-
Notifications
You must be signed in to change notification settings - Fork 156
/
res32_bilstm_ctc.py
418 lines (368 loc) · 16.1 KB
/
res32_bilstm_ctc.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
"""
##################################################################################################
# Copyright Info : Copyright (c) Davar Lab @ Hikvision Research Institute. All rights reserved.
# Filename : res32_bilstm_ctc.py
# Abstract : Base recognition Model, res32 bilstm ctc
# Current Version: 1.0.0
# Date : 2021-06-11
##################################################################################################
"""
# encoding=utf-8
_base_ = [
]
"""
1. model setting
description:
Text recognition model configuration information
Add keywords:
None
"""
# model type
type = 'RECOGNIZOR'
# recognition dictionary
character = '/data1/open-source/demo/text_recognition/__dictionary__/Scene_text_68.txt'
# model setting
model = dict(
type='GeneralRecognizor',
pretrained=None,
backbone=dict( # Backbone parameter
type='ResNet32',
input_channel=1,
output_channel=512,),
transformation=None,
neck=None, # Relation module parameter
sequence_module=dict( # Sequential module parameter
type='CascadeRNN',
rnn_modules=[
dict(
type='BidirectionalLSTM',
input_size=512,
hidden_size=256,
output_size=256,
with_linear=True,
bidirectional=True,),
dict(
type='BidirectionalLSTM',
input_size=256,
hidden_size=256,
output_size=512,
with_linear=True,
bidirectional=True,), ]),
sequence_head=dict( # Recognition head parameter
type='WarpCTCHead',
input_size=512,
use_1x1conv=False,
converter=dict( # Recognition Converter parameter
type='CTCLabelConverter',
character=character,
use_cha_eos=True,
# cates=10000),
),
loss_ctc=dict( # Model loss function
type='WarpCTCLoss',
blank=0, # when use_cha_eos=True:96 else:0
length_average=True,
size_average=True,
loss_weight=1.0))
)
# Model training and test parameter configuration
train_cfg = dict( # Dimensions remain or change
sequence=dict(),
keep_dim=False,
)
test_cfg = dict(
sequence=dict(),
keep_dim=False,
batch_max_length=25,
)
# Training dataset load type
dataset_type = 'DavarMultiDataset'
# dataset settings
# support the dataset type
ppld = {
'LMDB_Standard': 'LoadImageFromLMDB', # open-source LMDB data
# Davar dataset type
'LMDB_Davar': 'RCGLoadImageFromLMDB',
'File': 'RCGLoadImageFromFile',
'Loose': 'RCGLoadImageFromLoose',
'Tight': 'RCGLoadImageFromTight',
}
"""
Instruction manual:
data_types=['LMDB','File','Tight','File'] # corresponding to different
ann_files = ['train1|train2|train3',
'Datalist/train1.json|Datalist/train2.json',
'Datalist/train_xxx.json',
'Datalist/train_yyy.json'] # Separated by '|'
img_prefixes = ['xx/yy/zz/|aa/bb/cc/|mm/nn/',
'dd/ee/', 'ff/gg/hh/',
'ii/jj/kk/'] # Separated by '|', corresponding to the ann_files
batch_ratios = ['0.1|0.1|0.1',
'0.2|0.2',
'0.1',
'0.2'] # string format, corresponding to the ann_files
# sum of the batch_ratios equals to 1
"""
# Training dataset format
data_types = [
'LMDB_Standard',
'LMDB_Standard'
]
# File prefix path of the traning dataset
img_prefixes = [
'*******/TextRecognition/LMDB/BenchEn/train/', # path to the training dataset
'*******/TextRecognition/LMDB/BenchEn/train/', # path to the training dataset
]
# Dataset Name
ann_files = [
'MJ', 'SK'
]
# Normalization parameter
img_norm_cfg = dict(
mean=[127.5],
std=[127.5])
# training pipeline
train_pipelines = [
dict(
type=ppld["LMDB_Standard"],
character=character, # recognition dictionary
test_mode=False, # whether is in test mode
sensitive=False, # sensitive to Upper or Lower
color_types=["gray"], # color loading type, ["rgb", "bgr", "gray"]
fil_ops=True, # whether to filter unsupported text
),
dict(
type='ResizeNormalize',
size=(100, 32),
interpolation=2,
# Interpolation method of the Resize function
# 0 - INTER_NEAREST(default) # 1 - INTER_LINEAR
# 2 - INTER_CUBIC # 3 - INTER_AREA
mean=img_norm_cfg["mean"],
std=img_norm_cfg["std"], ),
dict(type='DavarDefaultFormatBundle'), # Uniform Training data tensor format
dict(type='DavarCollect', keys=['img', 'gt_text']), # Data content actually involved in training stage
]
print('train_piplines:', train_pipelines)
val_pipeline = [
dict(type=ppld["LMDB_Standard"],
character=character,
test_mode=True,
sensitive=False,
color_types=["gray"], # color loading type, ["rgb", "bgr", "gray"]
fil_ops=True, ),
dict(type='ResizeNormalize',
size=(100, 32),
interpolation=2,
mean=img_norm_cfg["mean"],
std=img_norm_cfg["std"],
),
dict(type='DavarDefaultFormatBundle'),
dict(type='DavarCollect', keys=['img', 'gt_text'], meta_keys=[]),
]
test_pipeline = [
dict(type=ppld["LMDB_Standard"],
character=character,
test_mode=True,
sensitive=False,
color_types=["gray"],
fil_ops=True, ),
dict(type='ResizeNormalize',
size=(100, 32),
interpolation=2,
mean=img_norm_cfg["mean"],
std=img_norm_cfg["std"],
),
dict(type='DavarDefaultFormatBundle'),
dict(type='DavarCollect', keys=['img'], meta_keys=[]),
]
data = dict(
samples_per_gpu=128, # batchsize=100->memory:6400M
workers_per_gpu=2,
sampler=dict(
type='DistBatchBalancedSampler', # BatchBalancedSampler and DistBatchBalancedSampler
mode=0,
# model 0: Balance in batch, calculate the epoch according to the first iterative data set
# model 1: Balance in batch, calculate the epoch according to the last iterative data set
# model 2: Balance in batch, record unused data
# model -1: Each dataset is directly connected and shuffled
),
train=dict(
type=dataset_type,
batch_ratios=['0.5', '0.5'],
dataset=dict(
type="DavarRCGDataset",
data_type=data_types,
ann_file=ann_files,
img_prefix=img_prefixes,
batch_max_length=25,
used_ratio=1,
test_mode=False,
pipeline=train_pipelines)
),
val=dict(
type=dataset_type,
batch_ratios=1,
samples_per_gpu=400,
test_mode=True,
dataset=dict(
type="DavarRCGDataset",
data_type="LMDB_Standard",
ann_file='mixture',
img_prefix='/path/to/validation',
batch_max_length=25,
used_ratio=1,
test_mode=True,
pipeline=val_pipeline,),
),
test=dict(
type=dataset_type,
data_type='LMDB_Standard',
ann_file='IIIT5k_3000',
img_prefix='/path/to/evaluation/',
batch_ratios=1,
batch_max_length=25,
used_ratio=1,
pipeline=test_pipeline, ),
)
"""
3. Training parameter settings
description:
Configure the corresponding learning rate and related strategy according to the dataset or model structure
Add keywords:
None
"""
# Optimizer parameter settings
# optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
# optimizer = dict(type='Adadelta', lr=1.0, rho=0.95, eps=1e-8)
# optimizer = dict(type='Adam', amsgrad=False, betas=(0.9, 0.999), eps=1e-8, lr=0.001, weight_decay=0)
optimizer = dict(type='AdamW', betas=(0.9, 0.999), eps=1e-8, lr=0.001, weight_decay=0)
optimizer_config = dict(grad_clip=dict(max_norm=5, norm_type=2))
# Learning rate parameter setting
lr_config = dict(
# policy='fixed',
policy='step',
warmup='linear',
warmup_iters=300,
warmup_ratio=1.0 / 3,
gamma=0.3,
step=[3, 4, 5]
)
# logger setting
# yapf:disable
log_config = dict(
interval=50,
hooks=[dict(type='TextLoggerHook'), ])
"""
======================================================================================================================
Evaluation & Checkpoint Instruction manual:
======================================================================================================================
1. Evaluation Setting
----------------------------------------------------------------------------------------------------------------------
$ General Type
evaluation = dict(
interval=1, # Evaluation interval By Epoch
model_type="RECOGNIZOR", # Evaluation Model Type,
# including["DETECTOR", "RECOGNIZOR", "SPOTTER"]
save_best="accuracy", # Save the best metric evaluation model
eval_mode="general", # Evaluation type,
# Note: general equals MMDetection Official
# Evaluation Hook
by_epoch=True, # by_epoch: True -- By Epoch
# False -- By Iteration
# Note: (Could not work together)
rule="greater", # the Metric rule, including "greater" or "lower"
metric=['accuracy', 'NED'], # Supported Metric Name
)
----------------------------------------------------------------------------------------------------------------------
$ Lightweight Type
evaluation = dict(
start=3, # Which epoch to start evaluation
start_iter=0.5, # the percentage of the training iteration to evaluate
save_best="accuracy", # Save the best metric evaluation model
iter_interval=1, # Evaluation interval By Epoch
model_type="RECOGNIZOR", # Evaluation Model Type,
# including["DETECTOR", "RECOGNIZOR", "SPOTTER"]
eval_mode="lightweight", # Evaluation type,
# Note: lightweight could evaluate the model by
# iterations and by epochs
by_epoch=True, # by_epoch: True -- By Epoch
by_iter=True, # by_iter: True -- By Iteration
# (independent with By_epoch,
# could work together)
rule="greater", # the Metric rule, including "greater" or "lower"
metric=['accuracy', 'NED'], # Supported Metric Name
)
======================================================================================================================
2. Checkpoint Setting
----------------------------------------------------------------------------------------------------------------------
$ General Type
checkpoint_config = dict(type="DavarCheckpointHook", # Checkpoint Hook Name
interval=1, # Checkpoint save interval By Epoch
by_epoch=True, # by_epoch: True -- By Epoch
# False -- By Iteration
# Note: (Could not work together)
filename_tmpl='ckpt/ace_e{}.pth', # Checkpoint Save Name format
metric="accuracy", # Save the best metric Name "Accuracy"
rule="greater", # the Metric rule, including "greater" or "lower"
save_mode="general", # General equals MMDetection Official Checkpoint Hook
)
----------------------------------------------------------------------------------------------------------------------
$ Lightweight Type
checkpoint_config = dict(type="DavarCheckpointHook", # Checkpoint Hook Name
interval=1, # Checkpoint save interval By Epoch
iter_interval=1, # Checkpoint save interval By Iteration
by_epoch=True, # by_epoch: True -- By Epoch
by_iter=True, # by_iter: True -- By Iteration
# (independent with By_epoch,
# could work together)
filename_tmpl='ckpt/ace_e{}.pth', # Checkpoint Save Name format
metric="accuracy", # Save the best metric Name "Accuracy"
rule="greater", # the Metric rule, including "greater" or "lower"
save_mode="lightweight", # Lightweight type, only save the best metric model and
# latest iteration and latest epoch model
init_metric=-1, # initial metric of the model
model_milestone=0.5 # the percentage of the
# training process to save checkpoint
)
======================================================================================================================
"""
checkpoint_config = dict(type="DavarCheckpointHook",
interval=1,
iter_interval=1,
by_epoch=True,
by_iter=True,
filename_tmpl='ckpt/res32_ace_e{}.pth',
metric="accuracy",
rule="greater",
save_mode="lightweight",
init_metric=-1,
model_milestone=0.5
)
# evaluation setting
evaluation = dict(start=3,
start_iter=0.5,
save_best="accuracy",
iter_interval=1,
model_type="RECOGNIZOR",
eval_mode="lightweight",
by_epoch=True,
by_iter=True,
rule="greater",
metric=['accuracy', 'NED'],
)
# === runtime settings ===
# yapf:enable
runner = dict(type='EpochBasedRunner', max_epochs=6) # Total training epoch
dist_params = dict(backend='nccl')
log_level = 'INFO'
# The path where the model is saved
work_dir = '/path/to/davar_opensource/ctc_base/'
# Load from Pre-trained model path
load_from = None
# Resume from Pre-trained model path
resume_from = None
# workflow setting
workflow = [('train', 1)]
# gpu number
gpus = 2