-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathatta_main.py
473 lines (406 loc) · 15.2 KB
/
atta_main.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
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
"""Main script for both training and evaluation."""
from __future__ import annotations
import argparse
import json
import math
import os
import pickle
import sys
import time
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import torch.cuda.amp as amp
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import wandb
from torch.distributed.elastic.multiprocessing.errors import record
from torchmetrics import JaccardIndex as IoU
from torchvision.utils import save_image
from part_model.attack import (
setup_eval_attacker,
setup_train_attacker,
setup_val_attacker,
)
from part_model.dataloader import COLORMAP, load_dataset
from part_model.models import build_model
from part_model.utils import (
AverageMeter,
ProgressMeter,
adjust_learning_rate,
dist_barrier,
get_compute_acc,
get_rank,
init_distributed_mode,
is_main_process,
pixel_accuracy,
save_on_master,
)
from part_model.utils.argparse import get_args_parser
from part_model.utils.atta import ATTA
from part_model.utils.image import get_seg_type
from part_model.utils.loss import get_train_criterion
best_acc1 = 0
@record
def main(args):
"""Main function."""
init_distributed_mode(args)
global best_acc1
# Fix the seed for reproducibility
seed = args.seed + get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
# Data loading code
print("=> creating dataset")
loaders = load_dataset(args)
train_loader, train_sampler, val_loader, test_loader = loaders
# Create model
print("=> creating model")
model, optimizer, scaler = build_model(args)
cudnn.benchmark = True
# Define loss function
criterion, train_criterion = get_train_criterion(args)
# Logging
if is_main_process():
logfile = open(os.path.join(args.output_dir, "log.txt"), "a")
logfile.write(str(args) + "\n")
logfile.flush()
if args.wandb:
wandb_id = os.path.split(args.output_dir)[-1]
wandb.init(
project="part-model", id=wandb_id, config=args, resume="allow"
)
print("wandb step:", wandb.run.step)
eval_attack = setup_eval_attacker(args, model)
no_attack = eval_attack[0][1]
train_attack = setup_train_attacker(args, model)
val_attack = setup_val_attacker(args, model)
save_metrics = {
"train": [],
"test": [],
}
# Initialize ATTA training if specified
atta: ATTA | None = None
if args.adv_train == "atta":
atta = ATTA(input_dim=args.input_dim, num_samples=args.num_train)
print(args)
if args.evaluate:
if args.resume:
load_path = args.resume
else:
load_path = f"{args.output_dir}/checkpoint_best.pt"
else:
print("=> beginning training")
val_stats = {}
for epoch in range(args.start_epoch, args.epochs):
is_best = False
if args.distributed:
train_sampler.set_epoch(epoch)
lr = adjust_learning_rate(optimizer, epoch, args)
print(f"=> lr @ epoch {epoch}: {lr:.2e}")
# Train for one epoch
train_stats = _train(
train_loader,
model,
train_criterion,
train_attack,
optimizer,
scaler,
epoch,
args,
atta=atta,
)
if (epoch + 1) % 2 == 0:
val_stats = _validate(
val_loader, model, criterion, no_attack, args
)
clean_acc1, acc1 = val_stats["acc1"], None
is_best = clean_acc1 > best_acc1
if args.adv_train != "none":
val_stats = _validate(
val_loader, model, criterion, val_attack, args
)
acc1 = val_stats["acc1"]
is_best = acc1 > best_acc1 and clean_acc1 >= acc1
save_dict = {
"epoch": epoch + 1,
"state_dict": model.state_dict(),
"optimizer": optimizer.state_dict(),
"scaler": scaler.state_dict(),
"best_acc1": best_acc1,
"args": args,
}
if is_best:
print("=> Saving new best checkpoint")
save_on_master(save_dict, args.output_dir, is_best=True)
best_acc1 = (
max(clean_acc1, best_acc1)
if acc1 is None
else max(acc1, best_acc1)
)
save_epoch = epoch + 1 if args.save_all_epochs else None
save_on_master(
save_dict, args.output_dir, is_best=False, epoch=save_epoch
)
log_stats = {
**{f"train_{k}": v for k, v in train_stats.items()},
**{f"test_{k}": v for k, v in val_stats.items()},
"epoch": epoch,
}
if is_main_process():
save_metrics["train"].append(log_stats)
if args.wandb:
wandb.log(log_stats)
logfile.write(json.dumps(log_stats) + "\n")
logfile.flush()
# Compute stats of best model after training
dist_barrier()
load_path = f"{args.output_dir}/checkpoint_best.pt"
print(f"=> loading checkpoint from {load_path}...")
if args.gpu is None:
checkpoint = torch.load(load_path)
else:
# Map model to be loaded to specified single gpu.
loc = "cuda:{}".format(args.gpu)
checkpoint = torch.load(load_path, map_location=loc)
model.load_state_dict(checkpoint["state_dict"])
# Running evaluation
for attack in eval_attack:
# Use DataParallel (not distributed) model for AutoAttack.
# Otherwise, DDP model can get timeout or c10d failure.
stats = _validate(test_loader, model, criterion, attack[1], args)
print(f"=> {attack[0]}: {stats}")
stats["attack"] = str(attack[0])
dist_barrier()
if is_main_process():
save_metrics["test"].append(stats)
if args.wandb:
wandb.log(stats)
logfile.write(json.dumps(stats) + "\n")
if is_main_process():
# Save metrics to pickle file if not exists else append
pkl_path = os.path.join(args.output_dir, "metrics.pkl")
if os.path.exists(pkl_path):
metrics = pickle.load(open(pkl_path, "rb"))
metrics.append(save_metrics)
else:
pickle.dump([save_metrics], open(pkl_path, "wb"))
last_path = f"{args.output_dir}/checkpoint_last.pt"
if os.path.exists(last_path):
os.remove(last_path)
logfile.close()
def _train(
train_loader,
model,
criterion,
attack,
optimizer,
scaler,
epoch,
args,
atta: ATTA | None = None,
):
use_seg: bool = get_seg_type(args) is not None
seg_only: bool = "seg-only" in args.experiment
tf_params_offset: int = 2 if not use_seg or seg_only else 3
use_atta: bool = args.adv_train == "atta"
batch_time = AverageMeter("Time", ":6.3f")
data_time = AverageMeter("Data", ":6.3f")
losses = AverageMeter("Loss", ":.4e")
top1 = AverageMeter("Acc@1", ":6.2f")
mem = AverageMeter("Mem (GB)", ":6.1f")
progress = ProgressMeter(
len(train_loader),
[batch_time, data_time, losses, top1, mem],
prefix="Epoch: [{}]".format(epoch),
)
compute_acc = get_compute_acc(args)
# Switch to train mode
model.train()
end = time.time()
for i, samples in enumerate(train_loader):
# Measure data loading time
data_time.update(time.time() - end)
segs: torch.Tensor | None = None
if not use_seg or seg_only:
# If training segmenter only, `targets` is segmentation mask
images, targets = samples[:tf_params_offset]
else:
images, segs, targets = samples[:tf_params_offset]
segs = segs.cuda(args.gpu, non_blocking=True)
if use_atta:
orig_images: torch.Tensor = images.clone()
tf_params: list[torch.Tensor] = samples[tf_params_offset:]
# Update image with saved perturbation
images = atta.apply(images, tf_params)
images = images.cuda(args.gpu, non_blocking=True)
targets = targets.cuda(args.gpu, non_blocking=True)
batch_size = images.size(0)
# Compute output
with amp.autocast(enabled=not args.full_precision):
if attack.use_mask:
# Attack for part models where both class and segmentation
# labels are used
images = attack(images, targets, segs)
if attack.dual_losses:
targets = torch.cat([targets, targets], dim=0)
segs = torch.cat([segs, segs], dim=0)
else:
# Attack for either classifier or segmenter alone
images = attack(images, targets)
if segs is None or seg_only:
outputs = model(images)
loss = criterion(outputs, targets)
elif "groundtruth" in args.experiment:
outputs = model(images, segs=segs)
loss = criterion(outputs, targets)
else:
outputs = model(images, return_mask=True)
loss = criterion(outputs, targets, segs)
outputs = outputs[0]
if args.adv_train in ("trades", "mat"):
outputs = outputs[batch_size:]
if not math.isfinite(loss.item()):
print("Loss is {}, stopping training".format(loss.item()))
sys.exit(1)
# Measure accuracy and record loss
acc1 = compute_acc(outputs, targets)
losses.update(loss.item(), batch_size)
top1.update(acc1.item(), batch_size)
# Compute gradient and do SGD step
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
# Measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
mem.update(torch.cuda.max_memory_allocated() // 1e9)
if i % args.print_freq == 0:
if is_main_process() and args.wandb:
wandb.log(
{
"acc": acc1.item(),
"loss": loss.item(),
"scaler": scaler.get_scale(),
}
)
progress.display(i)
if use_atta:
# Update saved perturbation in ATTA
perturbation: torch.Tensor = images.cpu() - orig_images
atta.update(perturbation, tf_params)
progress.synchronize()
return {
"acc1": top1.avg,
"loss": losses.avg,
"lr": optimizer.param_groups[0]["lr"],
}
def _validate(val_loader, model, criterion, attack, args):
seg_only = "seg-only" in args.experiment
batch_time = AverageMeter("Time", ":6.3f")
data_time = AverageMeter("Data", ":6.3f")
losses = AverageMeter("Loss", ":.4e")
top1 = AverageMeter("Acc@1", ":6.2f")
pacc = AverageMeter("PixelAcc", ":6.2f")
mem = AverageMeter("Mem (GB)", ":6.1f")
iou = AverageMeter("IoU", ":6.2f")
progress = ProgressMeter(
len(val_loader),
[batch_time, data_time, losses, top1, mem],
prefix="Test: ",
)
compute_acc = get_compute_acc(args)
compute_iou = IoU(args.seg_labels).cuda(args.gpu)
# switch to evaluate mode
model.eval()
end = time.time()
for i, samples in enumerate(val_loader):
# measure data loading time
data_time.update(time.time() - end)
segs: torch.Tensor | None = None
if len(samples) == 2:
images, targets = samples
elif seg_only:
images, targets, _ = samples
else:
images, segs, targets = samples
segs = segs.cuda(args.gpu, non_blocking=True)
# DEBUG
if args.debug:
save_image(COLORMAP[segs].permute(0, 3, 1, 2), "gt.png")
save_image(images, "test.png")
images = images.cuda(args.gpu, non_blocking=True)
targets = targets.cuda(args.gpu, non_blocking=True)
batch_size = images.size(0)
# DEBUG: fixed clean segmentation masks
if "clean" in args.experiment:
model(images, clean=True)
# compute output
with torch.no_grad():
if attack.use_mask:
images = attack(images, targets, segs)
else:
images = attack(images, targets)
# Need to duplicate segs and targets to match images expanded by
# image corruption attack
if images.shape[0] != targets.shape[0]:
ratio = images.shape[0] // targets.shape[0]
targets = targets.repeat(
(ratio,) + (1,) * (len(targets.shape) - 1)
)
if segs:
segs = segs.repeat((ratio,) + (1,) * (len(segs.shape) - 1))
if segs is None or "normal" in args.experiment or seg_only:
outputs = model(images)
elif "groundtruth" in args.experiment:
outputs = model(images, segs=segs)
loss = criterion(outputs, targets)
else:
outputs, masks = model(images, return_mask=True)
if "centroid" in args.experiment:
masks, _, _, _ = masks
pixel_acc = pixel_accuracy(masks, segs)
pacc.update(pixel_acc.item(), batch_size)
loss = criterion(outputs, targets)
# DEBUG
# if args.debug and isinstance(attack, PGDAttackModule):
if args.debug:
save_image(
COLORMAP[masks.argmax(1)].permute(0, 3, 1, 2),
"pred_seg_clean.png",
)
print(targets == outputs.argmax(1))
import pdb
pdb.set_trace()
# measure accuracy and record loss
acc1 = compute_acc(outputs, targets)
losses.update(loss.item(), batch_size)
top1.update(acc1.item(), batch_size)
if seg_only:
iou.update(compute_iou(outputs, targets).item(), images.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
mem.update(torch.cuda.max_memory_allocated() // 1e9)
if i % args.print_freq == 0:
progress.display(i)
# TODO: this should also be done with the ProgressMeter
progress.synchronize()
print(f" * Acc@1 {top1.avg:.3f}")
if pacc.count > 0:
pacc.synchronize()
print(f"Pixelwise accuracy: {pacc.avg:.4f}")
if seg_only:
iou.synchronize()
print(f"IoU: {iou.avg:.4f}")
return {"acc1": top1.avg, "loss": losses.avg, "pixel-acc": pacc.avg}
if __name__ == "__main__":
parser = argparse.ArgumentParser(
"Main script for part model (ATTA)", parents=[get_args_parser()]
)
args = parser.parse_args()
os.makedirs(args.output_dir, exist_ok=True)
main(args)