z|<W8b*030vniJ4eQF0l_7-;U$0RN7y~M^n=7
zOi6n#*cp04+G#)`=rtLCc|y6^<JI8D+{mFC%%vNisKR*P^wF$J82*EX&tQmABE)}D
z+tYWQ1WxjfDVni?5K2!NkB^mrp_K-_l{hg1tIZU;S!k7y3o5iKsxLVYPw)69LMQu<
zA%;F&ukZNoXAb5mJ#wCzY@S)QtDM@E8_!`rchA6=^V$6Nk)1858SP_sR!%Hmzheyk
zh4Wx1nB#=vM!uegU-UT$;Zoftxy1s3ioXnTf^90!19JNT#lD)GjFt9=7;FtAHC%mI
zT|}6-Gn8gGj8hw+wEY>*!>`3z0!BpQSpaF1-)muN4)T17RH&Bo@d%`=9t`J-X-oAg
zt&j%$d)?3*iovEVj$0h;B0BM+JKZ578L~eO(tg0!B?5I$Dw~q37J_=L{7942!oY|w
zOEYY}WVnY>oTVF+pK&UU>@Y$_dB$jAEQy~tOr$0^reyI`G!T}7H=8h-Fi>fTM?+;~
zVIKMr=4lA<wEf&eSetMB76G2N@1}ifuDtjJGZw^7T-+>e&Oiz*v}gC{a5+#IgorTm
ziE&!PlQ%n~zJaZ%Qk@seXb75H!Ci{C3Em>85nLhoA;DFGcL1ati#Bdh;YP~p7Zq-G
zC5xtwa7&gLNYhZm&ctJT&k&+zu6suVM`bd9iq446L_?Tn>!(2Gf-!H{Mj3YBMNE?s
z`DvM9+NK8AoHD~-6U25<4+|N{@}y?iMbz7f+Z(~?1X`2JbVS@zhl{iDNt?yxR22Hf
zwl;ANoX=U?8~Aq0+PaUS5mNV$H?_awJ`PRo!uSvqV>{kBV{KN7PDBOTX~tUO?CGo8
zhgWvp#HCRZXA=u%(b&yr?=L01q1d|xh_ddRxaYV&#PdLp&7{D4Xp*N(*|Ql!-7VZo
z9ZPn9QTA+Pu6?}Ad%b9f8-Xm~3Yy&@qzw`2w`0Z2YVk3sAbx~>OT8=IH7~o<+G`%u
z0LKL;zvf99wQ9z!F|W_V3Y2ukUB&0)!V_|pRS}8t;?5ZyTsuIJ^I?fPndtCJIuDxY
zX~sgAZs&Hl=ZhemRv(#DKKtmt7>NH4RfWqM@(b7zGYsjDo|ei`z<$Y9rc|aRi0z;r
z7M_&lNma&f#6kaRee}W+g$eo<oTPHK56wY7YUwG2;}qB_m6;01zaR|xBhPlI8*}My
zCz>%nKA$Ai3WvWAfM|Q1$5AmjLW)P_aqNflILeVbb-`K08w9+(J%%@j_^bF_`DS&&
zf24ypGUwdNDc3s|I-VfSdK_NVCNcV-(5ZxBobVVejHz+FD<n-6JswjD1oI)|t<ZhQ
zT(5l}dAuW=q5S_Dbz1&sv#mTW>8Gsxfon`jzblCCpdJ<;k>!b{A5stLKH@G^Msew(
z1Olxt7mI!aK-t<fU~BE@`PlT0eca`UIe=mL87}$A5vIMD&}*t*V6t92Mo57Ks+%TI
zH}e=@*eohoQgBAWSq0}5WX7<txu9Sf@WkdKZk0%9!?2x<Pi>w)&@%^xJEY!nyVP1_
zLq{((v)G}6Ov%T)m2OlC2XO`I+2MfqlhO0dAc_%NJfgpvXW}vFd)w0RBWY|$X>2JX
z!9WkuB9M20WZKx<R>`{RnKXl#Yu;SGT1UEXxw+h^HRbGb6yZ{YR838Yfe=kjU^Q!v
zW_o}55bqQ7AXJ=VH=4C;-Ysdb)^0SfG^Bm&>MEaw+WR-s0H`6YM$Nm4OK<b$)zw-<
zI`7nO)@$!#?3LQ`D!$EKyS`d`R~GB<UGrAnU2ZfmDAyhid-p}1TPUgY^mfqQxE)7u
z(2O9D_tSqxPgDVb(T6iqRkw8e$ggCm^c$__VZNE!0v6r?s7@=W(Jug0=s_QBB{n>H
z4z8&X8~diuykmiZs;<R3pH7R{3>zVBapwwTsGz+5r4NkV(g(TSncdmlx!w8Q1%|7X
zxuoP<%E#xEVltO3B=hj`v&qb!Loc7q?iTO9eoKoB=yj9)9y8g+EDJGu4tLq~&JW3W
zwKDA)&VRhC4CBa*(4t|imZa6|MzJ(U-J!I5gN=<qNUIyd6p+PAX>r?x6v=!LZuJAv
zjswAHP<gZTUjmB)R4u3R!$<7%2!2A)CQ$c^HMTkgJ^_g;ODeeYhHcUALlB1|xCcRe
zkNs(+u`SdUWu6#yPdQ@Fm`C_Ks>%&&#v+1@5~AW~!PysaEg3N$Q@s3tSy?+x#vW&8
z|H*xEgISm{o9*RknOW-GUr;Az_M|&6h^?Mh2}F59{nfIx6?gG(6q(Ea9&+kcI^*}b
z=|QkN(gaZ2C0@Z?xWX{YxfF)VLCS@6cxNCGln=u7ZhEzmg^eKgQUOZSKHL1*zNT{3
zN>0>`R$Ww8z?ljVy@J~F6a^KJnXZ-YrIfs{)00GJ0W92d%DCl}^paUH%P*7*<x}N1
J%FmQf{tx+f*o^=H

literal 0
HcmV?d00001

diff --git a/pycls/models/__pycache__/regnet.cpython-37.pyc b/pycls/models/__pycache__/regnet.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ba9edc0de1db0293eafc9d6366926717ff645341
GIT binary patch
literal 3809
zcmbtXTW=f372cV>kX(_vS+LW_b+?UUgaJw!c_`A<H7eVz8dMj<lHqodcCq3t>6Mnd
z)XY#UiDgk#DgyMOK;QcyqEG!3{R#a6^O`6Bg}k)q%&sU!77{dFV$WQ6=FB<g`_Ax5
zqv0Cx{Opha_}$5ZVSIs&i^qb>&*2p(AeiABj4+xIpX$==Tht&%-}W7XBW6<RSN*E4
zTS=|&`Ytgd=X@cm&xqe(HgnDlf00#K70M-6V=k1-Y?;+r<Ba$#n+98COJ@dKiriDv
zUuAU9Xs%2aUqro~NP7K@MTuy=0bPupCQ)Xm+k?W{NXIbTq1X=(mZ<nsc*PnBX-JYA
zIeFJOB{^9G3Cohn$4|5LB;rzdq1R2aP<k@+I!PuXA-yz9e>91B=EbS>SbP$T_#lbA
zgRytqYEorFx3YOQOc}IDSCQj?8oUadefiP)>#QHGPsWU;W3kRc8Hz}*55}EDl(IO;
zI*)p3BwK^AT09<xsf;I4pcT1=n(O0Gj?U?rAn?Is&5!nLXrLd?;DC6%Q~C^!HZ_@9
z_R-ud`zW)@J~OBDdG@^T;#*vYcG2q2WPO7j4h2|8zK*<3)*pm2K8TZ8j=ikwjbbKW
zgXL1@^}q}R(drP?5FS+4hIbEMaTi1nmi_J8XLL%Y6jUWSMIZc)9$R~b&6<ScshXyu
zt%r(?6gkkgEJO{4zH}cYv5=_7<L~0SXdTCxydP>MgOLb&BEaF+qa^EuiFgc^i!W|g
z4cF1CtzK2_lzv4qW%lE=X>ve@p-4~Jqu?Nu*R<n{F#0aMrCR^Te%jyVetZR%<jW{9
zvM#Nm_e0Bm?S40#9zpPx6PBjeCoFbs%s129qW$%b-@Duq(Km?)O-s>{Q075jDB2Ug
z1AchhJZj}FFTI0HY2Kl0r1<gM&^HIM1C0v28pMBq3dT4iJ#t6O+_(dQbfpdbr`D7*
z%FHu};SvcJwlq3S2JG`^`W>T;eoK;b3QVNOr711I3ITd35i=v__RN^txs}@*57o&h
zn-PFv1c(+006^|2<3}FQMih2ldqRf2$m{Z~?+qgE9Y7Qwc_(2qj6{ngH0D8p*udjy
zeB6hM4#K;iS%{CRIkgZWN9N#OZsyjJdBs^e0-AQ=syck>)6qZ??o&RvDMWDb3#5G5
zi)5gc6ai`(Xbr_T!Uch;HAR<cO_vL}<dlP^Z)sORg`;+g+z6ESFSXiv^B;KTmO<wD
zcKa#9I4XFP6&_sf*L3(Qv=mJT%3^W13mhJ%y-1l!lqyGNNEt8*30P4s{E8HTjQRsb
zrSL5QJcE*nYx}6%H(U1yxOEZc56B|9OCAvZL+E>h!0KQi;DZw|Aatz%3o01tM>fBY
z5q@Mr3307Ul3TZNjEWDW@h5^YoSQR9Zfk&sI}oqXW5Nqji|KBLGpwA>bbadN*tT<;
zJ2Rl*X{DU)<dxi-!K~cUz1UyPD>KA{b6%SnjNXB)I;~D?dF9-d3xJOL%*bt&&|d+Z
zRPt(G1FQh0H_b_FGfE>K%1B%R!$=hHD6{z><Y7OO5ijxa20_wnsKpH-AS=gN+U7jt
z9C=kaFd6*B(PZGP42XD;jUuib%*`y~_w=k`8XphAoBi<j@QILPfFcGI|2~MO>zkt>
zMDY+spJN8#$TEHfh^v!9PNzpgd*64{VZX$SUxW3E`gOlA^~!?=MKb>o78A(W2K|7n
zl6z#C+y=M5LwF1NIyi9gV6fvOJP?uL6S!Z&y~;N~H@cA3ftjkl^*P|dx51{Catgam
z8m0a8CE_ycVuHdKmlgj8fw!JRdI;eF66g*jga=6H)S7dRDJ#G>kp2PF*3>Sq6IHk*
z5KW;@d6~HEsmL0TVkLJ@P5q6-imS!tjH=$n<$uxTLQ0u2t?7DQ%Z8SV5HLuuuu_hr
zONYzM{<}MM<%(R*fu;{{b&cHpD}n1I5Kz<EnbQR3emth8CA{EiTB=WpG!Ef_S11V*
zDNY94wKsR#`>L|}V)y0q;8l~B2|c*xP@U&C;FAm5e6O%}D+uB=mO=2feEB~s#vh_~
zl_$IV!S>6oQY-!ubl#-blS{q6y)B9@oxS}RpFV4U(~iHm+_CAZngA*c5LC+1)sb>_
z7Ag9ufJ*>SBa$aaw<xm*mjg>g0SCtIG}kl;v_}dp=&KDyUp<6(TiMUr&-Z_+*^%Q#
zLth7!z1I$Qca#Oi3su|OZ*K?tzj@wPpg&Y!Dfh|N?$citefIY5v-Xx+*m=1fJl)#Z
z+iUMBTMU5$_&RD+z3^dk0hLO+zIA|vZfm3$R@b`!1h@YL1)hR`jN+#tlyfu+c~8LQ
zrxpaABorbj14!SIK7~2>t-y?9P)hEbi*%K&(q+0z*Gz}tk6!uJ>163zd^H`^GX-W0
z0_6q)pnRBMy&eScCm|_EDnY=q4(yJ5sz#|zCrXkFjWEN&y?P4@uC+rO{1{$=gjgjO
z)mDKi??JnHi$8}#p)CgEf^%=_zbvhxjNwlgP;M_yfp&FMRW8~z8LE|HVkhHK3rMsZ
z_X^aN3zUD`!gLIGCXT>WE~ZOd0(`TL9ZuQ6RH!{)1Ok$#R`UuR=66z6eN<TIF?Q*H
VdKRq$>()u#bj_+=tyODo{l5o^LP7ul

literal 0
HcmV?d00001

diff --git a/pycls/models/__pycache__/resnet.cpython-37.pyc b/pycls/models/__pycache__/resnet.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f90ccf07559970934a722a186d8e5bf554e9b3f9
GIT binary patch
literal 9154
zcmcIqO>7%UcJ4nmn-oP+lx)c!$Kh<8u{k>{kL}srWHP(4Y-?r=*c?ZaGZSKWdnr~+
zB1JW|)h$_~Ng!DduqOjK<dDN6kPslio^#AKrvSO;P(XkH$wjB!b4kAUs@bGyc|4n$
zZM%xqudA!8UcL9dpVGbQ>AZpKM}PVEKfC^pVf-iE3@-<byC~6rpb|#i5T>v?=B{NK
zrm<_+9aB>>3B|2v6N<N+tLM0Fck;Va^(oT`oZ~_;ePY&&!VwwaitMlL`i$^I?n|Ry
z`os`<G4-V(rh;61_Jv(9gHjOFpiC!}IZ%pX29%kEG7m~g%z`qTP%eN{7IUD?C6tSx
z%!>=4Tu3Mjpj;FSpe!VmMPaQO)y3ZIY7o_e8(&o8%5GN#ooMM43pZ+2Q`ybUt*Ucs
z<Mx!;a_XXzbX#@!M>2I`f_myv-)kw=KD|wZfU3Vm)ihDjWuedr_fewvP{l@U3iH?2
z!0a1+^Vm8ucHHP%Yzyny=^H2J3-gy|AI~lsn7Z9tS`A{^3J!v(5^o2UST@3FvnzKS
zajP3vHuu9OH#UwcVPiK~s#<DhqY<^5e%iA>ztN53P7nsooe{P7?)qzny8e7{y%Kig
zO0>VX*OhS~su@+hM}wxXsJVOTm?NUI6C6dVxEw`+q*<*<*_F!G)5XjjnqocAGxIU^
z=r@L(155tr_Ra0?Zg8`ABt&==-4u<u5e4zhy`yF)N@~4tH)RloLA<ngq^7rmSTBc$
zqWy_j;BCwF?4nsP<pt21DCtGJN>_9PRbbRjh`uGvy8SC-6AB>gFTJ`$LJYyp{0iGv
zcd-Yq%0hfT4H}|XU^$8_+qhY34u_-P3<~{5@1vqE?i(*G+Fa9!?PEunC!qM5*lxSn
zT#G*T(gr0fXq%l{)jB1Kk+hC=h^4d9YD85}Wv~f>lovr(p6|E9R_y!A3PUx!EE;=p
z>mXP@*m~0Kc5VyhH4e7?J+zhks1*hcsjSUr)mBavbT*aq46ozWCWg=XezVhvBH!l?
zkn{BI%cy`9Mno!Ln8d?gDOvi`@Y_uiXu9Z^R(&5OdKXpSIEFx<m=J99m<m{X0=Ay8
zs2|(IXC&4&Nvl2mUtCbL-$PqD2d&^4rrPQS^7lcgTCzfP^H5Kb1cm3RvJj|e4Jp10
zN<?~Sn0d2|(wiC1HtA8#64OZPz_h&YLx}b})Gqk`FZLUq<Vn`|MYjoF5C<d)hnh|p
z)C@YEA(LnZRa5=}bz+9muFxGb%rhOUkl~$bf|g#S{B%VG5aU88@+@{FZaIh@aFfI?
z!A+ONFAH8x$!WaUyUGeUq5mIN?i}8EkE$9=PLp>I-RU<K$v?tP-x0=+CrkZ|Fn6qI
zF*bxnNE4f}#iDAHsA3Bp63@PUVguwEf%M3|A5MI~BD#R2F_@@4^EWDuP*lSGU4JX8
zbRnGGFsf!`3E!8OsG^lHld&01JdEXjGp@S4(eiCPRGDXfcRyAx9<@ZEYz(SgJ>rn{
zsUYx3-^W{82&}T{LTGynXQ65&xx<~HUlJRZ++U&*8*PGI$ZH=kENtPtbfKFT=yu<V
zopz?}a-9_!)>;owa-ihfQ^I{onqA;AEE)qNjK-LL?HEOo?Hlc)Kf`SHYTDS4{(oF`
z)YMKO!=U4{IuNF)>|R5NR=6b>aR)H9gS|+3yTHvb-rt2zXItT3r_l`7t0nmsz2V;P
zh6jKJv!SN$HR9%W4R+(UkfgcFx>r;7%9Ax^KmP2oavJ_dsGQZ{(Nhps9;?FY%A=?L
z!`hQa%lB85+wf)3*;nQUfQfo~o2{_XK?~Yj!En5J@+P%tvJ)CBmq7anB`TsaEXSO;
zE}La*(Y$Q-E=_J%Di|r$*m#KwZ`f5d4i^b*PK*b}PYVMuy$`E7Fn^K%8~_BBG0Cud
z&Fxl4$RMoRdWzTaqRMm{N5E5-p^+ew%72P7j(is<<~4*AODjaxlU2MS3Hc-;ktE};
z5rme@Fe1N4t@L9<31;vhT_8wc;U(}WX1xMyNE6ibqIJ{7hGiT+&a+twh>XC)lf?1f
zjghr}1wn2d-eLrKuY%SY0LkK2`h0W^)mfVc_5;reDEdx6bCOBG0*rUGrqgFf%otbk
zF)6u4orniR#*B7^koeC~B34>b>no!*q#nmew}-$-dYwaIQTZ=4NH5zKfX^1@rWLxR
z<D1q`tbseo!Wt9Y!8-TdF`d;`+QFOIHfegcolEpSk2j`3DYT~(y)W_@4#vp#vlu4>
z^zdFnv;7&i_2-RJdS#T^WE4UVc0>{cD)ShGLz6R^6FH#AoIlU#k=HWGMiuBG?_l9t
z5kH`oMhr<%DSv<p?EWFOX>)<kXL~DmOLq77B7SSVdO^d5ySlPgTUl56we{794_5qJ
zU)*Alkz`!tA5--qRewSisk;0TRX?WcPpP^~)iPE0sG|Lm4^Sz0Bf*xYa+*LCxdJM*
zt3o}<CY&cQbuti;SN?_h8|j#~%%p)&Q6gHpVO>cf$1yLCpy%qjBK7UykNg-P9)}+J
z2?&$(B;aYGj04YG=gn$}KHn5|NH1&jK~fZUSsZVkL81vlda>5wE}~5xM;%f4DyKz<
zdi5+jFCLJ1T6u8E?sdA&o!%l^Edj5qvO)JBRt_uQt$cjru$uTT1UGciFU^jmGkOQr
zS3ytTJ#k+FA6noT_+-zMC6!CV5cviik~EPNhbWfh1&zRG$SwM%nUiowq-apw-_o&Q
zb#{MCb<T}cQ8|0E+t#{Bc78%3mGw1EwkrB?H~eU)wdd1_2ec#xS(B`HZ(+oRGyR(5
zyGZ?+7mv^&bFpLZ6$yxUDD;81**4@mFeS_*?Q_DMXtR>(yDu{Agdu#2;efTX971`_
zp`j7ZsXWcjXNi)l_>3|)H8&V_m<4x1X6R84c`?;<<w1T%y;4b-R8#_#K8lVp6IZir
z_U1->cojG3>*R$q4+si*#oyr-8^`ckPB<*oq0xddFvB8N&!+Xnq<gf%4;@lzza}Hs
zFpebCD4$TZP8C625--W?M3_q}_B9A*zC_WphnsxGDpI^e(Vdp%A-<%HeZ6->TxO5%
zjV=!9VR!ghc5xif^*l-RCQ5n@73lvRBQeaZ#f`0?SK^;ace8@U)sQRr*5T#PG4D6R
ziz!CJbiRA@*BIt3tnkc7urgGo0y?I!(#Fl9h!N1eFya1wF<`%A$C8)(c8JJh;E^AK
z`*-_h2-kR!qY$qRPX{h?KF;)W@Y=k7zE29~4yGVtuCR}36l($Lr~cH5Wg6s4v!XdV
zQs?NhXKy?oP)-7gqpg6`E9@M|hnSPft*<WE*8Gp3)_8wao^U}2+;SEjs)h4lSbmO2
zBTgx^rMwM4iT>akP|ku?v!0Vdl<4_{bdzrW4N63eGOVIiGD}tgI?CFaF!fdKw9%Og
zYv<RDDqgYbDx;xiVoIcN$0*}=os>;aW>VR{sZl3^d`Vv;DM|FKAc=ZL%V@E5MdY)M
z8}+XWC6au^$fV~v9kvz+yS*8z$29FsMvQctz8Dcqks)kzJl(c9{Q#ev(+`;#kv1}A
zSxP?$>VbYcuDlDm^vD#DG=>G-<%V}Gmv*gp6JbCr<dj1kL~$j_8di4qakGI`Dbl48
zQa=^28;Oc%D${IjHe@xgJ=zT9^<hnAA1tpgufel*)>qcn$8r@urp;QeN~2tb&qJuW
zF-mI}sDqs3(AP^{*7LRgH^dSxs5)O!FpCxgF)?$b{KL(n-&Fa@ryThdXA%uxnXEWY
zk`-)03qv1+hGxb7Iw8aKxQ_v4I(V&*p^ejo)6W(v3B*$lO`0iz3H>7q%keNxz<#6V
zO{&9s8_tMZ%A+qPd4zNh!nfBYblhP==PZxF<!OX+mdGNUBX6b1T+7u?36)RTuOR+?
z4c0M>Asu(neH4^cXGRQ_oAij|r*nL^*YZ>49ybn?)UQKn-&aI%lEDGK^G_&=7o+nq
zCA0SyLlcSLh{Z#>{zh4tSMc(+QP4lrw6ufCiDydX@~#iTGWU9qgycrj+TQ4qkW&8x
zeTj|=IQ+?bo|ltEZ=$3Z?JZs36f0dU15QY9TL0?fr=|3C$iyd7pghEl20$2aH@d3N
zgfvB;3!$B9BT0a#AK@u^#!N|`x-k+D08@{_l%8oY{W-xW(LxWce3I#%ijl@RF_Y*i
zgCmG~mf`Kz`^p>ce(xPdS>zoPcdCL@r{{x8>rC!95WPQK2L`{foBTP}SS@O|p@FvC
zpsGX_ISG<BL1&P5s6}ZYnQpJjH1_s7M>xaSR|P%_c^0&`w&Ms;TS3qB5%*v5Br2oA
zk%nE;1xFis%hLx2<Rhl18?+&GksgvX-bac44wYu6Fc)E^E*ipy6}y5hfLDbix^tWn
znPXS?A*UXvKH2d%-ZCgb8GDqV1E3(~?&sn>+F9grOp$+yqd$L&DO1QI4Jidsrf^z1
zqD;p{!MW!4jF{$eN(@ZJn%u@LC50QD2I_tAm`*w?BG`+!E75j$zaz*h!Y?Mn2(7{q
zV&w-RV({B+A$jY23-^a<dF<tf$)11j`G9jpr?gLqGe!VT>1yPgMuhk2`{dJ|R;ap*
zYWTu;s6Cq)-^ZUlSb3x}4_2P6f6Cjh+)j53UZrwcAtpz{pzKx{vufg42*!~;8f8Vd
zP;R5#(Z^y_wWp8$`;V5_)>hV(Lq~JNP>6lY3D1m#>EBVbL|AliG;htrX)N<U4s2N>
zMo7q<NmZ8cKQ%W0hMO1xtiWj(<?z~^J8V;C(M9f%>fDLL@<6#mJilc8#2J_aYhVu?
zex4oZcf5XP-Wa&?RFa7#87ahMgm9(~-^WF6G0SGG2r8cw!IO7IZr(_r%;1TKBuox@
zSyxPrJt^Ty9+H$pfbELuu_v=}84^Wvn-fKrD4kQBsr9a%BMfwi4tc=;!{>wd00sKo
zf^sCojizIOQ=pA@E|B{@M?6L@e~WaDg||ge$j9w1;;<x6565KTxghsJk_S{VT9B5?
zh^m+>jzjV!4xg-k_7rCuGoL?vu>Pt4WMx%T<X@m04*wq3{I&JvPgeW~Yb;-qpqIr<
z9#N}D)pJyJEBb(?Oj7VK2dI+tfwBbTP~64=@*R{9IFGGYH6d|inEo4u)941$Ck_H|
z`d4t~fk6ex-bKqY7jbZ{OYgEa7QcIItlT6(o0#$qOt*v$95=|tWK!R%JtmO<7vfE7
zk4s+Jds`B{dWnxz)ExgAz|wx)!ryHWNC25shHq7NcyHl!YG%A6JaT2Dk|BmtUfh+<
z?WHher%TQ1p3SZdmZ0aGtt|~pdHS0IjYHboV0rs1jKI;Za)GHhK-LmL2RNKL<#<<1
z4+jlwxkeo27Nud7_mO7zj|o4bE2tbqsbv_Wf}M8?wpS?6l&_TEE?+HQDqs6Qd-GgO

literal 0
HcmV?d00001

diff --git a/pycls/models/anynet.py b/pycls/models/anynet.py
new file mode 100644
index 0000000..a70d58d
--- /dev/null
+++ b/pycls/models/anynet.py
@@ -0,0 +1,380 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""AnyNet models."""
+
+import pycls.utils.logging as lu
+import pycls.utils.net as nu
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+logger = lu.get_logger(__name__)
+
+
+def get_stem_fun(stem_type):
+    """Retrives the stem function by name."""
+    stem_funs = {
+        "res_stem_cifar": ResStemCifar,
+        "res_stem_in": ResStemIN,
+        "simple_stem_in": SimpleStemIN,
+    }
+    assert stem_type in stem_funs.keys(), "Stem type '{}' not supported".format(
+        stem_type
+    )
+    return stem_funs[stem_type]
+
+
+def get_block_fun(block_type):
+    """Retrieves the block function by name."""
+    block_funs = {
+        "vanilla_block": VanillaBlock,
+        "res_basic_block": ResBasicBlock,
+        "res_bottleneck_block": ResBottleneckBlock,
+    }
+    assert block_type in block_funs.keys(), "Block type '{}' not supported".format(
+        block_type
+    )
+    return block_funs[block_type]
+
+
+class AnyHead(nn.Module):
+    """AnyNet head."""
+
+    def __init__(self, w_in, nc):
+        super(AnyHead, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(w_in, nc, bias=True)
+
+    def forward(self, x):
+        x = self.avg_pool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+        return x
+
+
+class VanillaBlock(nn.Module):
+    """Vanilla block: [3x3 conv, BN, Relu] x2"""
+
+    def __init__(self, w_in, w_out, stride, bm=None, gw=None, se_r=None):
+        assert (
+            bm is None and gw is None and se_r is None
+        ), "Vanilla block does not support bm, gw, and se_r options"
+        super(VanillaBlock, self).__init__()
+        self._construct(w_in, w_out, stride)
+
+    def _construct(self, w_in, w_out, stride):
+        # 3x3, BN, ReLU
+        self.a = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=stride, padding=1, bias=False
+        )
+        self.a_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 3x3, BN, ReLU
+        self.b = nn.Conv2d(w_out, w_out, kernel_size=3, stride=1, padding=1, bias=False)
+        self.b_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.b_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class BasicTransform(nn.Module):
+    """Basic transformation: [3x3 conv, BN, Relu] x2"""
+
+    def __init__(self, w_in, w_out, stride):
+        super(BasicTransform, self).__init__()
+        self._construct(w_in, w_out, stride)
+
+    def _construct(self, w_in, w_out, stride):
+        # 3x3, BN, ReLU
+        self.a = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=stride, padding=1, bias=False
+        )
+        self.a_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 3x3, BN
+        self.b = nn.Conv2d(w_out, w_out, kernel_size=3, stride=1, padding=1, bias=False)
+        self.b_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.b_bn.final_bn = True
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class ResBasicBlock(nn.Module):
+    """Residual basic block: x + F(x), F = basic transform"""
+
+    def __init__(self, w_in, w_out, stride, bm=None, gw=None, se_r=None):
+        assert (
+            bm is None and gw is None and se_r is None
+        ), "Basic transform does not support bm, gw, and se_r options"
+        super(ResBasicBlock, self).__init__()
+        self._construct(w_in, w_out, stride)
+
+    def _add_skip_proj(self, w_in, w_out, stride):
+        self.proj = nn.Conv2d(
+            w_in, w_out, kernel_size=1, stride=stride, padding=0, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+
+    def _construct(self, w_in, w_out, stride):
+        # Use skip connection with projection if shape changes
+        self.proj_block = (w_in != w_out) or (stride != 1)
+        if self.proj_block:
+            self._add_skip_proj(w_in, w_out, stride)
+        self.f = BasicTransform(w_in, w_out, stride)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        if self.proj_block:
+            x = self.bn(self.proj(x)) + self.f(x)
+        else:
+            x = x + self.f(x)
+        x = self.relu(x)
+        return x
+
+
+class SE(nn.Module):
+    """Squeeze-and-Excitation (SE) block"""
+
+    def __init__(self, w_in, w_se):
+        super(SE, self).__init__()
+        self._construct(w_in, w_se)
+
+    def _construct(self, w_in, w_se):
+        # AvgPool
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        # FC, Activation, FC, Sigmoid
+        self.f_ex = nn.Sequential(
+            nn.Conv2d(w_in, w_se, kernel_size=1, bias=True),
+            nn.ReLU(inplace=cfg.MEM.RELU_INPLACE),
+            nn.Conv2d(w_se, w_in, kernel_size=1, bias=True),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        return x * self.f_ex(self.avg_pool(x))
+
+
+class BottleneckTransform(nn.Module):
+    """Bottlenect transformation: 1x1, 3x3, 1x1"""
+
+    def __init__(self, w_in, w_out, stride, bm, gw, se_r):
+        super(BottleneckTransform, self).__init__()
+        self._construct(w_in, w_out, stride, bm, gw, se_r)
+
+    def _construct(self, w_in, w_out, stride, bm, gw, se_r):
+        # Compute the bottleneck width
+        w_b = int(round(w_out * bm))
+        # Compute the number of groups
+        num_gs = w_b // gw
+        # 1x1, BN, ReLU
+        self.a = nn.Conv2d(w_in, w_b, kernel_size=1, stride=1, padding=0, bias=False)
+        self.a_bn = nn.BatchNorm2d(w_b, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 3x3, BN, ReLU
+        self.b = nn.Conv2d(
+            w_b, w_b, kernel_size=3, stride=stride, padding=1, groups=num_gs, bias=False
+        )
+        self.b_bn = nn.BatchNorm2d(w_b, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.b_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # Squeeze-and-Excitation (SE)
+        if se_r:
+            w_se = int(round(w_in * se_r))
+            self.se = SE(w_b, w_se)
+        # 1x1, BN
+        self.c = nn.Conv2d(w_b, w_out, kernel_size=1, stride=1, padding=0, bias=False)
+        self.c_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.c_bn.final_bn = True
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class ResBottleneckBlock(nn.Module):
+    """Residual bottleneck block: x + F(x), F = bottleneck transform"""
+
+    def __init__(self, w_in, w_out, stride, bm=1.0, gw=1, se_r=None):
+        super(ResBottleneckBlock, self).__init__()
+        self._construct(w_in, w_out, stride, bm, gw, se_r)
+
+    def _add_skip_proj(self, w_in, w_out, stride):
+        self.proj = nn.Conv2d(
+            w_in, w_out, kernel_size=1, stride=stride, padding=0, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+
+    def _construct(self, w_in, w_out, stride, bm, gw, se_r):
+        # Use skip connection with projection if shape changes
+        self.proj_block = (w_in != w_out) or (stride != 1)
+        if self.proj_block:
+            self._add_skip_proj(w_in, w_out, stride)
+        self.f = BottleneckTransform(w_in, w_out, stride, bm, gw, se_r)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        if self.proj_block:
+            x = self.bn(self.proj(x)) + self.f(x)
+        else:
+            x = x + self.f(x)
+        x = self.relu(x)
+        return x
+
+
+class ResStemCifar(nn.Module):
+    """ResNet stem for CIFAR."""
+
+    def __init__(self, w_in, w_out):
+        super(ResStemCifar, self).__init__()
+        self._construct(w_in, w_out)
+
+    def _construct(self, w_in, w_out):
+        # 3x3, BN, ReLU
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=1, padding=1, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class ResStemIN(nn.Module):
+    """ResNet stem for ImageNet."""
+
+    def __init__(self, w_in, w_out):
+        super(ResStemIN, self).__init__()
+        self._construct(w_in, w_out)
+
+    def _construct(self, w_in, w_out):
+        # 7x7, BN, ReLU, maxpool
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=7, stride=2, padding=3, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+        self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class SimpleStemIN(nn.Module):
+    """Simple stem for ImageNet."""
+
+    def __init__(self, in_w, out_w):
+        super(SimpleStemIN, self).__init__()
+        self._construct(in_w, out_w)
+
+    def _construct(self, in_w, out_w):
+        # 3x3, BN, ReLU
+        self.conv = nn.Conv2d(
+            in_w, out_w, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn = nn.BatchNorm2d(out_w, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class AnyStage(nn.Module):
+    """AnyNet stage (sequence of blocks w/ the same output shape)."""
+
+    def __init__(self, w_in, w_out, stride, d, block_fun, bm, gw, se_r):
+        super(AnyStage, self).__init__()
+        self._construct(w_in, w_out, stride, d, block_fun, bm, gw, se_r)
+
+    def _construct(self, w_in, w_out, stride, d, block_fun, bm, gw, se_r):
+        # Construct the blocks
+        for i in range(d):
+            # Stride and w_in apply to the first block of the stage
+            b_stride = stride if i == 0 else 1
+            b_w_in = w_in if i == 0 else w_out
+            # Construct the block
+            self.add_module(
+                "b{}".format(i + 1), block_fun(b_w_in, w_out, b_stride, bm, gw, se_r)
+            )
+
+    def forward(self, x):
+        for block in self.children():
+            x = block(x)
+        return x
+
+
+class AnyNet(nn.Module):
+    """AnyNet model."""
+
+    def __init__(self, **kwargs):
+        super(AnyNet, self).__init__()
+        if kwargs:
+            self._construct(
+                stem_type=kwargs["stem_type"],
+                stem_w=kwargs["stem_w"],
+                block_type=kwargs["block_type"],
+                ds=kwargs["ds"],
+                ws=kwargs["ws"],
+                ss=kwargs["ss"],
+                bms=kwargs["bms"],
+                gws=kwargs["gws"],
+                se_r=kwargs["se_r"],
+                nc=kwargs["nc"],
+            )
+        else:
+            self._construct(
+                stem_type=cfg.ANYNET.STEM_TYPE,
+                stem_w=cfg.ANYNET.STEM_W,
+                block_type=cfg.ANYNET.BLOCK_TYPE,
+                ds=cfg.ANYNET.DEPTHS,
+                ws=cfg.ANYNET.WIDTHS,
+                ss=cfg.ANYNET.STRIDES,
+                bms=cfg.ANYNET.BOT_MULS,
+                gws=cfg.ANYNET.GROUP_WS,
+                se_r=cfg.ANYNET.SE_R if cfg.ANYNET.SE_ON else None,
+                nc=cfg.MODEL.NUM_CLASSES,
+            )
+        self.apply(nu.init_weights)
+
+    def _construct(self, stem_type, stem_w, block_type, ds, ws, ss, bms, gws, se_r, nc):
+        logger.info("Constructing AnyNet: ds={}, ws={}".format(ds, ws))
+        # Generate dummy bot muls and gs for models that do not use them
+        bms = bms if bms else [1.0 for _d in ds]
+        gws = gws if gws else [1 for _d in ds]
+        # Group params by stage
+        stage_params = list(zip(ds, ws, ss, bms, gws))
+        # Construct the stem
+        stem_fun = get_stem_fun(stem_type)
+        self.stem = stem_fun(3, stem_w)
+        # Construct the stages
+        block_fun = get_block_fun(block_type)
+        prev_w = stem_w
+        for i, (d, w, s, bm, gw) in enumerate(stage_params):
+            self.add_module(
+                "s{}".format(i + 1), AnyStage(prev_w, w, s, d, block_fun, bm, gw, se_r)
+            )
+            prev_w = w
+        # Construct the head
+        self.head = AnyHead(w_in=prev_w, nc=nc)
+
+    def forward(self, x):
+        for module in self.children():
+            x = module(x)
+        return x
diff --git a/pycls/models/effnet.py b/pycls/models/effnet.py
new file mode 100644
index 0000000..e310a21
--- /dev/null
+++ b/pycls/models/effnet.py
@@ -0,0 +1,235 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""EfficientNet models."""
+
+import pycls.utils.logging as logging
+import pycls.utils.net as nu
+import torch
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+logger = logging.get_logger(__name__)
+
+
+class EffHead(nn.Module):
+    """EfficientNet head."""
+
+    def __init__(self, w_in, w_out, nc):
+        super(EffHead, self).__init__()
+        self._construct(w_in, w_out, nc)
+
+    def _construct(self, w_in, w_out, nc):
+        # 1x1, BN, Swish
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=1, stride=1, padding=0, bias=False
+        )
+        self.conv_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.conv_swish = Swish()
+        # AvgPool
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        # Dropout
+        if cfg.EN.DROPOUT_RATIO > 0.0:
+            self.dropout = nn.Dropout(p=cfg.EN.DROPOUT_RATIO)
+        # FC
+        self.fc = nn.Linear(w_out, nc, bias=True)
+
+    def forward(self, x):
+        x = self.conv_swish(self.conv_bn(self.conv(x)))
+        x = self.avg_pool(x)
+        x = x.view(x.size(0), -1)
+        x = self.dropout(x) if hasattr(self, "dropout") else x
+        x = self.fc(x)
+        return x
+
+
+class Swish(nn.Module):
+    """Swish activation function: x * sigmoid(x)"""
+
+    def __init__(self):
+        super(Swish, self).__init__()
+
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class SE(nn.Module):
+    """Squeeze-and-Excitation (SE) block w/ Swish."""
+
+    def __init__(self, w_in, w_se):
+        super(SE, self).__init__()
+        self._construct(w_in, w_se)
+
+    def _construct(self, w_in, w_se):
+        # AvgPool
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        # FC, Swish, FC, Sigmoid
+        self.f_ex = nn.Sequential(
+            nn.Conv2d(w_in, w_se, kernel_size=1, bias=True),
+            Swish(),
+            nn.Conv2d(w_se, w_in, kernel_size=1, bias=True),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        return x * self.f_ex(self.avg_pool(x))
+
+
+class MBConv(nn.Module):
+    """Mobile inverted bottleneck block w/ SE (MBConv)."""
+
+    def __init__(self, w_in, exp_r, kernel, stride, se_r, w_out):
+        super(MBConv, self).__init__()
+        self._construct(w_in, exp_r, kernel, stride, se_r, w_out)
+
+    def _construct(self, w_in, exp_r, kernel, stride, se_r, w_out):
+        # Expansion ratio is wrt the input width
+        self.exp = None
+        w_exp = int(w_in * exp_r)
+        # Include exp ops only if the exp ratio is different from 1
+        if w_exp != w_in:
+            # 1x1, BN, Swish
+            self.exp = nn.Conv2d(
+                w_in, w_exp, kernel_size=1, stride=1, padding=0, bias=False
+            )
+            self.exp_bn = nn.BatchNorm2d(w_exp, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+            self.exp_swish = Swish()
+        # 3x3 dwise, BN, Swish
+        self.dwise = nn.Conv2d(
+            w_exp,
+            w_exp,
+            kernel_size=kernel,
+            stride=stride,
+            groups=w_exp,
+            bias=False,
+            # Hacky padding to preserve res  (supports only 3x3 and 5x5)
+            padding=(1 if kernel == 3 else 2),
+        )
+        self.dwise_bn = nn.BatchNorm2d(w_exp, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.dwise_swish = Swish()
+        # Squeeze-and-Excitation (SE)
+        w_se = int(w_in * se_r)
+        self.se = SE(w_exp, w_se)
+        # 1x1, BN
+        self.lin_proj = nn.Conv2d(
+            w_exp, w_out, kernel_size=1, stride=1, padding=0, bias=False
+        )
+        self.lin_proj_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        # Skip connection if in and out shapes are the same (MN-V2 style)
+        self.has_skip = (stride == 1) and (w_in == w_out)
+
+    def forward(self, x):
+        f_x = x
+        # Expansion
+        if self.exp:
+            f_x = self.exp_swish(self.exp_bn(self.exp(f_x)))
+        # Depthwise
+        f_x = self.dwise_swish(self.dwise_bn(self.dwise(f_x)))
+        # SE
+        f_x = self.se(f_x)
+        # Linear projection
+        f_x = self.lin_proj_bn(self.lin_proj(f_x))
+        # Skip connection
+        if self.has_skip:
+            # Drop connect
+            if self.training and cfg.EN.DC_RATIO > 0.0:
+                f_x = nu.drop_connect(f_x, cfg.EN.DC_RATIO)
+            f_x = x + f_x
+        return f_x
+
+
+class EffStage(nn.Module):
+    """EfficientNet stage."""
+
+    def __init__(self, w_in, exp_r, kernel, stride, se_r, w_out, d):
+        super(EffStage, self).__init__()
+        self._construct(w_in, exp_r, kernel, stride, se_r, w_out, d)
+
+    def _construct(self, w_in, exp_r, kernel, stride, se_r, w_out, d):
+        # Construct the blocks
+        for i in range(d):
+            # Stride and input width apply to the first block of the stage
+            b_stride = stride if i == 0 else 1
+            b_w_in = w_in if i == 0 else w_out
+            # Construct the block
+            self.add_module(
+                "b{}".format(i + 1),
+                MBConv(b_w_in, exp_r, kernel, b_stride, se_r, w_out),
+            )
+
+    def forward(self, x):
+        for block in self.children():
+            x = block(x)
+        return x
+
+
+class StemIN(nn.Module):
+    """EfficientNet stem for ImageNet."""
+
+    def __init__(self, w_in, w_out):
+        super(StemIN, self).__init__()
+        self._construct(w_in, w_out)
+
+    def _construct(self, w_in, w_out):
+        # 3x3, BN, Swish
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.swish = Swish()
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class EffNet(nn.Module):
+    """EfficientNet model."""
+
+    def __init__(self):
+        assert cfg.TRAIN.DATASET in [
+            "imagenet"
+        ], "Training on {} is not supported".format(cfg.TRAIN.DATASET)
+        assert cfg.TEST.DATASET in [
+            "imagenet"
+        ], "Testing on {} is not supported".format(cfg.TEST.DATASET)
+        super(EffNet, self).__init__()
+        self._construct(
+            stem_w=cfg.EN.STEM_W,
+            ds=cfg.EN.DEPTHS,
+            ws=cfg.EN.WIDTHS,
+            exp_rs=cfg.EN.EXP_RATIOS,
+            se_r=cfg.EN.SE_R,
+            ss=cfg.EN.STRIDES,
+            ks=cfg.EN.KERNELS,
+            head_w=cfg.EN.HEAD_W,
+            nc=cfg.MODEL.NUM_CLASSES,
+        )
+        self.apply(nu.init_weights)
+
+    def _construct(self, stem_w, ds, ws, exp_rs, se_r, ss, ks, head_w, nc):
+        # Group params by stage
+        stage_params = list(zip(ds, ws, exp_rs, ss, ks))
+        logger.info("Constructing: EfficientNet-{}".format(stage_params))
+        # Construct the stem
+        self.stem = StemIN(3, stem_w)
+        prev_w = stem_w
+        # Construct the stages
+        for i, (d, w, exp_r, stride, kernel) in enumerate(stage_params):
+            self.add_module(
+                "s{}".format(i + 1), EffStage(prev_w, exp_r, kernel, stride, se_r, w, d)
+            )
+            prev_w = w
+        # Construct the head
+        self.head = EffHead(prev_w, head_w, nc)
+
+    def forward(self, x):
+        for module in self.children():
+            x = module(x)
+        return x
diff --git a/pycls/models/regnet.py b/pycls/models/regnet.py
new file mode 100644
index 0000000..46b0158
--- /dev/null
+++ b/pycls/models/regnet.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""RegNet models."""
+
+import numpy as np
+import pycls.utils.logging as lu
+from pycls.core.config import cfg
+from pycls.models.anynet import AnyNet
+
+
+logger = lu.get_logger(__name__)
+
+
+def quantize_float(f, q):
+    """Converts a float to closest non-zero int divisible by q."""
+    return int(round(f / q) * q)
+
+
+def adjust_ws_gs_comp(ws, bms, gs):
+    """Adjusts the compatibility of widths and groups."""
+    ws_bot = [int(w * b) for w, b in zip(ws, bms)]
+    gs = [min(g, w_bot) for g, w_bot in zip(gs, ws_bot)]
+    ws_bot = [quantize_float(w_bot, g) for w_bot, g in zip(ws_bot, gs)]
+    ws = [int(w_bot / b) for w_bot, b in zip(ws_bot, bms)]
+    return ws, gs
+
+
+def get_stages_from_blocks(ws, rs):
+    """Gets ws/ds of network at each stage from per block values."""
+    ts_temp = zip(ws + [0], [0] + ws, rs + [0], [0] + rs)
+    ts = [w != wp or r != rp for w, wp, r, rp in ts_temp]
+    s_ws = [w for w, t in zip(ws, ts[:-1]) if t]
+    s_ds = np.diff([d for d, t in zip(range(len(ts)), ts) if t]).tolist()
+    return s_ws, s_ds
+
+
+def generate_regnet(w_a, w_0, w_m, d, q=8):
+    """Generates per block ws from RegNet parameters."""
+    assert w_a >= 0 and w_0 > 0 and w_m > 1 and w_0 % q == 0
+    ws_cont = np.arange(d) * w_a + w_0
+    ks = np.round(np.log(ws_cont / w_0) / np.log(w_m))
+    ws = w_0 * np.power(w_m, ks)
+    ws = np.round(np.divide(ws, q)) * q
+    num_stages, max_stage = len(np.unique(ws)), ks.max() + 1
+    ws, ws_cont = ws.astype(int).tolist(), ws_cont.tolist()
+    return ws, num_stages, max_stage, ws_cont
+
+
+class RegNet(AnyNet):
+    """RegNet model."""
+
+    def __init__(self):
+        # Generate RegNet ws per block
+        b_ws, num_s, _, _ = generate_regnet(
+            cfg.REGNET.WA, cfg.REGNET.W0, cfg.REGNET.WM, cfg.REGNET.DEPTH
+        )
+        # Convert to per stage format
+        ws, ds = get_stages_from_blocks(b_ws, b_ws)
+        # Generate group widths and bot muls
+        gws = [cfg.REGNET.GROUP_W for _ in range(num_s)]
+        bms = [cfg.REGNET.BOT_MUL for _ in range(num_s)]
+        # Adjust the compatibility of ws and gws
+        ws, gws = adjust_ws_gs_comp(ws, bms, gws)
+        # Use the same stride for each stage
+        ss = [cfg.REGNET.STRIDE for _ in range(num_s)]
+        # Use SE for RegNetY
+        se_r = cfg.REGNET.SE_R if cfg.REGNET.SE_ON else None
+        # Construct the model
+        kwargs = {
+            "stem_type": cfg.REGNET.STEM_TYPE,
+            "stem_w": cfg.REGNET.STEM_W,
+            "block_type": cfg.REGNET.BLOCK_TYPE,
+            "ss": ss,
+            "ds": ds,
+            "ws": ws,
+            "bms": bms,
+            "gws": gws,
+            "se_r": se_r,
+            "nc": cfg.MODEL.NUM_CLASSES,
+        }
+        super(RegNet, self).__init__(**kwargs)
diff --git a/pycls/models/resnet.py b/pycls/models/resnet.py
new file mode 100644
index 0000000..f86482c
--- /dev/null
+++ b/pycls/models/resnet.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""ResNe(X)t models."""
+
+import pycls.utils.logging as lu
+import pycls.utils.net as nu
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+logger = lu.get_logger(__name__)
+
+
+# Stage depths for ImageNet models
+_IN_STAGE_DS = {50: (3, 4, 6, 3), 101: (3, 4, 23, 3), 152: (3, 8, 36, 3)}
+
+
+def get_trans_fun(name):
+    """Retrieves the transformation function by name."""
+    trans_funs = {
+        "basic_transform": BasicTransform,
+        "bottleneck_transform": BottleneckTransform,
+    }
+    assert (
+        name in trans_funs.keys()
+    ), "Transformation function '{}' not supported".format(name)
+    return trans_funs[name]
+
+
+class ResHead(nn.Module):
+    """ResNet head."""
+
+    def __init__(self, w_in, nc):
+        super(ResHead, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.fc = nn.Linear(w_in, nc, bias=True)
+
+    def forward(self, x):
+        x = self.avg_pool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+        return x
+
+
+class BasicTransform(nn.Module):
+    """Basic transformation: 3x3, 3x3"""
+
+    def __init__(self, w_in, w_out, stride, w_b=None, num_gs=1):
+        assert (
+            w_b is None and num_gs == 1
+        ), "Basic transform does not support w_b and num_gs options"
+        super(BasicTransform, self).__init__()
+        self._construct(w_in, w_out, stride)
+
+    def _construct(self, w_in, w_out, stride):
+        # 3x3, BN, ReLU
+        self.a = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=stride, padding=1, bias=False
+        )
+        self.a_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 3x3, BN
+        self.b = nn.Conv2d(w_out, w_out, kernel_size=3, stride=1, padding=1, bias=False)
+        self.b_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.b_bn.final_bn = True
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class BottleneckTransform(nn.Module):
+    """Bottleneck transformation: 1x1, 3x3, 1x1"""
+
+    def __init__(self, w_in, w_out, stride, w_b, num_gs):
+        super(BottleneckTransform, self).__init__()
+        self._construct(w_in, w_out, stride, w_b, num_gs)
+
+    def _construct(self, w_in, w_out, stride, w_b, num_gs):
+        # MSRA -> stride=2 is on 1x1; TH/C2 -> stride=2 is on 3x3
+        (str1x1, str3x3) = (stride, 1) if cfg.RESNET.STRIDE_1X1 else (1, stride)
+        # 1x1, BN, ReLU
+        self.a = nn.Conv2d(
+            w_in, w_b, kernel_size=1, stride=str1x1, padding=0, bias=False
+        )
+        self.a_bn = nn.BatchNorm2d(w_b, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.a_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 3x3, BN, ReLU
+        self.b = nn.Conv2d(
+            w_b, w_b, kernel_size=3, stride=str3x3, padding=1, groups=num_gs, bias=False
+        )
+        self.b_bn = nn.BatchNorm2d(w_b, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.b_relu = nn.ReLU(inplace=cfg.MEM.RELU_INPLACE)
+        # 1x1, BN
+        self.c = nn.Conv2d(w_b, w_out, kernel_size=1, stride=1, padding=0, bias=False)
+        self.c_bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.c_bn.final_bn = True
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class ResBlock(nn.Module):
+    """Residual block: x + F(x)"""
+
+    def __init__(self, w_in, w_out, stride, trans_fun, w_b=None, num_gs=1):
+        super(ResBlock, self).__init__()
+        self._construct(w_in, w_out, stride, trans_fun, w_b, num_gs)
+
+    def _add_skip_proj(self, w_in, w_out, stride):
+        self.proj = nn.Conv2d(
+            w_in, w_out, kernel_size=1, stride=stride, padding=0, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+
+    def _construct(self, w_in, w_out, stride, trans_fun, w_b, num_gs):
+        # Use skip connection with projection if shape changes
+        self.proj_block = (w_in != w_out) or (stride != 1)
+        if self.proj_block:
+            self._add_skip_proj(w_in, w_out, stride)
+        self.f = trans_fun(w_in, w_out, stride, w_b, num_gs)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def forward(self, x):
+        if self.proj_block:
+            x = self.bn(self.proj(x)) + self.f(x)
+        else:
+            x = x + self.f(x)
+        x = self.relu(x)
+        return x
+
+
+class ResStage(nn.Module):
+    """Stage of ResNet."""
+
+    def __init__(self, w_in, w_out, stride, d, w_b=None, num_gs=1):
+        super(ResStage, self).__init__()
+        self._construct(w_in, w_out, stride, d, w_b, num_gs)
+
+    def _construct(self, w_in, w_out, stride, d, w_b, num_gs):
+        # Construct the blocks
+        for i in range(d):
+            # Stride and w_in apply to the first block of the stage
+            b_stride = stride if i == 0 else 1
+            b_w_in = w_in if i == 0 else w_out
+            # Retrieve the transformation function
+            trans_fun = get_trans_fun(cfg.RESNET.TRANS_FUN)
+            # Construct the block
+            res_block = ResBlock(b_w_in, w_out, b_stride, trans_fun, w_b, num_gs)
+            self.add_module("b{}".format(i + 1), res_block)
+
+    def forward(self, x):
+        for block in self.children():
+            x = block(x)
+        return x
+
+
+class ResStem(nn.Module):
+    """Stem of ResNet."""
+
+    def __init__(self, w_in, w_out):
+        assert (
+            cfg.TRAIN.DATASET == cfg.TEST.DATASET
+        ), "Train and test dataset must be the same for now"
+        super(ResStem, self).__init__()
+        if "cifar" in cfg.TRAIN.DATASET:
+            self._construct_cifar(w_in, w_out)
+        else:
+            self._construct_imagenet(w_in, w_out)
+
+    def _construct_cifar(self, w_in, w_out):
+        # 3x3, BN, ReLU
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=3, stride=1, padding=1, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+
+    def _construct_imagenet(self, w_in, w_out):
+        # 7x7, BN, ReLU, maxpool
+        self.conv = nn.Conv2d(
+            w_in, w_out, kernel_size=7, stride=2, padding=3, bias=False
+        )
+        self.bn = nn.BatchNorm2d(w_out, eps=cfg.BN.EPS, momentum=cfg.BN.MOM)
+        self.relu = nn.ReLU(cfg.MEM.RELU_INPLACE)
+        self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+    def forward(self, x):
+        for layer in self.children():
+            x = layer(x)
+        return x
+
+
+class ResNet(nn.Module):
+    """ResNet model."""
+
+    def __init__(self):
+        assert cfg.TRAIN.DATASET in [
+            "cifar10",
+            "imagenet",
+        ], "Training ResNet on {} is not supported".format(cfg.TRAIN.DATASET)
+        assert cfg.TEST.DATASET in [
+            "cifar10",
+            "imagenet",
+        ], "Testing ResNet on {} is not supported".format(cfg.TEST.DATASET)
+        super(ResNet, self).__init__()
+        if "cifar" in cfg.TRAIN.DATASET:
+            self._construct_cifar()
+        else:
+            self._construct_imagenet()
+        self.apply(nu.init_weights)
+
+    def _construct_cifar(self):
+        assert (
+            cfg.MODEL.DEPTH - 2
+        ) % 6 == 0, "Model depth should be of the format 6n + 2 for cifar"
+        logger.info("Constructing: ResNet-{}".format(cfg.MODEL.DEPTH))
+        # Each stage has the same number of blocks for cifar
+        d = int((cfg.MODEL.DEPTH - 2) / 6)
+        # Stem: (N, 3, 32, 32) -> (N, 16, 32, 32)
+        self.stem = ResStem(w_in=3, w_out=16)
+        # Stage 1: (N, 16, 32, 32) -> (N, 16, 32, 32)
+        self.s1 = ResStage(w_in=16, w_out=16, stride=1, d=d)
+        # Stage 2: (N, 16, 32, 32) -> (N, 32, 16, 16)
+        self.s2 = ResStage(w_in=16, w_out=32, stride=2, d=d)
+        # Stage 3: (N, 32, 16, 16) -> (N, 64, 8, 8)
+        self.s3 = ResStage(w_in=32, w_out=64, stride=2, d=d)
+        # Head: (N, 64, 8, 8) -> (N, num_classes)
+        self.head = ResHead(w_in=64, nc=cfg.MODEL.NUM_CLASSES)
+
+    def _construct_imagenet(self):
+        logger.info(
+            "Constructing: ResNe(X)t-{}-{}x{}, {}".format(
+                cfg.MODEL.DEPTH,
+                cfg.RESNET.NUM_GROUPS,
+                cfg.RESNET.WIDTH_PER_GROUP,
+                cfg.RESNET.TRANS_FUN,
+            )
+        )
+        # Retrieve the number of blocks per stage
+        (d1, d2, d3, d4) = _IN_STAGE_DS[cfg.MODEL.DEPTH]
+        # Compute the initial bottleneck width
+        num_gs = cfg.RESNET.NUM_GROUPS
+        w_b = cfg.RESNET.WIDTH_PER_GROUP * num_gs
+        # Stem: (N, 3, 224, 224) -> (N, 64, 56, 56)
+        self.stem = ResStem(w_in=3, w_out=64)
+        # Stage 1: (N, 64, 56, 56) -> (N, 256, 56, 56)
+        self.s1 = ResStage(w_in=64, w_out=256, stride=1, d=d1, w_b=w_b, num_gs=num_gs)
+        # Stage 2: (N, 256, 56, 56) -> (N, 512, 28, 28)
+        self.s2 = ResStage(
+            w_in=256, w_out=512, stride=2, d=d2, w_b=w_b * 2, num_gs=num_gs
+        )
+        # Stage 3: (N, 512, 56, 56) -> (N, 1024, 14, 14)
+        self.s3 = ResStage(
+            w_in=512, w_out=1024, stride=2, d=d3, w_b=w_b * 4, num_gs=num_gs
+        )
+        # Stage 4: (N, 1024, 14, 14) -> (N, 2048, 7, 7)
+        self.s4 = ResStage(
+            w_in=1024, w_out=2048, stride=2, d=d4, w_b=w_b * 8, num_gs=num_gs
+        )
+        # Head: (N, 2048, 7, 7) -> (N, num_classes)
+        self.head = ResHead(w_in=2048, nc=cfg.MODEL.NUM_CLASSES)
+
+    def forward(self, x):
+        for module in self.children():
+            x = module(x)
+        return x
diff --git a/pycls/utils/__init__.py b/pycls/utils/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/pycls/utils/__pycache__/__init__.cpython-37.pyc b/pycls/utils/__pycache__/__init__.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..96f4f6f1bcfdd4536461b9519cdfa59dd4615e7d
GIT binary patch
literal 138
zcmZ?b<>g`k0^5egaUl9Jh=2h`Aj1KOi&=m~3PUi1CZpd<h9ZzKg7{^upOK%Ns$W%^
zl9E?hte=utl31KtqF+#%oKp-XOG`3=?D+W1yv&mLc)fzkTO2mI`6;D2sdga4J_9iW
E0PT7puK)l5

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/benchmark.cpython-37.pyc b/pycls/utils/__pycache__/benchmark.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c0ea13df6940d93448da99a37e90e03357496067
GIT binary patch
literal 1877
zcmZWp&2HO95aup_B~fmyxK4o<jSg)Z0fpQ{kW*XKaU9eK8-e2>0FjiSxRgwqBJJfW
zwgjpOpj`6?J;>;(Z_-C7_S#ckp{LF)X+=ROa5%gBX?JJ7Z{|_E-PGXu>CK<N{R7^=
zIJkLqaDIWe@W3$5(->hiC1Xl8qK$RWAk~cNnXsa#*4XxJHD{#GxZyR37MaWDwZ=?l
zz10@PYcZQSZ#B<l4c3Hjo6%=luQmJT$s`L!oM%OMnDg!+%EHrez|Z1r)XgGsne(%v
z{~m6k9rcLR!{G=%=1Dw`ct~*?9s^#y@gBUz?_f%8sa?ZOuf7sQ&_qv+1$nH!a_6+v
zO0pnCyV!oA2{W;lHsXY{(B}Hp!y4UC=zD<HF$MzrRsZ_%4_Im8%EjaLeM&uXAu2Na
zVJ{z_Pei1Gx(qn$f-kxPuHC(#<=qhv*!SJ&b&z()Ig5JzURxSClqL_dQN%Gc*?4-q
zzvrJE9PZ23i{0bH=TH5Ell^09ikye1(g-IkkgZ}0DdTw-&mxXlkj~Nb!wRU2IF!1`
zL+J*uBOZ(*AFt3eq>0C-(uy;<W+5AXIN^RQBJP=p;D2en1AG1VVCOU+M?14AW7)LW
zVSxyWNbH<X!?dXFM8x2)i0hwE<yKgQ?+-713|yUh3-&9%0HYZW?GTqXsYN=(fnO6x
zgm1xGh-Bi^fERs5>fhk3NbP?_>eq7v=%EWTH%n5QrM{qZtF)G;q6cZV5_{<=nj1i~
zEzmkwgL;>yBKsJ2u`5bH-R;8;fN_-|oElr7vxaag#Kp_?lS?~kmrl|iqUg!&?kDmX
z1mRhw6UctQw*egUAkMf6fh#d^3q2c*bZ)@h!8uG6$5K}8^A>t8nl_pZFuex<3_j9H
z^P=#FlZ+!5(n^Ctloosw2j)4CSRoCBcn<Q0GVq*1?I}4_<~HIxYZTX*doW+z2czjo
za7`<a3^db<<|c7!PZ9nlU{_U)!C-uf@HgNHEg=b=Q=u;n@S$26DQQWsjkzxL#IEMR
za~OR`fj~WJFr69i$XqXtm2Ze9P>QlJMGGsrgw=)d5@Z5maF=ZmjaAzC-hqDghj9)4
zfmuM5J=<Op1?dUnB2p#Xe~)#BeTa$adGrK4H_J1BkmeyauAbp-BEHgoolCD%O(CtS
zE+D@j&DB}-KK~e3PapL*s-|F`9Pb_+N&Dc?e|F&QD-B2kx&~7ck<CZDCwsrDHEBXi
z$O~!4nRxi15(?HB$Gl6LPUJ?KkNl*7yB8u5MbA+};dfx&v*WQ}U{gV{a4cNUU9V@)
zzKP7f5ec@8HI1y=K%G3LseQmyZ9sR(7TKb%D*g^NC?)vg-vI(D!1jHXhrTawt8UX*
zeWni=9zlwwsfN<7$NXD78YLt*D|ndms2}FpFdnJI0VUY9>TQ&|t9EH;xsUx=wI7ul
ib%Qm*tFV;nogblZPSWUSJgLCiGL+IG2L0BybN7Fa%-%Qv

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/checkpoint.cpython-37.pyc b/pycls/utils/__pycache__/checkpoint.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..44434ad469c82e9fb7a099608fa29c06144c9d70
GIT binary patch
literal 2798
zcmZuzTW{Mo6ecN3krT(cv}@O*1JvDGV}sM(fTGBXb!oZ<+tMOwhr!q;s3H@~u_TY8
zi^i3KVd?80_Pz%Pd)i;(>wp3M3w_znp&HxCCP0T2Ns;F}-?{Mf#l?97&!<2C@#EhM
zg#3jcCyxW2&)_Ydf)J!l1Qo1D+tju!br_)}<!zT%GhW++8K*aw`fVS^c)KRt_Pp@K
z+;h@i5Wc8ETNm?U0op}T7mLs~L_;h=yCjyy3bf1OyjT_Ip40Y<V7sLG+VJMRK^E&I
z&x$~IA{}(1O!Q<>L{E}zKY)**mq!Bo`CbrrWPH%ilS~(_7oa!UZc^jKd;9HrnhV(r
zWj~KQp;?$&7~{H=heKxr-eMVqCS5va+Uaub9-}sNxV9s;O5_uX+DJG$j!?H?FIPdK
zbA@`$?A>a5hUJCf{YZBV?}BEE!Omvx{-cNC-K`x{Lm6)0`DXJ^qG~W@#mi6DI(aJB
zhKE9AhsBzRbW}*a)<2AU#Va|`3G}Cjwfcu<bzkamIuwdTT|_i^CSeLVCNaf6JUBLi
zdj&oO9TQDUa?D0lv#v7+*hi$K7CIYVKMfrb1v(F+8N9(>(vz*>;<M4STQ@!yBTz|q
zzyuzcCylhJTcF{Coy~h&-zhv?xhOmos9cj7*HUD77Nv4(5QUCv8mHl*(gp59Se+_l
zY{I1j9}{pN!-pnEWMn<{Ie;89fJyI?BMMNRk{*HQULqr=X_t@Pl9bGX=EL>Vpwqn2
zfs(Pz^epV_on*fwVan<rWP|jvRL$0ZOuY+=6t9EWCnfn|8O{dt9(anE^f&es->t}R
zQnd(cz!t1^!_d7mR(LtT-b)G{=V|};ySO0|>5<_zG)^xnW*wlljvm@8ZbRqfpBMkY
z)AX#pEeB=}=fG8rSL8}dF(swxn8s;$QfS%s`7`H)|Hsz3|54~oRM-bWn8(gjpbgD2
zp$&Z#Jz@jiq6MO&4QOFw2PrEq2)au&7fd*oNFb{s@(9M8eAv1xwNz=6L4+oI7^F%D
z5xi3vMG!rSl3w%}eD&Tbju2ac?*$ORP~(K5Zz0$H#00?{04e1aKWYV*sZ|t{$hD68
zvbMf2Gx@ZyZr{M|s!$tV?L@^)bk2}-vU-*l%YyObQy`}>PLv(0*I+X^9YgxeiKcOS
z@AvT7S9S&~bE+^{CScLdsFbzydFYf78W(NC>#_?(umenY&T##-HBZHQ(a8rr5j?iJ
zSc(9$U_4L?L&Jta>jj>oP&Az?>$wki?rd$Fm5nbqH@<pse{1_;_+WEq>;7HH(lz+G
zO3!DZol9n|R;~F6c0I{zSD_yxb@*T^nRrkaJ~SEAV^Y+u-#}K={0Q&~=T}CFcDtV7
z2VC7KNq27SACt^03AQ7rWF;g9>}z*iJ0`;YosYPlFPY#MF>$zhVO%eH38jIO(gk~Y
z!*jb4mWEk7wYDDuI^WV5<_GEx(5cB)0~)hX6{SK)TAHQ2uah(xN)_4xYqI8|VPaq!
zV1g7v8|L~dkE^it)964#*cMeNBcI0Ke)LWF#e+w?#)G{FJ!$6cW>5)YZMYbueVZta
z3&kpik_2^BLWXv&CZ;ygp;C|)!pu)2$1F|zA?EtCR4w6lGcJA;bh5!j9G`Nw%&s_$
zGV0N~Q^(*(8Yk2tZIhfW;G?R)Hta8J2g(!2{t^9!B5~ldmd25z=eoY7@n%W8wTzW?
z#7jP&FCF1tf<J&BroJzk9r0d`IAZP+#<M?cehn9hs-92b1$!z_ZH?bgo=7P0wz6H?
zm|QS|YtKg4f-Kii`v;kTTSk8XMAy@(AD&zbEZOQUfTG?;fuuJVW(wbB7%+`HMIjYl
zBQ}-Fm8#>?_i={f)yDdPx{iZQ99+V|%0(=zO;>%0lb9~m8j25a8B2@*uw(z6V%&yx
z_M$`BD)pTPB;*F;v_aJ^n88f9YUaXF<Z&2QE_cyhKw9?d(GrlDRPuNL6~nl8$Sll^
zsdrEpi({qwEzXr}#d)@u>{p<$%qh%Ll?P)5xAnj#D}`}wJ;bUr1I8v>dm-@GD=)u|
Zn-pkG&U^$z>zp(0_^$8yHQ$H!!oQgrw*>$I

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/distributed.cpython-37.pyc b/pycls/utils/__pycache__/distributed.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..002be998868f3e48ba035c983a34930519ea6ae5
GIT binary patch
literal 1883
zcmaJ>&2QT_6elIwRum^#inQ3UVIVLdn6b_tdf1`pz8Vi}Ff<L8bO8cv0wdBk9a&OH
zDydx?*kRZH13M+yY5$V0JN35nPJ55ClWr(RfREG@-}mpm#~T|Ff%U~NzyJJOi;zF@
zW3@cEd=1+?0wTx>5mb07J)v;=sW<UYd`d(h8o!d0Kr}@Q&V~rZ8k|iLiFG(z!aF3L
z_H4VajMeJ2v{JB<Obe+^_ciDx`yE>O$zYh^$7-SJUD)PbAeM~5<e698v)<Uh3a*Lp
z&QPt-f24%;rImW3GHI9^FgubgDYcfFWrfZYX@G4Ud%_L8TyeWm#9%%B7ZaaqX^K2E
z>NJ&XkZYFa!=cKC_t~XZ7A}lu0+?LjJsHX5tjJYnP1ot$%W7L?Y8CtL;5jqWq7;Zr
z#*I<K<%8O;&cTY!os8z$8T;(kKXmVSB;md0Dno0farPM=AP4I2qokJTcb=tFAfRsb
zd>Owq&a=yrN=EoF!RQ1?iMOOhF5`F&nI#%Yr=hVNRk;8BX$&S0tA@>WGOE_aZB46i
zC~b@p{6TbxqDcP!yf?}xvNxLwkxflcaLXYdy<(cAW=Tt{;QofHZZWMkm5J+e#TdjU
z=rZU82{qz1ad)`j*@W$|@&GP~l`C&ft|`U4KMrhT+!XYx1!o9*?b@5uIUwj&0O0ij
zMZs)mFN4bRRLuZG2p}N`916H91sm$TEV{FGn-mXwy_buJFBhGlYMt`rOlG25hw9jP
zBJC&_Rdku_RKy09RRJbszZ2;OxK;&8DR>n@0%9RAR3bIHUfprh%UZ?GRGsj4fA8=p
ze*Deu_uuaKt0rJ^!fh2id3JbI1qaW5IO=#OYbfXPt~Ws0ZR+(Jer^GI54LFoA%5Fy
z(tvJKkGel^Z#hhMoyKbdDWZ@fZ}w4NLjUH||I5#Q$9I9&VciwAgexF);WDxtY}Bqg
zFYkfH9oTh=J&oT9LJwB}8`I}XEZ{OE3;JSbPWt4h4;LPKI`-xs&Zbx1!k^PCf9_-X
z&!c1VVtY>ET^smrZ~*rKa6fR;+Y6ApY6#yp$F27rWX_))lQ}_9@Vnr^fz+7wI%t#d
z6k3L-X$-AXCX(ea#kf^@25skX1I4(2&Vs&@naN?o)idS@Af>XRv@1U2(6W<UsDXmV
zD*LoZc_O=Pw;o*3$EGZbT-#+=BPLBbac%5u*G8TBTz457sC^Rbf~O_G7%eP4KAk$}
zRCWPF7r1~-mJggVRJ=r94RZFeJTDdK!1)Pt2?ld_)%Trx!X3fD4!N0TNt_o$C<v^-
zzUqW;)IPdRQq|-|flU%zas}v}l#o7bTaAUqE?VZQif$q^RkLPRYd5_AAiLVCgNtu1
zeFFN-9UvqOXhh$kn_fiQu(zpZ!0m)pD~<)UOI-KkmK#IeB-av1Hc-RbF`O{2KO7L9
h>>Jh{fDyJ;_y|f^rt(W<xeg1wkhcABJ=_Qz{{Tmr1_l5C

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/error_handler.cpython-37.pyc b/pycls/utils/__pycache__/error_handler.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a275b4cac7508806403fea1c081e15b5760b621f
GIT binary patch
literal 1945
zcmaJ?L2uhO6ecNIk{u^4(yiONuFx)93=Bzo7_dV(6zhtvz!nTmh5`aMLJ?_Ojw~r8
z4Kzjy3`>BX_78TjvD5w%uRHB8?6mhN*^7%}H26e5lX~BK-$&}}?d=|c_T?{s{QN^m
z$X_^FH6I3#pldXOBt%fb3Yt*&WC?RmFY%rd;R*kW2w(b_UebYiM+7hrWB~I}up=@E
z=l8y!6(+A$l}WAhauiFYDix2@QWR3{{|!;dVTN|qymp{t7)Sz<2}Vu4*JKD-gm)Ds
z{%hisqyrIryYqOQ7vjmwOx7l^%K6?4mDV~=%UG^_@laLMIE^!08Mk2eGuB4JCq~7t
z-vMKYp%>(mE!YKJQVMl|Nc3>v*^ZvokOA(n5$AcC8_oyJ`dSu4yX69KqtZ+csO@r|
z6{*&o|4IJ&;$U1&<-z=1i1J(?h}5K7nuGd0D|BmTCWo;r5MLFjK37P|E<$s7>HNLT
zg!k7l0)$kiQ^~pQaXziYtiW-f^Rro6wEu*hiz?&Xz1MEH7&BQE;EN!RGXttX8bo!`
zVBj0jHA0D~&zvWN$-3c^N6__sFxOdgjFH3_zUW*5XCe?GJOdGlE<8iggW7J{{u9*2
zw~Z|3zyGgl;=L27ortS49;Ug{^6XPR*zfgXXy4`9$Wq5c5Sgo_j`MPpc!}Z=ZZ$}u
zPDiN$ml@005qFYir!cLYr%shD%~~>U(}r%aB9AfehhQ#1HJpqHSZdhByC4%E&&zKN
zJRO*I7QvEGV!}me!bt?4?vl0o@(zzdNP$V!NE#c6R8FgM(6d{OAkJoTChfLU7+~Uc
zF0}2Mv688Pb7g~LH?%?FDv<0n+nSkFpj>z}yjSO=GA(R)^z7-6N6#PFoyM^e+<>VL
zWTH@T|EZyo0p|J<bnUJW7Qk<p_SpR0>)fs7xKSoNTCPktVE}291#OQO1^U_i&T~0J
z+zPnaWNZHrAl-spHc0DQmZC*yJ*dDN2yHhNg1Zv!Ax?K-*!=bezV(uf!WwM2Y|Z0m
zFlhiT$t&O!ptiSw@fJSx&XSm55<1=*@SnV0bf_ptLhQJML?TyK>mk8jJu{Y7+WMz?
zQP=={n9%zWUVQ*&MMQv(^|5S-xD7tUsqs(Z^;n<HZ*3&m(4Jx0v0-*dtD^u<=*zp9
z**D~;-B+HWkfCS1iH|v*MK6d!tp<y5NiONHaPvEgZX7w~tgmiN2YFF3nvc;ScpHGS
zo*f>$E25FnJ#Y_vr>i95pf_I5rb(ZJBC}IIRB&UdySV?2H1B}#bLefFeY(r$A8w?%
zsa-^b(o|b$`e?S%VD<`ytFUkiUwbW-3+k>pyy9?$(|6SIS+o`TQCFcKUhUo$!?+;{
nKNPmR{_Zs2grGW^$jrDcZ;N50ht0M86^3<zAPj|{cQgDK`pnar

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/io.cpython-37.pyc b/pycls/utils/__pycache__/io.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8a2532424822361aab2e50edf1e20e3c6086dd4b
GIT binary patch
literal 2286
zcmb7F&2Jk;6rb5|Z@x+!N~Kn1kf^9795;zjqNt^5Lxd`Yv?QR?1-fi@#_=ZW9d~Bj
zG|pN=O@+jX8)q9%{7ZA?)PJET-mIOHa$~J|Z~WeyH($T^oBJ~}A%gGD&woCZ(EHo`
zIs0trtV5GOfuJZt6jLk55w5fqS(UaUyV6eNfVOkD<3&D33=acV8{sIVPBcYb>b*kI
zH1%l!dWP26Ov|DnoqC0%S!(T~`gH%sgNLN2vOH56ljLSXldfWvv;^;vFPLJf61=Ft
zhQ*-mI@Z2m&w5O%{>N>lx^k`Ar1?T?FUh*Sy*x`>S<d7_$~(>Oahl82GUMR$CT#<>
z{2hcsL)cLX`oJ35)H<vUtkRO#O0<hgG_Xs1Xe(!g<*IV2J@icf6u-n^_mxxPku`8s
za8MhDC3wU)rL&8w|5N7wz@7M9$nabR!|4)w7^#`zEOmxnX+xA%S>{at)WXPQsJs0?
zzTihi&J#*hn-Q}gBD*ZjS{a<m-f?C9c4vzyUUkzXZL^qWf+Phcf~lS;j3XrQbraQo
zClJmR5-yMjt!i*0_In*xD3VE1a77XVnLi~m>vZ$uvt$|AWtA?I3zhEAKD~K&&76N-
z-h~_3?%%CWoA3gLB|XJEiOSL>hkw=jq{%BuR;`STV|SOyUN1dlYK@!>`s8faS+Nfj
zcetAD{bXP^a5AW0(3W7@h0JMNhY!Boez>!_ac_51hf*a%$)ilQ^<4bu$;Q@hd><S!
z%&BMZNy&uDc(Ex27usua0VxcOx_Oc^ZE>j`b57c0FEXj5_Tey#q{DR3Ne&tOk=o1o
z{yr1h$%+=&({yq}aZluRFPZ`sa}k@PjcRW%B=Rz8BD4)6nj70qOl+o$X3rU5lE09c
z2UxyoG~2wxn*C!+i(}cOiAp3>V>C~ssTSGH_(JzsPo{$zE*oSM5Xhak9PB`w#tsgw
zX`?F{S{HGMg^7HpFaj`Ch4Bk?C>~e}DO_45nn%hi0rpqXDae7XfC|oAseKA%K%LT_
zN7RK8_7xcO0Rs5j2hI@6AE54*P%hNJicA~X73huEA0YreD)FH!UQXfz<E1qygs}f=
zhjqD7<m|LbSMYto<OGBp0`5kwRe(IB0zg7~$e%OO%K1^n0u@)l?aSu!;>z;!(ygXq
zFI3|!Rbwv^jjU)u!8B5c;yG(Bt*ovrE-o%tnP1<ABiV-r-;?#%7}Ohl=M8T3uM_h9
z$>2wT<_*GI<YaAOu{HSW{<vbvOokkUG!D1xfwr?kX;<)GLA47@1N_<Y7+@t8<vnm6
ziA*t3gXG5FlkK{v-4x=%r?^V;$cKf-IV++N5R@}&`a0+`kwW=OZG*eQ&se-~GR?$i
zyNtoR1SWYA1acjG#k!2=Ee~{v11vrPYlaA*0@hr8=6Nv7d5a1|mn8yXbD(F9DXw&f
z_6QB(0T|(-J3<3bd8G$9@u)TMl@GTL`1oyo5WHB0*q{W$8`Mf))ymo_UO@xDtjQl9
zq0$xyK!QGuH8dCirNR?}(X;i(=rN+s*JuZQhf2Sq$6wEgf&k_#&Z`cO080T+P^qdi
ztV(Q5R{s|#OShI*KU-XZT2pOaG+>rI+gth<F>THdCoH(Z3kApD*kZ-LY6}AnZ5{1x
z*Qay{<0dba_IJilmiEEOyQ~oJnMiw|!SzZdvaYyh436<>N3euy8}3(p1lD>eE`fr_
z^ZeFyn6`VxVJx#gi)P;5vss(AVO||Tp>zSzOotVK#*x!|$BM~V^0<`<javXRljc9j
z5e6ok(ByRx$PIxsE&@&riCEA#v|>RE0Y6o2eNOvvOnDl|+BZ}QBrc|4sIZOr5QO%|
z2Pw>hEb7KmgZ*!fc{PY*!?<xQE*Q6Ogkh-)nbj2MmS|{FR0t1lcR20k>~lk4(%2m<
RfQ6dhv>jTG9b5}8{R0oQgY5tS

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/logging.cpython-37.pyc b/pycls/utils/__pycache__/logging.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7d70c65240da862903fcf8f4ac6bf7dfb454a452
GIT binary patch
literal 3790
zcmZ`*TW=f372cWMC07)6u_N0x>NcCi2}Bg7(iBZwLown+wrjwV5m{*hHt}|~v!qsD
z?y@sOH)1HDfK=EoO;PmK4-)d!zcjCX>R<3vzcWixm%1e8va_>uzB%VRmroWJ8V-IR
z{pE)z^G(P3H$6;$E+!wNsh8-4({Th7JZBwdN8aIfbUUscy^d!`zvJ7n)~|DiIsKqB
z$Eq(Iod({x`Fy|GX=2UmEC|1|C~BQ0Q5V6H(^(dCqJi;}m={fqm&Jlu#Q26-63ZA@
z#3gYV;}vmLtcWW|taDZH$4=|oD0o=*dRftH{|8u|%@#9m((R2_|M2c=H_Ou^?x!tv
zD|~k~&x*7tTfY?F&{IEED(>OcsG+ndhq|4Yy-{;dl|`g<tno=g2hxwB`DmY_sSq7q
zQ#*UOlJH)7!WXq8e&CE9%_CRT0V^2&{PA!wkPwopa3Hfn17g^fWk1xrX_yS9Obd-S
zWr98JS1dHFaj)eWFH-3MeOd_PM|1;Ii@Nzx?OHe<q!Sc^^d*|21>ry>+x(Z)@+G~L
zHR|fr1?Wr=l{<TBqLrMZHR@h{u(n(F)3wpQ5XHV)6S0m}s@DelNv<aSP-mD!EmpmQ
zeN)SNMJdy_@x9{c95#A0Qj>c|_LlASyiDR;-ELQVH*qCb?JHBOv}}T%VU|M~%J}w!
z7L#CK1*(EKk4Lj~;oLW|Xz%K{S6z7tPv@*dTCmo6f{AvH*@+{!d$7lz8+qt`T@&mj
zY^3XZf#6TM{6Np`HO7u`*PUWT&+j#l7w~RTFP(6#dh7J<au+uJo-wDW=}U*ux$vRr
z+Gyo*s#Q1~SW`~*)8fX@(`RXJYTZ)yV=Whfh3+lWxNm|f88=&(4Hv`CVy2>gtaK`&
zxiT^J$w9`GWvNZD8y6x^r83@w&3lhbJ-YX3>$5xCk`!-(hmY<@8+RXX%LP1El6Z-|
zgIMnxZ?DV>Q-AcucC_(e%gjZz``&|xcg_4xtg__Ovgl?#tHl=UEKUSRNwqxVh_sei
z04;`tXtHS);?y!a#|_+oH+h3K`3hg;@*S)tL`(Y7=%QJhP@Ex~jGYrkL>*n(O0~?O
zD>c(#D%&kCuVJT|>!tc3NS8|ZvPb@VO4|xj13N6BRIcNTii=#nhxrTvNhde32tLMy
zeD(<TJa$f;4d;-73v7>U?UEP&%{^e+)gIRQm>>JXKd$Nen4Rzw*k9CuyFOY5EEQ6S
z2yKXHpOZ%Z_83jQk1m6xZ&ttN2RKhVW3JgD0BU~h3g_4x<J1#YK+>4~lmE%HTB&_^
ztNmg3pyiuJrr;ntE|S#v-Moyosf#qp`f+aR8`VU<0Vyrl@b>$LJvHpv1tq;|tcxts
zN!cIV-XK)On;IXz4MRq!QSQ`VPq%M_%5$oeI(nM!tIqsTrO`zEDERJ{I-s`M3XGrW
zw69D<W&J^((xJu|!~Q^7Qy}a_WW6XdURJ;-2r(kt_MQ8{?OP3+#wiq2BaWr=!bJ@6
z^9<rUbQznv+#2||%mOBVi?x|8C;}3Pt>tfH0{z4Khy3r02W-qJHXQIT9l<|$ww$lX
z{vLS2Ehk*Wa(Z&@lQ$;I)JCfh%UD!~uW~~b$m0-Z>g8z&fFZee+c3E~!(kN;@s<;9
zOnStL_rNn}%%0YyH@0Xryh4K`hwK^x`Cs{U=N}VBY_?qaD|~0XZiZ~&BALNQenu-S
zzn~?UJI_>Q?$5EKGPg5&2ka4Nw0Fkdv;F4~llCm)tC(2C(3<sLFt$QwMFYt%2#HSu
z)C8CO2ul}`Nix9}+5{5F<&}f9Fms5Df|9ao`JI4_Y_lEf;!IlYRNSz0`Ro5Z*eqSR
z+h-)iGJ`kSWwwaFB<d#=3+P8#g67oJhv=R=&so<NGe^O=iVE%vFybj!H!*sYC%_1b
z8x<qmv1=Leo4em?87De?7Ux5Rv7|&kSA((;l;(9A%JR>V-XElEaG?;@MOb!;CZ`dD
z*g<&6DHKIjtLu;$DxB-i3kKQViy2d1U~S3^)CF|()C`soE(AieF2tE&D!j4;MR99{
zuSitI%F)e1EaBFt-uF6Uz;BVBjGqq5WcM6PD6(3X{B(-^YQ^hgpg%#GcpZG6;N`E_
znxuq4_)rMJ^#^&Tr&M;K!q-uzbBeaQDGCW_Q2q|<@;*A_;mfL|!CACYy|Y<s`QrV4
zPsdR;<nR@?Ow=GZF-}4Xp&t>Snk5JA1~qKGdlYD-ievr)7dvD&F!1LN-nk$~Fbf<W
z(L{qFV}4i#6Aw?`_YRVPi==|}E>*1lXybmW!#K}paYKb!5sFNrL@VV!$W-V!D~L&I
zxPx^?>;d246(a1HBF)wRA<~b4+Y)K{fSr=(jN6_g4J(jg=WUFJ>%r6)2=hmrHYLnk
zlbmavRK1QZ=LkZ{&{F%x6wwM1WY6kxj8uE#o<Y^}<WI0q5}PEIdh#uF)1X2`15hfE
zB530R$QW-YAZG~3hO<C+6AxA=H>T;BP$0E48-ih#vq;=iB&L>ri<qhqhRMg|C)5#`
z3W2SYX3n*P^=X9jQ1u~wrox^D;5Y@qjXTN)DAB*5Beu~QpLCB(m;#^b)Vfg-A$iil
zxch9|qz*f0fF~rbigh;->sDBk$*<`<CMnjouAyLzM43d9T*9+#QAg!fbv>#Y%{5!>
zw{4l<7QiU8ogq*g-!7T?Gi#F4K^5^YRi7c`rR_<{NIoQgOdZ+HCM{cC+M>`F12#q5
xXl(t?Vm+hAnJ!JR4uTK!^fvjbBHMX9K#sS6ei>ua^IZA|bHPgB2ThE@e*yDteDVMQ

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/meters.cpython-37.pyc b/pycls/utils/__pycache__/meters.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6ff94ec6049fdf543873f0e66aed7950276c7131
GIT binary patch
literal 7415
zcmc&(OOG5^6|Q$zKd0y68OM%bfB<Po#&JRj2}y9`JRu&N#6Z%DQfhbAjJxT_<W`N7
zY4=D-lWq1OMQBE>Sh0_U#ExIUrdYDdf<M5D?>o1;x@OXj2@;`N_uNO_syg>PzkB<|
zdc9)cdhR!Wc<<A%8OA@@nOzPVS5V^Lf`mrf2+hzMm~9iaJ+Ov$+cx!>({|;aH!N9(
zX$<{#IlEKIMyl<qj5veZu->jq+Z{B9OYJ4Jy|DC=(QdwKgnn55$Oy~P@{!$MLBA4K
z(XUE>75!RRN53xpwa|LYXf-CKYf%!Zc<VE;X>7O5)D5He_aijy?#?dzH+#cKbxq#K
z<z?k6>i7pBQ=@Mjnu&?p3e6+yUF*;~w9&UiOM1?s8`{VA)SOyVXX>6f_^vZ~`s#Re
zCsIkg*%@s{NoO;`XToT})_AkCxzqbF3LpKT7bdqi<3#mFyPG>>HS8o?y=KEOK6@6Q
zo?N{9*fWnke)!1U%dDQc`?#*2VAXO`C+r-=sTcO{498(wilgp$6k<TXsO6;|J{%;W
z;-cF2Z55}siiW8(?2Y2KgZc3v4OPd@e_g(GdpwLTO%B3vbP!((J4q*wl1qCB-9enu
z{iKKXP>XZ#AT9CbV0{hG$3BQ*d1ljWm}&{_E(d0pg~k<>m?Sb_Bt{#;Y(s>h4dpta
z6S}C~Oxd1Rb-Sch-1fC<+vTtU;Vz}sx4NA{M~P}Dm#;;gcwa@3^kyvMn|C^c{b=*y
zPT1K?qHuFZjfa~rf$?sl#-r9&m;CTzZKGU8i5~<>j42HGV<<Z?LJJhe3~Gf2l}hvR
zhMf-w(P$Tw<x(`9maX`z)J?`oXOOzx@%|`jSt&$3*lAZk=#9eh2SMDML?UfXpk*WV
zgP=F+B|-3Y^kVACFl(k~PSy%b+?q|wt&|_-7lrW)XwXWJ%@g=nNaihJ-WIm)?K8|e
z8MEC%)KO{$PpegsmMN+{%b^R&M#)2%kg-=Gd&@=kWX58lK9ch`HBU?vd^+MsPl5da
z0!9y>BR$Qnwf&h$G-J~1?CnLPP#6?eTb6K{x}xK=%$7T07|5O0oZ%X5a&oT7a6apl
zKyyo|;n@V@q|2X}(7d_bvQulc2kl2;uQL(>-a}rE-6#p5@jy@AA|GOJ=Hx<=htt!(
z1S5)u3LlQXn2+l01D!j&mkUg6EMfw4PH~Xi7DYkZLNPDW2_OpVzBVH=*d2^-bp`=m
z%`kdT8O*0`w<`39QD+zhL0Sm{M78|^+qEEgf4?)x?v#Qc9Cw32ox^+82FXPdq$ope
zl6-@Najih;ZmJ1F@x|=<repc8{A}NrdtD6Y7wzE+O56rn7`4_zM68cY#H~{3BYt^d
zIjo>Ah1IZz+7Iht19drE3Y(}a;c~cwx*D#AYp84CIf!&4t=v?d-bjbD^LaQ+ILO2h
zC!Hid9kp@=-2!h&J&0F{*|+FSYfx6;B<>MwFDYT%KjMtCgoaAU@oHhbR%q8Ha5RK2
z32iP8BFkwt+8cLo2R)`&MOmfp^&2-|zm}G-z3_JM@{Q|PUrT*NOPQgF7t#tR2h4;@
z#GO{&xc+Kz<K;JBzy4C{490PsR!9RKN3DTMtHWF9?OhJ=$gK0EbZVphp}sZj#p$f~
zWWZjP_u|fQZxF><#8BS_2WRbp7W8A3cpbz*U~8DZ#9z;<nKk4ygu2zaa`Hg@%Bg6V
zD`OiknVPCGkH!okg}96uN9ADD)uU*6*?p#)l8D?FT3#&piuyjjqRswMwisd5C6dQT
z=<Dink|#*$dwI;%lAk+<_yis3rzkNKxnVkH1wXEQR!^D*3FuhQ5Y6jU9^%{PD;Mr1
zT9Dt=PI}#YL`*-kitaBlaTX_vQlH&l9HQ>?+3|foJMO;og`XXFyONeMi8bx`Dj!fF
zdbP96MomwVFqq}~Hrlqdp>1j$na+iyx@(4HJ6A?LV8+>qCnKe)osX1c#GktIgg=vL
zSu0v%o@iUlKrCRccd-glzNZtWpL-Eo;zBBaFfGB&7U7fH{vNPk1e}Ku^e)%U5Ngmn
z=qJ{S$=_t7U?IgPx{M3FN+?v6i0&8-imopimeH$kBWhV2052QHd1K80z;^m>;!W)%
zY%4xC51piR?4KBf+M#n?CQ92dl1jhYubo&j0zA1Pf<J*HW$!ln^`tR%82$Q7$ITPo
zU(V{4tX@sLWbODIfV+urxt~-Iy{U(&>}{*_;1f1WY}>9#G?rL4aqk{px_j8Frf!fN
z>_w?7kuY^=r>RdTW7bQ{(h7S!JE;w<te(Mp6r)W+?)ESmw${a^(^>}`hR$xJ1LXHO
zkufXvsR6!Azfm?3sofhT&{~A8LpM(AyL<Zp9K&Eg#tT%L&r<a2VG?2?Y!<+XKyz(A
zk}c1ek46^Vs5fYpwAGZm`zp}O^VyUoYUCev6}N|JIsb0zikX}VvZ<f}@nN}r8xO^s
zAdI&ctfuK(7Hp^~F~>vv<<Bu!tjPlft0J12zazIRx?}E=jDJ(Hg9Ns<4g~*eTdmV=
zrpB?w4cZB}ZuYUGT{8=?qGyfcT`e`DR`n)GY7O?&#$db~^yBeJdqK<5IQbeUZIk?*
z<aAWk56u}q#1y?r!crZ$&Z))m1M|6gZlM|pz{F{|&&~}Si<z1Gnsyi~a~Lag7&~(q
z=9)4}<szI1bv0`v+VpE#yChCk&z!2!U+On!PLwRu1*h)Rn<5(ZSKutG$7@q3bC&X<
zKlMNFELH6+VmC8yDXZtP0`(J;7f86OMLwZk!j0BBo#tNVq}&zME9|~X@*2tOB+Okp
z;W6CltS8ZLffs1|mrzMtS8sr{$_0I1=P1ou-5|+bw(Y~C17eef+Fqi0YCI%1euRhP
zZ-UUq*J<MtFKc4y%&nIE#L6$6vGT%ea!a5uvruL+!x9V{NaMsjv=S4hEm>!lA8f@$
zs@@{G34+}zQ@av_U1Y0Ew7<e<e}~yIy~+Rvm=yUtGZ7o&S6X$&g{p6nJVf$sk~ubi
ziruG4o+F`Ksb7!?m%n0b?mN<AqEO!f`4Ni15`pSVm6{Kr?$`WtzUQyYj4o#67bSEB
z<$n*@%QvGq(Rh7hj@L1EKe$e_oxybh=vE@BP5`=BXqN=6`$Eft2P;CWLTg#*+QsW|
zy>mr}uex>>s!qEPlXW4aF+d7HOlvp-n&awOPosi4h!FgsH{ANob47`csN@rrcm;%}
z?$2PehR!RwnC^$1>GMF&O<Au#!w_(AZ1YS5k;lnfu5O7OJfSnx%>8M8+A)(!SrOkU
z?%@i^IhVv@sND7`A8a||f^Eu@iI)F_k}cVhtVN`(G4o_;zU2QBDytvj^^2jhqAz{Q
zuXD>`n3IBlC4T^9e|eCs=w|o$ApQ2wDEIsz;ITvs%wf$UL24TfsEwy6*}+c{9_wSD
zDNvt2`VpY(01VGYW`Nn&V0jkWytB~vjICy=IS<?FS+H?dg$zS~L5UgXdE8dhAa;GO
zx`i0l#q27ItnWQEct!+FYV&xT_#(7{N$oz6D=;ZXaJng&)RS~t>idvLNvGrEs-V#t
z>jWtEc!`(<$=mtFKXj%%M0U5^<peRM2s$+;^$WT53K-D|^+mSkOQ%b<nuQ#>C?85{
zl?|n2+Mn?i;u~{#D4BMiE!{#gp^_z0w)M__5&V;Y|5r@R_LoJpN2CXwcYZGT{}0S7
z<jVg9+)4WS8<*sR7&AmySjIg79XF7<1$~GPJp=_uYlqymu&@n`iWdCgh=&4>^jrWj
z_OL5DF2U*?u{uAqx^f?TA+fSrzfM1r=dtBX8VPnj8DOJrxl>ncE3Jra6*1c9jZ1*q
z+@$7_RE%r+4D3p6M9o6Dox`(5rZ=~`GZv;dtQxM&zkLJZT7<m>M*JN&<p5v1bA#tN
z8P+1wa%TI9S=Qb<gV(+qprr_Q{z+CDCIwtQFgN|L5L>;Cum7HW5ElxzYE-I%txA3a
z*lI;_1~9llIfCxA+8qxD_yeUk9>v-J2RLKfJJ6@mtMW(Dmi$Gug@x<_2c@ob(^A$K
zCreMw%(kx*qta#lj_z1RTlgbur?;z56|-^uZheA9@d=bqz45I+|JG@8gpg4<FGTzt
hP^SK~`aJo0I;|Y|;rJJad<a4KAq&{P-+ZvS^KVCs+1>yE

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/metrics.cpython-37.pyc b/pycls/utils/__pycache__/metrics.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..23233a917bcb9b1add4a1137b999471a7288d552
GIT binary patch
literal 3724
zcmbtXTaVku6`tWulvdiCw`+TyI?TpREH{x;_mb4zIQBXh7>M1VPEk~FO3<8HEmsmL
z&rs`KfdY!!LEVP}&OeaQr~Z?oKVeYxrB8nAb6(nWhSX)(a)6=~=FEk|Gv|Es%{imp
zTCGCh`SACD`Q7RYA^*b8<T0Re6N>DBV5CVHWk!cK4PBc}Q`c74ZrT*znH{HFYL=jF
zHC<*i=P7BHS&6w&S6G=<pk882tO|9N)z~uBHMYW5p<ZTdY#r(qX51(BGXw9_ejKJz
z5=(C{5nh;dd;K(uTV9u^A_`^W1=vLH)u}SWy;elXDhW*-f~Qmz>`MV6BO|bHOf#C-
zZ9~46f-g_#A82YEm>Ef}40_CD7X0i#+YgMPIkYlsWRD5dbmYK0XKaj0nKh<E8{W8?
zJ*Iyme<XWU`{-uI$K(keI++6=lm|Pvwar|5=|1=3e)l04Ub5$<N$=e@I3xrQQ?DmD
zi{>2VLAdX=8-jw_gT`)<LYGBd{BVA15Hqh6Jmekeb^9{)x;XI#e)MJCR3?0(N}WKo
zxJ;EJ6OnTE;y>_&I0YZvTL;P+D6#>f1t@*>){s7fozCdj27~%*W6wBWU^k*;GBU>G
z8{?4${u%cK!VHU9hY^3Q?7dDBq{@{2ZrxBu<2^;&usLhst=|of{I*|s{8#erwX6F{
zmtP$mGZr7qt1L(Z$<wR7<FF$O340Cg$zJ!4zpd<aBAFr$Zd^i>bNdHxfd00e=x*$E
zk}&AV8_<}{u9uZ1qXAdTx1{6(DfNjEiBNW$h;Uz7s6m-vqoS#BLANUNqaFb42T~dQ
zQN5&WgkCBqM_d_g*<A5wo*lxH&E;7?ikWP#3O@0?_^fuqR<VrmQBi|Rz{w-xH4qz6
zBno0wX@xq5MbDY8v2Ik1b*g7fS#H53ESlvPp^=hjuyep;b7*Q-0>huWnTf1~9Ww`)
zPB0Rz0UmP1LLae$c_$cxTdWzf4GqnZtszBzJS$jnft(}E*qE@wY{5Ku{{4|AaMAni
zP4IIKhU<n{f=baN7&s3t4MOhyDs+DcWwNucBU-*TPcGco9FDq`E80SB7eJ^rIBK{#
zGg?fb)+-{^YH3dSWn2`2M99VKQ2!qgzr6tBAPoB=2#;Th^7I_yE^3rFhY+hY#$-${
zK<t_#Y<w56;`^XrF|<E|Hxp>z)ZQ0nMf`t)SD=kA1G;irB`!Aw)q~!YR~Uh8D7T?V
z<a|m-6f?<FjMETdkgsDpGRH)-cKZaRvpCudL;wMT3)z_J#;Aa+-bdYdj5L~6L~}$Y
zv8%9N*`SMerkw2MWW-}RW8}|JQ4woAy&?view;3(q@|3wCorL^%FH}vMJew>Ftob~
z<DFB>sOr4hpG?*xhIkv=Q_nCk;c3*UJpc?-azGD^TpDCZhme(r=5u;<9Z&@9J|@rU
zSMNX`vxYWgH#@ZtoF|6(Co?ju?T81Nnb`*=DDDOUSsqo!<Scog49!Q|8Dv5jTVnJ)
zM5MF$zKZR0*s5V`3v&Ef0=dp8W?FMIp<T>`Rxz`bnd$Oq1-@U!l`<zQwahWa^%!K+
zc1ipL$8t{|Z5I}=Ju#kBab;r9#Fkupp&24xZo|U2##Gljc)zW^J_A;4yzKRFNM??{
zr&;F=F290xkmh>QlWCAfkZC%{MW!|-Z@fUB2n@#qyU`QLL48x;kg~u0)vY`CRO!xL
z|NfokC#n_$ahVTn=y$kOl}JW0XpBP+{E3xwJBbgkF|`!Mez+gRF=TmFP5SAqXFZGp
zDc(TkrCz`oT#VFGn+urg>tyFhxFs<z-oz;u7{iqFm`APsRF&@rM_(jK2Uc)Ck76DO
zz0w{J(!Kz5^h%SS_yH)WyW%a>L-!ShOt1N`Xsl>g+50{irtHU9oo4ou1}i@=KL-W!
z5(ol`%%YpLN?p1|4SEHBThOl3WqOf%P+z2C7v?Rd{s-Sh{mnV*yDw4y?*;Wg&#8Zc
z>h2uX%TuaX7Es+~Gpd)dt;txy)*RKBMpYnLX#vguwt(g}P4mry=5QeOkt^q#Ptknw
z+ZoN5r!-&5J=C;bEoj|%1+4?Pf*t~=@xHx~+yc|&339J#a*H~SUqOM{TfBqfT@-k!
z6z`$9hT_L4FuZ9_il1WZeH0&{xQ=3uYvLC;u!G`56gN<OgyJTOTPSp-o`!2x0UwIs
z{7ZQK64@#R*)~s*ZJ7!TsUZ~o0KN3x7zgD#2#m0pm4=X=Qu6}lBs#c$0>;n2_%-}3
z=_caX>&u|Ub)v^hKjeME<J4pPFbcVM`R;DL0T&j$Q7J3zvp{X&-8lc&hj0S7SGcJw
z*WdN;-1A{NZLYvuU#s;W9;aNMcDvT<^-)7Uhojkf-mOsa8T9I_s^t4D34LE@yxwuX
z0on*x-+>!(thY?9=__~xZt5IFycf0dJ^mE*=*y?R9tjV-n4R=F)%fY9=1iMyi@TG@
cxf__l<T((Q;p#1@T9)Z<xDNbmchlYcFWZ@G9RL6T

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/multiprocessing.cpython-37.pyc b/pycls/utils/__pycache__/multiprocessing.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f8e84243278fa325a0e1720084b0fd1c34862959
GIT binary patch
literal 1234
zcmZ8gU2hyU6!pjKtY>zUwxN{Lrv-$>BC+9tCse8mS}6z-6^hg{QW;Ig&c?}n#rCM#
zl~+P)UdwOT6;J#nzw*Ri;E8K*kdj)u*LSY%W8Ztue7m!gA{gKN{^xIxV}$<pAO8;u
z!DBFe0FI*!am-1DGZJQ+(J)8#fFO+OI2&TWgI26gvIOFY4=z!b@`MjLyTsXuljmrf
zF7N-?S4MWKErr&ynNLnc)d{6%|G+)yaEjgdsZy;vC>mY~Rl>W%@G}GZ4oq`!%U>1`
z3|eE0EZz`|EVYpx{DCogfts%(WXM^nzQpi4jDAJU$B?0rdA7}1vX0DP9dFPEPQW^O
z3FkNl#Nh@`yo4CuOO~Jh+&6ksOpg1eG_q|b$EvO2yF8H<pIp^3JDf5{c<=UPBTasl
z&F8A^J9p;?u{dfA#lLTiP^#~YOP*eoqVsT$c0GI>Rg~hWD9>Da+^V`Txwt6Z`&?+F
z+Qm)4Y3vezL#~Pj_P=gb#d9r}!i@mFRrz@@df_O%*fDT<q2?NJAv$~Q<G+x)0}cH5
z>;03q7W>Nu=gmUz^THHbnEh^1R{F~KMnYWQq-@qL994beLG*2Kh(u&ecJVHW!0uD^
z9^}FuW#pG(c%da=fxb%!EC}d~sj);W>S^4dRb+^b-r&_>j#lx-fuR<?!B5buhpVAM
zr$Cn-0@;zpmRNcmbMls~n2p!5Wu8RKw$$0dnpt+5_}uyDFVMwZ8w1LCmDpsRdUDS@
zPc_g>m!G{$uWW$aQ1BPN$rI>^0ht2B#Ki{MCya}H7pEgf>duXx%et$?Prm1hKw)kG
zyv&6gKHF+MG{eP3*NKL^fd&kdB?aeskPrw!au=PpvYDoiHGLh}R=WVDZh7ZBtaiX<
zJ6nEFw!cP3I_<E`XqQWeyU_2g+k?$+Hxv=Tav<vy$ms{*5M?+5V-(235xz%K%<yfR
zV)YTMPD#KxoQ~Zv&pF5}cemaVfVu}8mH!*v2SF{fppzMwAXj<R8v%uelp9^kDGDdD
s53irNQBY{#Ry+PB*9!7VQ=uyMux@!@iAVl+Jq8yMsD<p|F=c!I0R!kyxc~qF

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/net.cpython-37.pyc b/pycls/utils/__pycache__/net.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ee0440bdd1fd026e997861dc4fdc20f363de620f
GIT binary patch
literal 3408
zcmbtW-)|dB6`q;h^*S5JNz>4qrlputez><dfW$3GrS~RHS}L`sN{WK4TUHy-##v{*
z>&&dv#H=Mm${Y6qi9f&=PrUHR1AofAA)YGn!1H~9Z^n+3M!DjFwP$vI?U^&@eCIo}
zJ~!tPJU{#GpMU+`5+VQ8owLV6<1-Z5Ma4;jaLQRs8`Qj6gPFJ0u*}=;I1PvDHC9~c
zxD6L=n>+7FqZnGlnGxXy?!F_ZwBhk0pTYPn_xLQ{C0^olc+c@NpU1n***2*z^zS|G
zCCw~K6It0yMWquYQMVTdS(LOYNtnG&#euB7hi;@^rOImVwVG7-@MzTwC{Ix2-%({G
zCmC(CQzkZY(uT@6F@iU{NtSfWULpgEx6>}PU2bK?;S9HPdW(!GcfMehBw`}NgN#39
zvpEiQ%F;c2Q%d);xpuidKVV}C!-brQV{RpkyCX`q{A=yPaB)N+=l(>_b;$YK`HnD(
z+2`U~J_AE|{Rgi?0Sk0wWh<2RF#EpDPU5igHjG;PStUxMEDGYN56h%UO=ugu*HsI}
zJZFv1s!Z9PFi4aoGp>qzQ4+-dZZetw_ouDjEdBM_Kl<-Jtyh<n7s)7*S&%eC#gar7
zo}|gqkGPsW2t^XczKr^zDy6;5Z|(;PMx-i?RZ=z&vOu{BbjAkQGsyOpEe}PeX4iwP
zxnIW&M7jHc46;lJ?JL1hDYjl$Uu$e`{lb4<fBxEk`n<mO!hg2*^5vRx_M%uz+l>OL
zoPH=$Ulm?NNf?N#rD#XFdqLtup~k$K)Q0g}L8lY^g@{`a{$*`-KkbC8{S(fU6S=Cj
zlwr2oJ!!^rQune5ZNRJ6Jy9hiVtgpM2nlit6=5Z|LhsNLTV^+CnU?4x6<^1AQ=dRT
zb0|d=1Hb$5MP|U)wsP9GS4igM^pu{mfr+4ZfXA>fA{&T}fyDtyj98{;BZAI*4tFl}
z3df6j)f;c%@<Q(V?qOqx93N}o<}_y`hPnOE*`PR>$w^)udLziE<ai6vwN5E!pJEnY
z=We@%GRMeZc2LTl+;a$mqm<9)vrFV7lCxXnAtLG#$?eJ9?D^Q?pLWRM5q4gm%Kc;_
z7m^?c-<;J;WSKzIxyhL8m-2#{)!P4Shj15gFZMrvl6JbiER>b52%C`%E9><NfSAcj
znp83oz$>5t?Y-JG;<f1=lx-Aw2UQDZeRE?#5p7gQe%nI4A)^1t4(;tR9;&u*u~Io%
zDw_L3drVwIg)m{YABdZ}uW>W3Zlm?3_b85J)`WtOtLRI;ZGrS3=n<&wPc%kQ`9|Z~
zqd08_v3!ii*@su_ybMW3=LN0vkOC@f^ntdP(C1Q~2DkL%=m6m-dWklMucNxk#5eQ;
zwwqj1TU0RqG0H^W|4#$)O})$P+15u0zwyrETbQ_N>jz)4wWev~FdbXw@E%B*w#h{&
zQB~|9YLl$j`Ct3g&i9#eTD@9URwqgfy@=~rp|00|`O@EBd->|c=C=Q8YxBwT?M>zF
zYGN9*niYqMayu#S#i116*5AxTStznJ)hrjuI07l!%^nZbOqldKp$IU27YoHbRMk1<
z;Oj6}g*Xj(C>lj9^0kPDwVOyN)RPTM9?FJ|vq@t($W4GD*>L?h*wu7OBZEOjk1iQi
zEq@i>v<IZtgV@V#k(Ftgm6=C95LJmSf~;=S+jy@~kBT2-)dxthF<>Z5`|J2Spe3Sd
z>CFld3Abd3p|*`E@^U(KbS^ve5VRCQ%lgYmO(7z5f{JO#*6P)QxDR<I)C_{p15I5Q
z`YZnoN+x9B<W;@?ittnA@%ppqTEjZ2a|h8UrQ#u4Qxevv(TreD^a2fzGc1z<%Sc;i
zdZu$f=oY`0jlP8f!u`S;Sh+QHK%pjc4-3<ln_J^CoY?8#-4JQF(oB*1npx$00S~%a
z$SZpy?Nm0_wq9>nYf})Kq%1Vt!7UEYPP(Bg@bCyQsuss#?S!3fcA^})A9O?U6D$=!
zMFm_U-KM=bj*$9zbfnX1FZQdpqQ|&OK?fh;i_(U75QdYs&bad6GH6O>@kHUDL$<_?
z52h>9yMTlK32o3RB#^hx4Y-1irX_vXGZ~AGGZwc-`nG2GSD$J32XS06$e-L!FoJ&t
zH;EQ*EPJ@URgPe;RD8(X+J}Y+^frbF7RN+@4dJ&D8xum6z5frMi6Mnfv&ue-!nbGC
zN4u8}ZJ6&h)}1l+y%zAM_3$TY{VIkw?y64s)@lZs^16WtIw1mBR_Pd@A7%!%a9Qxy
z*{+0%XJ}7>sX?jpjsfQ9XmC1U$LrH@`$Pk<O^2}Xz&0)18*+x9{So;+Ksz*}BWl2i
z(T4!_VR;9c8i9$g4w(E^=KKzxf@iG$54W%Oj6v#^zA@+5f%_ENjm_+6mW!3sOuHw(
za<$QYT*Jz1Qi2z8@!S+56;;RBrct_RVMPzb1_UX3IWMUN9sL*A?qJ;nmNp{VqZM4Q
zL>;5mdFA>(Pn*6kbSe|`x-uEDX!3V0Ner{7Yvw*zYo;PZmFz{WG56~$n!c3!EfH|T
xeugj%=#7WWJ{JIFJgn2cDn1$~gU9-gAa9_u5jI7X5-nN9YYTUZZqX|`{{>gaFz)~W

literal 0
HcmV?d00001

diff --git a/pycls/utils/__pycache__/timer.cpython-37.pyc b/pycls/utils/__pycache__/timer.cpython-37.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0840c242e0930f0f259591752a01acaa2bae3785
GIT binary patch
literal 1031
zcmaJ=O^?$s5VaF0O}p)~5*$`SLJ_CUrQI7BK;=W?)(f0^zzP|6NlBVwyGV_CKw^Ig
zh!cOwS5EsEI5Fe2D^{v-<nh>^%zJ(wzZ{JQ4BNA>Ki_`}8T(C>Llfvsv6~bPusH(`
zLB;3%4TBKGH3Ol7Z{a+I;4RCdHhx#uN{??an#~HbhZdkS#cnRqC^qMqEdWk`LU_&Q
z0t7@jMi4^}#~2dm<Jf}%W<GI|&)i<TNKIL<E0tP8N-uKA*H%Hg&`q7bR#p{OH>+%1
zaKxcTBZ0rdZiZ-f>^u8`%fK)(pJm)dS{Y^05k^&uUra|B_UCf)v8mOh-2$w(W&*j*
zF>12j7M0oCjV;mlOvdZ2OQbAUrIqrGun3g#_Oz4auyvf*DTZ3lYsVx?(=iL2pxka?
zY_6>&KjWJ2#53$nj8hldvbc{<2O75h4is1K;NkUlI!w_y;4S$%Y9fOj6Komo*fMy6
zf=TpnE`a-)z;)Ua4=FbC41?JXiJQeb!9n-YI05Bi;Rd#`dF2VasK~3zxMBW9>HJFl
z{}fGej#&)lWBd%iQ+|7v=I&W;@ANUo0!LO1@C>LP`B++LLvMv=8_upMD>dzDM|wm)
zN$<I9Dh_=Q&2f=6srLxG@$ZGl`>#Fe#P&rr%MvPsRbDG8-9XB^flWpJP|DAnyxQ;d
zq=crBQuhdy5^5?MkMYmtQXy)(VNJ&+z8*Z`B1pm{N@BmB3H^X#NH!NYAC~k8=!avO
QT=KNz^qj&CRTCxq3l8$i%m4rY

literal 0
HcmV?d00001

diff --git a/pycls/utils/benchmark.py b/pycls/utils/benchmark.py
new file mode 100644
index 0000000..44b9949
--- /dev/null
+++ b/pycls/utils/benchmark.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Functions for benchmarking networks."""
+
+import pycls.utils.logging as lu
+import torch
+from pycls.core.config import cfg
+from pycls.utils.timer import Timer
+
+
+@torch.no_grad()
+def compute_fw_test_time(model, inputs):
+    """Computes forward test time (no grad, eval mode)."""
+    # Use eval mode
+    model.eval()
+    # Warm up the caches
+    for _cur_iter in range(cfg.PREC_TIME.WARMUP_ITER):
+        model(inputs)
+    # Make sure warmup kernels completed
+    torch.cuda.synchronize()
+    # Compute precise forward pass time
+    timer = Timer()
+    for _cur_iter in range(cfg.PREC_TIME.NUM_ITER):
+        timer.tic()
+        model(inputs)
+        torch.cuda.synchronize()
+        timer.toc()
+    # Make sure forward kernels completed
+    torch.cuda.synchronize()
+    return timer.average_time
+
+
+def compute_fw_bw_time(model, loss_fun, inputs, labels):
+    """Computes forward backward time."""
+    # Use train mode
+    model.train()
+    # Warm up the caches
+    for _cur_iter in range(cfg.PREC_TIME.WARMUP_ITER):
+        preds = model(inputs)
+        loss = loss_fun(preds, labels)
+        loss.backward()
+    # Make sure warmup kernels completed
+    torch.cuda.synchronize()
+    # Compute precise forward backward pass time
+    fw_timer = Timer()
+    bw_timer = Timer()
+    for _cur_iter in range(cfg.PREC_TIME.NUM_ITER):
+        # Forward
+        fw_timer.tic()
+        preds = model(inputs)
+        loss = loss_fun(preds, labels)
+        torch.cuda.synchronize()
+        fw_timer.toc()
+        # Backward
+        bw_timer.tic()
+        loss.backward()
+        torch.cuda.synchronize()
+        bw_timer.toc()
+    # Make sure forward backward kernels completed
+    torch.cuda.synchronize()
+    return fw_timer.average_time, bw_timer.average_time
+
+
+def compute_precise_time(model, loss_fun):
+    """Computes precise time."""
+    # Generate a dummy mini-batch
+    im_size = cfg.TRAIN.IM_SIZE
+    inputs = torch.rand(cfg.PREC_TIME.BATCH_SIZE, 3, im_size, im_size)
+    labels = torch.zeros(cfg.PREC_TIME.BATCH_SIZE, dtype=torch.int64)
+    # Copy the data to the GPU
+    inputs = inputs.cuda(non_blocking=False)
+    labels = labels.cuda(non_blocking=False)
+    # Compute precise time
+    fw_test_time = compute_fw_test_time(model, inputs)
+    fw_time, bw_time = compute_fw_bw_time(model, loss_fun, inputs, labels)
+    # Log precise time
+    lu.log_json_stats(
+        {
+            "prec_test_fw_time": fw_test_time,
+            "prec_train_fw_time": fw_time,
+            "prec_train_bw_time": bw_time,
+            "prec_train_fw_bw_time": fw_time + bw_time,
+        }
+    )
diff --git a/pycls/utils/checkpoint.py b/pycls/utils/checkpoint.py
new file mode 100644
index 0000000..1c4cc45
--- /dev/null
+++ b/pycls/utils/checkpoint.py
@@ -0,0 +1,91 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Functions that handle saving and loading of checkpoints."""
+
+import os
+
+import pycls.utils.distributed as du
+import torch
+from pycls.core.config import cfg
+
+
+# Common prefix for checkpoint file names
+_NAME_PREFIX = "model_epoch_"
+# Checkpoints directory name
+_DIR_NAME = "checkpoints"
+
+
+def get_checkpoint_dir():
+    """Retrieves the location for storing checkpoints."""
+    return os.path.join(cfg.OUT_DIR, _DIR_NAME)
+
+
+def get_checkpoint(epoch):
+    """Retrieves the path to a checkpoint file."""
+    name = "{}{:04d}.pyth".format(_NAME_PREFIX, epoch)
+    return os.path.join(get_checkpoint_dir(), name)
+
+
+def get_last_checkpoint():
+    """Retrieves the most recent checkpoint (highest epoch number)."""
+    checkpoint_dir = get_checkpoint_dir()
+    # Checkpoint file names are in lexicographic order
+    checkpoints = [f for f in os.listdir(checkpoint_dir) if _NAME_PREFIX in f]
+    last_checkpoint_name = sorted(checkpoints)[-1]
+    return os.path.join(checkpoint_dir, last_checkpoint_name)
+
+
+def has_checkpoint():
+    """Determines if there are checkpoints available."""
+    checkpoint_dir = get_checkpoint_dir()
+    if not os.path.exists(checkpoint_dir):
+        return False
+    return any(_NAME_PREFIX in f for f in os.listdir(checkpoint_dir))
+
+
+def is_checkpoint_epoch(cur_epoch):
+    """Determines if a checkpoint should be saved on current epoch."""
+    return (cur_epoch + 1) % cfg.TRAIN.CHECKPOINT_PERIOD == 0
+
+
+def save_checkpoint(model, optimizer, epoch):
+    """Saves a checkpoint."""
+    # Save checkpoints only from the master process
+    if not du.is_master_proc():
+        return
+    # Ensure that the checkpoint dir exists
+    os.makedirs(get_checkpoint_dir(), exist_ok=True)
+    # Omit the DDP wrapper in the multi-gpu setting
+    sd = model.module.state_dict() if cfg.NUM_GPUS > 1 else model.state_dict()
+    # Record the state
+    checkpoint = {
+        "epoch": epoch,
+        "model_state": sd,
+        "optimizer_state": optimizer.state_dict(),
+        "cfg": cfg.dump(),
+    }
+    # Write the checkpoint
+    checkpoint_file = get_checkpoint(epoch + 1)
+    torch.save(checkpoint, checkpoint_file)
+    return checkpoint_file
+
+
+def load_checkpoint(checkpoint_file, model, optimizer=None):
+    """Loads the checkpoint from the given file."""
+    assert os.path.exists(checkpoint_file), "Checkpoint '{}' not found".format(
+        checkpoint_file
+    )
+    # Load the checkpoint on CPU to avoid GPU mem spike
+    checkpoint = torch.load(checkpoint_file, map_location="cpu")
+    # Account for the DDP wrapper in the multi-gpu setting
+    ms = model.module if cfg.NUM_GPUS > 1 else model
+    ms.load_state_dict(checkpoint["model_state"])
+    # Load the optimizer state (commonly not done when fine-tuning)
+    if optimizer:
+        optimizer.load_state_dict(checkpoint["optimizer_state"])
+    return checkpoint["epoch"]
diff --git a/pycls/utils/distributed.py b/pycls/utils/distributed.py
new file mode 100644
index 0000000..5c6116a
--- /dev/null
+++ b/pycls/utils/distributed.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Distributed helpers."""
+
+import torch
+from pycls.core.config import cfg
+
+
+def is_master_proc():
+    """Determines if the current process is the master process.
+
+    Master process is responsible for logging, writing and loading checkpoints.
+    In the multi GPU setting, we assign the master role to the rank 0 process.
+    When training using a single GPU, there is only one training processes
+    which is considered the master processes.
+    """
+    return cfg.NUM_GPUS == 1 or torch.distributed.get_rank() == 0
+
+
+def init_process_group(proc_rank, world_size):
+    """Initializes the default process group."""
+    # Set the GPU to use
+    torch.cuda.set_device(proc_rank)
+    # Initialize the process group
+    torch.distributed.init_process_group(
+        backend=cfg.DIST_BACKEND,
+        init_method="tcp://{}:{}".format(cfg.HOST, cfg.PORT),
+        world_size=world_size,
+        rank=proc_rank,
+    )
+
+
+def destroy_process_group():
+    """Destroys the default process group."""
+    torch.distributed.destroy_process_group()
+
+
+def scaled_all_reduce(tensors):
+    """Performs the scaled all_reduce operation on the provided tensors.
+
+    The input tensors are modified in-place. Currently supports only the sum
+    reduction operator. The reduced values are scaled by the inverse size of
+    the process group (equivalent to cfg.NUM_GPUS).
+    """
+    # Queue the reductions
+    reductions = []
+    for tensor in tensors:
+        reduction = torch.distributed.all_reduce(tensor, async_op=True)
+        reductions.append(reduction)
+    # Wait for reductions to finish
+    for reduction in reductions:
+        reduction.wait()
+    # Scale the results
+    for tensor in tensors:
+        tensor.mul_(1.0 / cfg.NUM_GPUS)
+    return tensors
diff --git a/pycls/utils/error_handler.py b/pycls/utils/error_handler.py
new file mode 100644
index 0000000..8ef0702
--- /dev/null
+++ b/pycls/utils/error_handler.py
@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Multiprocessing error handler."""
+
+import os
+import signal
+import threading
+
+
+class ChildException(Exception):
+    """Wraps an exception from a child process."""
+
+    def __init__(self, child_trace):
+        super(ChildException, self).__init__(child_trace)
+
+
+class ErrorHandler(object):
+    """Multiprocessing error handler (based on fairseq's).
+
+    Listens for errors in child processes and
+    propagates the tracebacks to the parent process.
+    """
+
+    def __init__(self, error_queue):
+        # Shared error queue
+        self.error_queue = error_queue
+        # Children processes sharing the error queue
+        self.children_pids = []
+        # Start a thread listening to errors
+        self.error_listener = threading.Thread(target=self.listen, daemon=True)
+        self.error_listener.start()
+        # Register the signal handler
+        signal.signal(signal.SIGUSR1, self.signal_handler)
+
+    def add_child(self, pid):
+        """Registers a child process."""
+        self.children_pids.append(pid)
+
+    def listen(self):
+        """Listens for errors in the error queue."""
+        # Wait until there is an error in the queue
+        child_trace = self.error_queue.get()
+        # Put the error back for the signal handler
+        self.error_queue.put(child_trace)
+        # Invoke the signal handler
+        os.kill(os.getpid(), signal.SIGUSR1)
+
+    def signal_handler(self, _sig_num, _stack_frame):
+        """Signal handler."""
+        # Kill children processes
+        for pid in self.children_pids:
+            os.kill(pid, signal.SIGINT)
+        # Propagate the error from the child process
+        raise ChildException(self.error_queue.get())
diff --git a/pycls/utils/io.py b/pycls/utils/io.py
new file mode 100644
index 0000000..d9d98d8
--- /dev/null
+++ b/pycls/utils/io.py
@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""IO utilities (adapted from Detectron)"""
+
+import logging
+import os
+import re
+import sys
+from urllib import request as urlrequest
+
+
+logger = logging.getLogger(__name__)
+
+_PYCLS_BASE_URL = "https://dl.fbaipublicfiles.com/pycls"
+
+
+def cache_url(url_or_file, cache_dir):
+    """Download the file specified by the URL to the cache_dir and return the
+    path to the cached file. If the argument is not a URL, simply return it as
+    is.
+    """
+    is_url = re.match(r"^(?:http)s?://", url_or_file, re.IGNORECASE) is not None
+
+    if not is_url:
+        return url_or_file
+
+    url = url_or_file
+    assert url.startswith(_PYCLS_BASE_URL), (
+        "pycls only automatically caches URLs in the pycls S3 bucket: {}"
+    ).format(_PYCLS_BASE_URL)
+
+    cache_file_path = url.replace(_PYCLS_BASE_URL, cache_dir)
+    if os.path.exists(cache_file_path):
+        return cache_file_path
+
+    cache_file_dir = os.path.dirname(cache_file_path)
+    if not os.path.exists(cache_file_dir):
+        os.makedirs(cache_file_dir)
+
+    logger.info("Downloading remote file {} to {}".format(url, cache_file_path))
+    download_url(url, cache_file_path)
+    return cache_file_path
+
+
+def _progress_bar(count, total):
+    """Report download progress.
+    Credit:
+    https://stackoverflow.com/questions/3173320/text-progress-bar-in-the-console/27871113
+    """
+    bar_len = 60
+    filled_len = int(round(bar_len * count / float(total)))
+
+    percents = round(100.0 * count / float(total), 1)
+    bar = "=" * filled_len + "-" * (bar_len - filled_len)
+
+    sys.stdout.write(
+        "  [{}] {}% of {:.1f}MB file  \r".format(bar, percents, total / 1024 / 1024)
+    )
+    sys.stdout.flush()
+    if count >= total:
+        sys.stdout.write("\n")
+
+
+def download_url(url, dst_file_path, chunk_size=8192, progress_hook=_progress_bar):
+    """Download url and write it to dst_file_path.
+    Credit:
+    https://stackoverflow.com/questions/2028517/python-urllib2-progress-hook
+    """
+    req = urlrequest.Request(url)
+    response = urlrequest.urlopen(req)
+    total_size = response.info().get("Content-Length").strip()
+    total_size = int(total_size)
+    bytes_so_far = 0
+
+    with open(dst_file_path, "wb") as f:
+        while 1:
+            chunk = response.read(chunk_size)
+            bytes_so_far += len(chunk)
+            if not chunk:
+                break
+            if progress_hook:
+                progress_hook(bytes_so_far, total_size)
+            f.write(chunk)
+
+    return bytes_so_far
diff --git a/pycls/utils/logging.py b/pycls/utils/logging.py
new file mode 100644
index 0000000..c9f1565
--- /dev/null
+++ b/pycls/utils/logging.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Logging."""
+
+import builtins
+import decimal
+import logging
+import os
+import sys
+
+import pycls.utils.distributed as du
+import simplejson
+from pycls.core.config import cfg
+
+
+# Show filename and line number in logs
+_FORMAT = "[%(filename)s: %(lineno)3d]: %(message)s"
+
+# Log file name (for cfg.LOG_DEST = 'file')
+_LOG_FILE = "stdout.log"
+
+# Printed json stats lines will be tagged w/ this
+_TAG = "json_stats: "
+
+
+def _suppress_print():
+    """Suppresses printing from the current process."""
+
+    def ignore(*_objects, _sep=" ", _end="\n", _file=sys.stdout, _flush=False):
+        pass
+
+    builtins.print = ignore
+
+
+def setup_logging():
+    """Sets up the logging."""
+    # Enable logging only for the master process
+    if du.is_master_proc():
+        # Clear the root logger to prevent any existing logging config
+        # (e.g. set by another module) from messing with our setup
+        logging.root.handlers = []
+        # Construct logging configuration
+        logging_config = {"level": logging.INFO, "format": _FORMAT}
+        # Log either to stdout or to a file
+        if cfg.LOG_DEST == "stdout":
+            logging_config["stream"] = sys.stdout
+        else:
+            logging_config["filename"] = os.path.join(cfg.OUT_DIR, _LOG_FILE)
+        # Configure logging
+        logging.basicConfig(**logging_config)
+    else:
+        _suppress_print()
+
+
+def get_logger(name):
+    """Retrieves the logger."""
+    return logging.getLogger(name)
+
+
+def log_json_stats(stats):
+    """Logs json stats."""
+    # Decimal + string workaround for having fixed len float vals in logs
+    stats = {
+        k: decimal.Decimal("{:.6f}".format(v)) if isinstance(v, float) else v
+        for k, v in stats.items()
+    }
+    json_stats = simplejson.dumps(stats, sort_keys=True, use_decimal=True)
+    logger = get_logger(__name__)
+    logger.info("{:s}{:s}".format(_TAG, json_stats))
+
+
+def load_json_stats(log_file):
+    """Loads json_stats from a single log file."""
+    with open(log_file, "r") as f:
+        lines = f.readlines()
+    json_lines = [l[l.find(_TAG) + len(_TAG) :] for l in lines if _TAG in l]
+    json_stats = [simplejson.loads(l) for l in json_lines]
+    return json_stats
+
+
+def parse_json_stats(log, row_type, key):
+    """Extract values corresponding to row_type/key out of log."""
+    vals = [row[key] for row in log if row["_type"] == row_type and key in row]
+    if key == "iter" or key == "epoch":
+        vals = [int(val.split("/")[0]) for val in vals]
+    return vals
+
+
+def get_log_files(log_dir, name_filter=""):
+    """Get all log files in directory containing subdirs of trained models."""
+    names = [n for n in sorted(os.listdir(log_dir)) if name_filter in n]
+    files = [os.path.join(log_dir, n, _LOG_FILE) for n in names]
+    f_n_ps = [(f, n) for (f, n) in zip(files, names) if os.path.exists(f)]
+    files, names = zip(*f_n_ps)
+    return files, names
diff --git a/pycls/utils/lr_policy.py b/pycls/utils/lr_policy.py
new file mode 100644
index 0000000..9c751db
--- /dev/null
+++ b/pycls/utils/lr_policy.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Learning rate policies."""
+
+import numpy as np
+from pycls.core.config import cfg
+
+
+def lr_fun_steps(cur_epoch):
+    """Steps schedule (cfg.OPTIM.LR_POLICY = 'steps')."""
+    ind = [i for i, s in enumerate(cfg.OPTIM.STEPS) if cur_epoch >= s][-1]
+    return cfg.OPTIM.BASE_LR * (cfg.OPTIM.LR_MULT ** ind)
+
+
+def lr_fun_exp(cur_epoch):
+    """Exponential schedule (cfg.OPTIM.LR_POLICY = 'exp')."""
+    return cfg.OPTIM.BASE_LR * (cfg.OPTIM.GAMMA ** cur_epoch)
+
+
+def lr_fun_cos(cur_epoch):
+    """Cosine schedule (cfg.OPTIM.LR_POLICY = 'cos')."""
+    base_lr, max_epoch = cfg.OPTIM.BASE_LR, cfg.OPTIM.MAX_EPOCH
+    return 0.5 * base_lr * (1.0 + np.cos(np.pi * cur_epoch / max_epoch))
+
+
+def get_lr_fun():
+    """Retrieves the specified lr policy function"""
+    lr_fun = "lr_fun_" + cfg.OPTIM.LR_POLICY
+    if lr_fun not in globals():
+        raise NotImplementedError("Unknown LR policy:" + cfg.OPTIM.LR_POLICY)
+    return globals()[lr_fun]
+
+
+def get_epoch_lr(cur_epoch):
+    """Retrieves the lr for the given epoch according to the policy."""
+    lr = get_lr_fun()(cur_epoch)
+    # Linear warmup
+    if cur_epoch < cfg.OPTIM.WARMUP_EPOCHS:
+        alpha = cur_epoch / cfg.OPTIM.WARMUP_EPOCHS
+        warmup_factor = cfg.OPTIM.WARMUP_FACTOR * (1.0 - alpha) + alpha
+        lr *= warmup_factor
+    return lr
diff --git a/pycls/utils/meters.py b/pycls/utils/meters.py
new file mode 100644
index 0000000..c0d9883
--- /dev/null
+++ b/pycls/utils/meters.py
@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Meters."""
+
+import datetime
+from collections import deque
+
+import numpy as np
+import pycls.utils.logging as lu
+import pycls.utils.metrics as metrics
+from pycls.core.config import cfg
+from pycls.utils.timer import Timer
+
+
+def eta_str(eta_td):
+    """Converts an eta timedelta to a fixed-width string format."""
+    days = eta_td.days
+    hrs, rem = divmod(eta_td.seconds, 3600)
+    mins, secs = divmod(rem, 60)
+    return "{0:02},{1:02}:{2:02}:{3:02}".format(days, hrs, mins, secs)
+
+
+class ScalarMeter(object):
+    """Measures a scalar value (adapted from Detectron)."""
+
+    def __init__(self, window_size):
+        self.deque = deque(maxlen=window_size)
+        self.total = 0.0
+        self.count = 0
+
+    def reset(self):
+        self.deque.clear()
+        self.total = 0.0
+        self.count = 0
+
+    def add_value(self, value):
+        self.deque.append(value)
+        self.count += 1
+        self.total += value
+
+    def get_win_median(self):
+        return np.median(self.deque)
+
+    def get_win_avg(self):
+        return np.mean(self.deque)
+
+    def get_global_avg(self):
+        return self.total / self.count
+
+
+class TrainMeter(object):
+    """Measures training stats."""
+
+    def __init__(self, epoch_iters):
+        self.epoch_iters = epoch_iters
+        self.max_iter = cfg.OPTIM.MAX_EPOCH * epoch_iters
+        self.iter_timer = Timer()
+        self.loss = ScalarMeter(cfg.LOG_PERIOD)
+        self.loss_total = 0.0
+        self.lr = None
+        # Current minibatch errors (smoothed over a window)
+        self.mb_top1_err = ScalarMeter(cfg.LOG_PERIOD)
+        self.mb_top5_err = ScalarMeter(cfg.LOG_PERIOD)
+        # Number of misclassified examples
+        self.num_top1_mis = 0
+        self.num_top5_mis = 0
+        self.num_samples = 0
+
+    def reset(self, timer=False):
+        if timer:
+            self.iter_timer.reset()
+        self.loss.reset()
+        self.loss_total = 0.0
+        self.lr = None
+        self.mb_top1_err.reset()
+        self.mb_top5_err.reset()
+        self.num_top1_mis = 0
+        self.num_top5_mis = 0
+        self.num_samples = 0
+
+    def iter_tic(self):
+        self.iter_timer.tic()
+
+    def iter_toc(self):
+        self.iter_timer.toc()
+
+    def update_stats(self, top1_err, top5_err, loss, lr, mb_size):
+        # Current minibatch stats
+        self.mb_top1_err.add_value(top1_err)
+        self.mb_top5_err.add_value(top5_err)
+        self.loss.add_value(loss)
+        self.lr = lr
+        # Aggregate stats
+        self.num_top1_mis += top1_err * mb_size
+        self.num_top5_mis += top5_err * mb_size
+        self.loss_total += loss * mb_size
+        self.num_samples += mb_size
+
+    def get_iter_stats(self, cur_epoch, cur_iter):
+        eta_sec = self.iter_timer.average_time * (
+            self.max_iter - (cur_epoch * self.epoch_iters + cur_iter + 1)
+        )
+        eta_td = datetime.timedelta(seconds=int(eta_sec))
+        mem_usage = metrics.gpu_mem_usage()
+        stats = {
+            "_type": "train_iter",
+            "epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
+            "iter": "{}/{}".format(cur_iter + 1, self.epoch_iters),
+            "time_avg": self.iter_timer.average_time,
+            "time_diff": self.iter_timer.diff,
+            "eta": eta_str(eta_td),
+            "top1_err": self.mb_top1_err.get_win_median(),
+            "top5_err": self.mb_top5_err.get_win_median(),
+            "loss": self.loss.get_win_median(),
+            "lr": self.lr,
+            "mem": int(np.ceil(mem_usage)),
+        }
+        return stats
+
+    def log_iter_stats(self, cur_epoch, cur_iter):
+        if (cur_iter + 1) % cfg.LOG_PERIOD != 0:
+            return
+        stats = self.get_iter_stats(cur_epoch, cur_iter)
+        lu.log_json_stats(stats)
+
+    def get_epoch_stats(self, cur_epoch):
+        eta_sec = self.iter_timer.average_time * (
+            self.max_iter - (cur_epoch + 1) * self.epoch_iters
+        )
+        eta_td = datetime.timedelta(seconds=int(eta_sec))
+        mem_usage = metrics.gpu_mem_usage()
+        top1_err = self.num_top1_mis / self.num_samples
+        top5_err = self.num_top5_mis / self.num_samples
+        avg_loss = self.loss_total / self.num_samples
+        stats = {
+            "_type": "train_epoch",
+            "epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
+            "time_avg": self.iter_timer.average_time,
+            "eta": eta_str(eta_td),
+            "top1_err": top1_err,
+            "top5_err": top5_err,
+            "loss": avg_loss,
+            "lr": self.lr,
+            "mem": int(np.ceil(mem_usage)),
+        }
+        return stats
+
+    def log_epoch_stats(self, cur_epoch):
+        stats = self.get_epoch_stats(cur_epoch)
+        lu.log_json_stats(stats)
+
+
+class TestMeter(object):
+    """Measures testing stats."""
+
+    def __init__(self, max_iter):
+        self.max_iter = max_iter
+        self.iter_timer = Timer()
+        # Current minibatch errors (smoothed over a window)
+        self.mb_top1_err = ScalarMeter(cfg.LOG_PERIOD)
+        self.mb_top5_err = ScalarMeter(cfg.LOG_PERIOD)
+        # Min errors (over the full test set)
+        self.min_top1_err = 100.0
+        self.min_top5_err = 100.0
+        # Number of misclassified examples
+        self.num_top1_mis = 0
+        self.num_top5_mis = 0
+        self.num_samples = 0
+
+    def reset(self, min_errs=False):
+        if min_errs:
+            self.min_top1_err = 100.0
+            self.min_top5_err = 100.0
+        self.iter_timer.reset()
+        self.mb_top1_err.reset()
+        self.mb_top5_err.reset()
+        self.num_top1_mis = 0
+        self.num_top5_mis = 0
+        self.num_samples = 0
+
+    def iter_tic(self):
+        self.iter_timer.tic()
+
+    def iter_toc(self):
+        self.iter_timer.toc()
+
+    def update_stats(self, top1_err, top5_err, mb_size):
+        self.mb_top1_err.add_value(top1_err)
+        self.mb_top5_err.add_value(top5_err)
+        self.num_top1_mis += top1_err * mb_size
+        self.num_top5_mis += top5_err * mb_size
+        self.num_samples += mb_size
+
+    def get_iter_stats(self, cur_epoch, cur_iter):
+        mem_usage = metrics.gpu_mem_usage()
+        iter_stats = {
+            "_type": "test_iter",
+            "epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
+            "iter": "{}/{}".format(cur_iter + 1, self.max_iter),
+            "time_avg": self.iter_timer.average_time,
+            "time_diff": self.iter_timer.diff,
+            "top1_err": self.mb_top1_err.get_win_median(),
+            "top5_err": self.mb_top5_err.get_win_median(),
+            "mem": int(np.ceil(mem_usage)),
+        }
+        return iter_stats
+
+    def log_iter_stats(self, cur_epoch, cur_iter):
+        if (cur_iter + 1) % cfg.LOG_PERIOD != 0:
+            return
+        stats = self.get_iter_stats(cur_epoch, cur_iter)
+        lu.log_json_stats(stats)
+
+    def get_epoch_stats(self, cur_epoch):
+        top1_err = self.num_top1_mis / self.num_samples
+        top5_err = self.num_top5_mis / self.num_samples
+        self.min_top1_err = min(self.min_top1_err, top1_err)
+        self.min_top5_err = min(self.min_top5_err, top5_err)
+        mem_usage = metrics.gpu_mem_usage()
+        stats = {
+            "_type": "test_epoch",
+            "epoch": "{}/{}".format(cur_epoch + 1, cfg.OPTIM.MAX_EPOCH),
+            "time_avg": self.iter_timer.average_time,
+            "top1_err": top1_err,
+            "top5_err": top5_err,
+            "min_top1_err": self.min_top1_err,
+            "min_top5_err": self.min_top5_err,
+            "mem": int(np.ceil(mem_usage)),
+        }
+        return stats
+
+    def log_epoch_stats(self, cur_epoch):
+        stats = self.get_epoch_stats(cur_epoch)
+        lu.log_json_stats(stats)
diff --git a/pycls/utils/metrics.py b/pycls/utils/metrics.py
new file mode 100644
index 0000000..ea095a1
--- /dev/null
+++ b/pycls/utils/metrics.py
@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Functions for computing metrics."""
+
+import numpy as np
+import torch
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+# Number of bytes in a megabyte
+_B_IN_MB = 1024 * 1024
+
+
+def topks_correct(preds, labels, ks):
+    """Computes the number of top-k correct predictions for each k."""
+    assert preds.size(0) == labels.size(
+        0
+    ), "Batch dim of predictions and labels must match"
+    # Find the top max_k predictions for each sample
+    _top_max_k_vals, top_max_k_inds = torch.topk(
+        preds, max(ks), dim=1, largest=True, sorted=True
+    )
+    # (batch_size, max_k) -> (max_k, batch_size)
+    top_max_k_inds = top_max_k_inds.t()
+    # (batch_size, ) -> (max_k, batch_size)
+    rep_max_k_labels = labels.view(1, -1).expand_as(top_max_k_inds)
+    # (i, j) = 1 if top i-th prediction for the j-th sample is correct
+    top_max_k_correct = top_max_k_inds.eq(rep_max_k_labels)
+    # Compute the number of topk correct predictions for each k
+    topks_correct = [top_max_k_correct[:k, :].view(-1).float().sum() for k in ks]
+    return topks_correct
+
+
+def topk_errors(preds, labels, ks):
+    """Computes the top-k error for each k."""
+    num_topks_correct = topks_correct(preds, labels, ks)
+    return [(1.0 - x / preds.size(0)) * 100.0 for x in num_topks_correct]
+
+
+def topk_accuracies(preds, labels, ks):
+    """Computes the top-k accuracy for each k."""
+    num_topks_correct = topks_correct(preds, labels, ks)
+    return [(x / preds.size(0)) * 100.0 for x in num_topks_correct]
+
+
+def params_count(model):
+    """Computes the number of parameters."""
+    return np.sum([p.numel() for p in model.parameters()]).item()
+
+
+def flops_count(model):
+    """Computes the number of flops statically."""
+    h, w = cfg.TRAIN.IM_SIZE, cfg.TRAIN.IM_SIZE
+    count = 0
+    for n, m in model.named_modules():
+        if isinstance(m, nn.Conv2d):
+            if "se." in n:
+                count += m.in_channels * m.out_channels + m.bias.numel()
+                continue
+            h_out = (h + 2 * m.padding[0] - m.kernel_size[0]) // m.stride[0] + 1
+            w_out = (w + 2 * m.padding[1] - m.kernel_size[1]) // m.stride[1] + 1
+            count += np.prod([m.weight.numel(), h_out, w_out])
+            if ".proj" not in n:
+                h, w = h_out, w_out
+        elif isinstance(m, nn.MaxPool2d):
+            h = (h + 2 * m.padding - m.kernel_size) // m.stride + 1
+            w = (w + 2 * m.padding - m.kernel_size) // m.stride + 1
+        elif isinstance(m, nn.Linear):
+            count += m.in_features * m.out_features + m.bias.numel()
+    return count.item()
+
+
+def acts_count(model):
+    """Computes the number of activations statically."""
+    h, w = cfg.TRAIN.IM_SIZE, cfg.TRAIN.IM_SIZE
+    count = 0
+    for n, m in model.named_modules():
+        if isinstance(m, nn.Conv2d):
+            if "se." in n:
+                count += m.out_channels
+                continue
+            h_out = (h + 2 * m.padding[0] - m.kernel_size[0]) // m.stride[0] + 1
+            w_out = (w + 2 * m.padding[1] - m.kernel_size[1]) // m.stride[1] + 1
+            count += np.prod([m.out_channels, h_out, w_out])
+            if ".proj" not in n:
+                h, w = h_out, w_out
+        elif isinstance(m, nn.MaxPool2d):
+            h = (h + 2 * m.padding - m.kernel_size) // m.stride + 1
+            w = (w + 2 * m.padding - m.kernel_size) // m.stride + 1
+        elif isinstance(m, nn.Linear):
+            count += m.out_features
+    return count.item()
+
+
+def gpu_mem_usage():
+    """Computes the GPU memory usage for the current device (MB)."""
+    mem_usage_bytes = torch.cuda.max_memory_allocated()
+    return mem_usage_bytes / _B_IN_MB
diff --git a/pycls/utils/multiprocessing.py b/pycls/utils/multiprocessing.py
new file mode 100644
index 0000000..380104a
--- /dev/null
+++ b/pycls/utils/multiprocessing.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Multiprocessing helpers."""
+
+import multiprocessing as mp
+import traceback
+
+import pycls.utils.distributed as du
+from pycls.utils.error_handler import ErrorHandler
+
+
+def run(proc_rank, world_size, error_queue, fun, fun_args, fun_kwargs):
+    """Runs a function from a child process."""
+    try:
+        # Initialize the process group
+        du.init_process_group(proc_rank, world_size)
+        # Run the function
+        fun(*fun_args, **fun_kwargs)
+    except KeyboardInterrupt:
+        # Killed by the parent process
+        pass
+    except Exception:
+        # Propagate exception to the parent process
+        error_queue.put(traceback.format_exc())
+    finally:
+        # Destroy the process group
+        du.destroy_process_group()
+
+
+def multi_proc_run(num_proc, fun, fun_args=(), fun_kwargs=None):
+    """Runs a function in a multi-proc setting."""
+
+    if fun_kwargs is None:
+        fun_kwargs = {}
+
+    # Handle errors from training subprocesses
+    error_queue = mp.SimpleQueue()
+    error_handler = ErrorHandler(error_queue)
+
+    # Run each training subprocess
+    ps = []
+    for i in range(num_proc):
+        p_i = mp.Process(
+            target=run, args=(i, num_proc, error_queue, fun, fun_args, fun_kwargs)
+        )
+        ps.append(p_i)
+        p_i.start()
+        error_handler.add_child(p_i.pid)
+
+    # Wait for each subprocess to finish
+    for p in ps:
+        p.join()
diff --git a/pycls/utils/net.py b/pycls/utils/net.py
new file mode 100644
index 0000000..9abed63
--- /dev/null
+++ b/pycls/utils/net.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Functions for manipulating networks."""
+
+import itertools
+import math
+
+import torch
+import torch.nn as nn
+from pycls.core.config import cfg
+
+
+def init_weights(m):
+    """Performs ResNet-style weight initialization."""
+    if isinstance(m, nn.Conv2d):
+        # Note that there is no bias due to BN
+        fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+        m.weight.data.normal_(mean=0.0, std=math.sqrt(2.0 / fan_out))
+    elif isinstance(m, nn.BatchNorm2d):
+        zero_init_gamma = (
+            hasattr(m, "final_bn") and m.final_bn and cfg.BN.ZERO_INIT_FINAL_GAMMA
+        )
+        m.weight.data.fill_(0.0 if zero_init_gamma else 1.0)
+        m.bias.data.zero_()
+    elif isinstance(m, nn.Linear):
+        m.weight.data.normal_(mean=0.0, std=0.01)
+        m.bias.data.zero_()
+
+
+@torch.no_grad()
+def compute_precise_bn_stats(model, loader):
+    """Computes precise BN stats on training data."""
+    # Compute the number of minibatches to use
+    num_iter = min(cfg.BN.NUM_SAMPLES_PRECISE // loader.batch_size, len(loader))
+    # Retrieve the BN layers
+    bns = [m for m in model.modules() if isinstance(m, torch.nn.BatchNorm2d)]
+    # Initialize stats storage
+    mus = [torch.zeros_like(bn.running_mean) for bn in bns]
+    sqs = [torch.zeros_like(bn.running_var) for bn in bns]
+    # Remember momentum values
+    moms = [bn.momentum for bn in bns]
+    # Disable momentum
+    for bn in bns:
+        bn.momentum = 1.0
+    # Accumulate the stats across the data samples
+    for inputs, _labels in itertools.islice(loader, num_iter):
+        model(inputs.cuda())
+        # Accumulate the stats for each BN layer
+        for i, bn in enumerate(bns):
+            m, v = bn.running_mean, bn.running_var
+            sqs[i] += (v + m * m) / num_iter
+            mus[i] += m / num_iter
+    # Set the stats and restore momentum values
+    for i, bn in enumerate(bns):
+        bn.running_var = sqs[i] - mus[i] * mus[i]
+        bn.running_mean = mus[i]
+        bn.momentum = moms[i]
+
+
+def reset_bn_stats(model):
+    """Resets running BN stats."""
+    for m in model.modules():
+        if isinstance(m, torch.nn.BatchNorm2d):
+            m.reset_running_stats()
+
+
+def drop_connect(x, drop_ratio):
+    """Drop connect (adapted from DARTS)."""
+    keep_ratio = 1.0 - drop_ratio
+    mask = torch.empty([x.shape[0], 1, 1, 1], dtype=x.dtype, device=x.device)
+    mask.bernoulli_(keep_ratio)
+    x.div_(keep_ratio)
+    x.mul_(mask)
+    return x
+
+
+def get_flat_weights(model):
+    """Gets all model weights as a single flat vector."""
+    return torch.cat([p.data.view(-1, 1) for p in model.parameters()], 0)
+
+
+def set_flat_weights(model, flat_weights):
+    """Sets all model weights from a single flat vector."""
+    k = 0
+    for p in model.parameters():
+        n = p.data.numel()
+        p.data.copy_(flat_weights[k : (k + n)].view_as(p.data))
+        k += n
+    assert k == flat_weights.numel()
diff --git a/pycls/utils/plotting.py b/pycls/utils/plotting.py
new file mode 100644
index 0000000..77164a2
--- /dev/null
+++ b/pycls/utils/plotting.py
@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Plotting functions."""
+
+import colorlover as cl
+import matplotlib.pyplot as plt
+import plotly.graph_objs as go
+import plotly.offline as offline
+import pycls.utils.logging as lu
+
+
+def get_plot_colors(max_colors, color_format="pyplot"):
+    """Generate colors for plotting."""
+    colors = cl.scales["11"]["qual"]["Paired"]
+    if max_colors > len(colors):
+        colors = cl.to_rgb(cl.interp(colors, max_colors))
+    if color_format == "pyplot":
+        return [[j / 255.0 for j in c] for c in cl.to_numeric(colors)]
+    return colors
+
+
+def prepare_plot_data(log_files, names, key="top1_err"):
+    """Load logs and extract data for plotting error curves."""
+    plot_data = []
+    for file, name in zip(log_files, names):
+        d, log = {}, lu.load_json_stats(file)
+        for phase in ["train", "test"]:
+            x = lu.parse_json_stats(log, phase + "_epoch", "epoch")
+            y = lu.parse_json_stats(log, phase + "_epoch", key)
+            d["x_" + phase], d["y_" + phase] = x, y
+            d[phase + "_label"] = "[{:5.2f}] ".format(min(y) if y else 0) + name
+        plot_data.append(d)
+    assert len(plot_data) > 0, "No data to plot"
+    return plot_data
+
+
+def plot_error_curves_plotly(log_files, names, filename, key="top1_err"):
+    """Plot error curves using plotly and save to file."""
+    plot_data = prepare_plot_data(log_files, names, key)
+    colors = get_plot_colors(len(plot_data), "plotly")
+    # Prepare data for plots (3 sets, train duplicated w and w/o legend)
+    data = []
+    for i, d in enumerate(plot_data):
+        s = str(i)
+        line_train = {"color": colors[i], "dash": "dashdot", "width": 1.5}
+        line_test = {"color": colors[i], "dash": "solid", "width": 1.5}
+        data.append(
+            go.Scatter(
+                x=d["x_train"],
+                y=d["y_train"],
+                mode="lines",
+                name=d["train_label"],
+                line=line_train,
+                legendgroup=s,
+                visible=True,
+                showlegend=False,
+            )
+        )
+        data.append(
+            go.Scatter(
+                x=d["x_test"],
+                y=d["y_test"],
+                mode="lines",
+                name=d["test_label"],
+                line=line_test,
+                legendgroup=s,
+                visible=True,
+                showlegend=True,
+            )
+        )
+        data.append(
+            go.Scatter(
+                x=d["x_train"],
+                y=d["y_train"],
+                mode="lines",
+                name=d["train_label"],
+                line=line_train,
+                legendgroup=s,
+                visible=False,
+                showlegend=True,
+            )
+        )
+    # Prepare layout w ability to toggle 'all', 'train', 'test'
+    titlefont = {"size": 18, "color": "#7f7f7f"}
+    vis = [[True, True, False], [False, False, True], [False, True, False]]
+    buttons = zip(["all", "train", "test"], [[{"visible": v}] for v in vis])
+    buttons = [{"label": l, "args": v, "method": "update"} for l, v in buttons]
+    layout = go.Layout(
+        title=key + " vs. epoch<br>[dash=train, solid=test]",
+        xaxis={"title": "epoch", "titlefont": titlefont},
+        yaxis={"title": key, "titlefont": titlefont},
+        showlegend=True,
+        hoverlabel={"namelength": -1},
+        updatemenus=[
+            {
+                "buttons": buttons,
+                "direction": "down",
+                "showactive": True,
+                "x": 1.02,
+                "xanchor": "left",
+                "y": 1.08,
+                "yanchor": "top",
+            }
+        ],
+    )
+    # Create plotly plot
+    offline.plot({"data": data, "layout": layout}, filename=filename)
+
+
+def plot_error_curves_pyplot(log_files, names, filename=None, key="top1_err"):
+    """Plot error curves using matplotlib.pyplot and save to file."""
+    plot_data = prepare_plot_data(log_files, names, key)
+    colors = get_plot_colors(len(names))
+    for ind, d in enumerate(plot_data):
+        c, lbl = colors[ind], d["test_label"]
+        plt.plot(d["x_train"], d["y_train"], "--", c=c, alpha=0.8)
+        plt.plot(d["x_test"], d["y_test"], "-", c=c, alpha=0.8, label=lbl)
+    plt.title(key + " vs. epoch\n[dash=train, solid=test]", fontsize=14)
+    plt.xlabel("epoch", fontsize=14)
+    plt.ylabel(key, fontsize=14)
+    plt.grid(alpha=0.4)
+    plt.legend()
+    if filename:
+        plt.savefig(filename)
+        plt.clf()
+    else:
+        plt.show()
diff --git a/pycls/utils/timer.py b/pycls/utils/timer.py
new file mode 100644
index 0000000..14f0bc4
--- /dev/null
+++ b/pycls/utils/timer.py
@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Timer."""
+
+import time
+
+
+class Timer(object):
+    """A simple timer (adapted from Detectron)."""
+
+    def __init__(self):
+        self.reset()
+
+    def tic(self):
+        # using time.time instead of time.clock because time time.clock
+        # does not normalize for multithreading
+        self.start_time = time.time()
+
+    def toc(self):
+        self.diff = time.time() - self.start_time
+        self.total_time += self.diff
+        self.calls += 1
+        self.average_time = self.total_time / self.calls
+
+    def reset(self):
+        self.total_time = 0.0
+        self.calls = 0
+        self.start_time = 0.0
+        self.diff = 0.0
+        self.average_time = 0.0
diff --git a/test.py b/test.py
new file mode 100644
index 0000000..3441599
--- /dev/null
+++ b/test.py
@@ -0,0 +1,78 @@
+
+import torch
+import argparse
+import numpy as np
+
+from PIL import Image
+
+from RegDanbooru2019_8G import RegDanbooru2019
+
+parser = argparse.ArgumentParser(description='Test RegDeepDanbooru')
+parser.add_argument('--model', default='', type=str, help='trained model')
+parser.add_argument('--image', default='', type=str, help='image to test')
+parser.add_argument('--size', default=768, type=int, help='canvas size')
+parser.add_argument('--threshold', default=0.5, type=float, help='threshold')
+args = parser.parse_args()
+
+DANBOORU_LABEL_MAP = {}
+
+def load_danbooru_label_map() :
+    print(' -- Loading danbooru2019 labels')
+    global DANBOORU_LABEL_MAP
+    with open('danbooru_labels.txt', 'r') as fp :
+        for l in fp :
+            l = l.strip()
+            if l :
+                idx, tag = l.split(' ')
+                DANBOORU_LABEL_MAP[int(idx)] = tag
+
+def test(model, image_resized) :
+    print(' -- Running model on GPU')
+    image_resized_torch = torch.from_numpy(image_resized).float() / 127.5 - 1.0
+    if len(image_resized_torch.shape) == 3 :
+        image_resized_torch = image_resized_torch.unsqueeze(0).permute(0, 3, 1, 2)
+    elif len(image_resized_torch.shape) == 4 :
+        image_resized_torch = image_resized_torch.permute(0, 3, 1, 2)
+    image_resized_torch = image_resized_torch.cuda()
+    with torch.no_grad() :
+        danbooru_logits = model(image_resized_torch)
+        danbooru = danbooru_logits.sigmoid().cpu()
+    return danbooru
+
+def load_and_resize_image(img_path, canvas_size = 512) :
+    img = Image.open(img_path).convert('RGB')
+    old_size = img.size
+    w, h = old_size
+    w, h = float(w), float(h)
+    ratio = float(canvas_size) / max(old_size)
+    new_size = tuple([int(round(x * ratio)) for x in old_size])
+    print(f'Test image size: {new_size}')
+    return np.array(img.resize(new_size, Image.ANTIALIAS))
+
+def translate_danbooru_labels(probs, threshold = 0.8) :
+    global DANBOORU_LABEL_MAP
+    choosen_indices = (probs > threshold).nonzero()
+    result = []
+    for i in range(probs.size(0)) :
+        prob_single = probs[0].numpy()
+        indices_single = choosen_indices[choosen_indices[:, 0] == i][:, 1].numpy()
+        tag_prob_map = {DANBOORU_LABEL_MAP[idx]: prob_single[idx] for idx in indices_single}
+        result.append(tag_prob_map)
+    return result
+
+def main() :
+    model = RegDanbooru2019().cuda()
+    model.load_state_dict(torch.load(args.model)['model'])
+    model.eval()
+    torch.save(model, 'RegNetY-8G.pth',)
+
+    test_img = load_and_resize_image(args.image, args.size)
+
+    danbooru = test(model, test_img)
+
+    tags = translate_danbooru_labels(danbooru, args.threshold)
+    print(tags)
+
+if __name__ == "__main__":
+    load_danbooru_label_map()
+    main()