File size: 24,622 Bytes
e1aaaac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
import sys

from train.datasets import COCOFlickrDataset, ImageNetDataset
from CLIP_eval.eval_utils import load_clip_model

sys.path.append("open_flamingo")
import os
import shutil
import time
import string
import random

import numpy as np
import open_clip
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
from training.scheduler import cosine_lr
from torchvision import transforms
from open_flamingo.eval.classification_utils import IMAGENET_1K_CLASS_ID_TO_LABEL
from train.pgd_train import pgd
from train.apgd_train import apgd_train as apgd
import wandb
from train.utils import init_wandb, AverageMeter
from train.sam_data import SamData
from open_flamingo.eval.models.utils import unwrap_model
from train.utils import str2bool

import argparse

parser = argparse.ArgumentParser()
parser.add_argument('--clip_model_name', type=str, default='ViT-L-14', help='ViT-L-14, ViT-B-32')
parser.add_argument('--pretrained', type=str, default='openai')
parser.add_argument('--dataset', type=str, default='imagenet')
parser.add_argument('--template', type=str, default='std')
parser.add_argument('--imagenet_root', type=str, default='/mnt/datasets/imagenet', help='Imagenet dataset root directory')
parser.add_argument('--output_normalize', type=str2bool, default=False, help='Whether the embedding is normalized')
parser.add_argument('--start_step', type=int, default=0, help='Start step for training')
parser.add_argument('--optimizer_state', type=str, default='', help='Optimizer state file path')
parser.add_argument('--steps', type=int, default=20000, help='Number of training steps')
parser.add_argument('--warmup', type=int, default=14000, help='Warmup steps')
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--loss', type=str, default='l2', help='ce, l2')
parser.add_argument('--loss_clean', type=str, default='none', help='ce, l2')
parser.add_argument('--clean_weight', type=float, default=0., help='Weight for clean loss')
parser.add_argument('--trades', type=str2bool, default=False, help='Use TRADES')
parser.add_argument('--opt', type=str, default='adamw', help='Optimizer type; sgd, adamw')
parser.add_argument('--momentum_sgd', type=float, default=0.9, help='Momentum for SGD optimizer')
parser.add_argument('--lr', type=float, default=1e-5, help='Learning rate')
parser.add_argument('--wd', type=float, default=1e-4, help='Weight decay')
parser.add_argument('--attack', type=str, default='apgd', help='Adversarial attack type')
parser.add_argument('--inner_loss', type=str, default='l2', help='Inner loss function for adversarial training')
parser.add_argument('--norm', type=str, default='linf', help='Norm for adversarial perturbation')
parser.add_argument('--eps', type=float, default=4, help='Epsilon for adversarial perturbation')
parser.add_argument('--iterations_adv', type=int, default=10, help='Iterations for adversarial attack')
parser.add_argument('--stepsize_adv', type=float, default=1., help='Step size for adversarial attack (no effect for apgd)')
parser.add_argument('--wandb', type=str2bool, default=True, help='Use Weights & Biases for logging')
parser.add_argument('--experiment_name', type=str, default='')
parser.add_argument('--overwrite', type=str2bool, default=False, help='Overwrite existing directory')
parser.add_argument('--log_freq', type=int, default=1, help='Logging frequency')
parser.add_argument('--eval_freq', type=int, default=50, help='Evaluation frequency')
parser.add_argument('--output_dir', type=str, default='', help='Output directory')
parser.add_argument('--save_checkpoints', type=str2bool, default=True, help='Save 10 training checkpoints')
parser.add_argument('--devices', type=str, default='', help='Device IDs for CUDA')


def main(args):
    # setup wandb
    if args.wandb:
        init_wandb(
            project_name='clip-finetune',
            model_name=args.finetuned_model_name,
            config=vars(args)
        )
    else:
        wandb.init(mode='disabled')

    # print args
    print(f"Arguments:\n{'-' * 20}")
    for arg, value in vars(args).items():
        print(f"{arg}: {value}")
    print(f"{'-' * 20}")

    # setup dirs
    if args.overwrite:
        shutil.rmtree(args.output_dir, ignore_errors=True)
    os.makedirs(os.path.join(args.output_dir, 'checkpoints'), exist_ok=False)

    # write args to file
    with open(os.path.join(args.output_dir, 'args.txt'), 'w') as f:
        f.write(str(args))

    main_device = 0
    # get models
    model_orig, _, image_processor = open_clip.create_model_and_transforms(
        args.clip_model_name, pretrained='openai' # 可选 output_tokens=True,返回token + patches
    )
    if args.optimizer_state != '':
        assert args.start_step > 0
        assert str(args.start_step) in args.optimizer_state
        assert args.pretrained in ['', 'none']
        args.pretrained = args.optimizer_state.replace('_opt', '')
    model, _, _ = load_clip_model(args.clip_model_name, args.pretrained)

    # Remove the Normalize transform by creating a new Compose object
    preprocessor_without_normalize = transforms.Compose(image_processor.transforms[:-1])
    normalize = image_processor.transforms[-1]
    del image_processor
    print(f'[preprocessor_without_normalize] {preprocessor_without_normalize}')
    print(f'[normalize] {normalize}')
    # preprocessor_without_normalize contains following transforms:
    # - Resize(size=224, interpolation=bicubic, max_size=None, antialias=warn)
    # - CenterCrop(size=(224, 224))
    # - ToTensor()
    # normalize:
    # Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))

    # get data
    if args.dataset == 'imagenet':
        dataset = ImageNetDataset(
            root=args.imagenet_root + '/train',
            transform=preprocessor_without_normalize,
        )

    elif args.dataset == 'segment_anything':
        dataset = SamData('/data/naman_deep_singh/datasets/newSAM', transform=preprocessor_without_normalize)

        print(dataset.__len__())
    elif args.dataset == 'coco':
        if os.path.exists('/mnt/datasets/coco'):
            image_dir_path = '/mnt/datasets/coco/train2017'
            annotations_path = '/mnt/datasets/coco/annotations/captions_train2017.json'
        elif os.path.exists('/mnt/lustre'):
            image_dir_path = '/mnt/lustre/hein/cschlarmann37/datasets/coco/train2017'
            annotations_path = '/mnt/lustre/hein/cschlarmann37/datasets/coco/annotations/captions_train2017.json'
        else:
            raise ValueError('COCO dataset not found')
        dataset = COCOFlickrDataset(
            image_dir_path=image_dir_path,
            annotations_path=annotations_path,
            transform=preprocessor_without_normalize
        )
    dataset_eval = ImageNetDataset(
        root=args.imagenet_root + '/val',
        transform=preprocessor_without_normalize,
    )
    dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, drop_last=True)
    dataloader_eval = DataLoader(dataset_eval, batch_size=args.batch_size, shuffle=True, num_workers=8, drop_last=True)

    # Get text label embeddings of all ImageNet classes
    if args.template == 'std':
        template = 'This is a photo of a {}'
    elif args.template == 'blurry':
        template = 'This is a blurry photo of a {}'
    else:
        raise ValueError(f'Unknown template: {args.template}')
    print(f'template: {template}')
    texts = [template.format(c) for c in IMAGENET_1K_CLASS_ID_TO_LABEL.values()]
    text_tokens = open_clip.tokenize(texts)
    model_orig.to(main_device)
    with torch.no_grad():
        embedding_text_labels_norm = []
        for el in (text_tokens[:500], text_tokens[500:]):
            # we need to split the text tokens into two batches because otherwise we run out of memory
            # note that we are accessing the model directly here, not the CustomModel wrapper
            # thus its always normalizing the text embeddings
            embedding_text_labels_norm.append(
                model_orig.encode_text(el.to(main_device), normalize=True).detach().cpu()
            )
        embedding_text_labels_norm = torch.cat(embedding_text_labels_norm).T.to(main_device)
        assert torch.allclose(
            F.normalize(embedding_text_labels_norm, dim=0),
            embedding_text_labels_norm
        )
        if args.clip_model_name == 'ViT-B-32':
            assert embedding_text_labels_norm.shape == (512, 1000), embedding_text_labels_norm.shape
        elif args.clip_model_name in ('ViT-L-14', 'ViT-L-14-336'):
            assert embedding_text_labels_norm.shape == (768, 1000), embedding_text_labels_norm.shape
        else:
            raise ValueError(f'Unknown model: {args.clip_model_name}')

    model_orig.cpu()
    model_orig = ClipVisionModel(model=model_orig.visual, args=args, normalize=normalize)
    if num_gpus > 1:
        model_orig = torch.nn.DataParallel(model_orig)
    model_orig.cuda()

    model = ClipVisionModel(model=model.visual, args=args, normalize=normalize)
    if num_gpus > 1:
        model = torch.nn.DataParallel(model)
    model.cuda()

    # set optimizer (all params have requires_grad=True)
    params = unwrap_model(model).model.parameters()

    if args.opt == 'adamw':
        optimizer = torch.optim.AdamW(params, lr=args.lr, weight_decay=args.wd)
    elif args.opt == 'sgd':
        optimizer = torch.optim.SGD(
            params,
            lr=args.lr,
            momentum=args.momentum_sgd,
            weight_decay=args.wd
        )
    else:
        raise ValueError(f'Optimizer {args.optimizer} not supported.')
    if args.optimizer_state != '':
        optimizer.load_state_dict(torch.load(args.optimizer_state))

    # set scheduler
    scheduler = cosine_lr(optimizer, args.lr, args.warmup, args.steps)

    # compute amount of epochs
    total_epochs = args.steps / len(dataloader)
    print(f'train for {total_epochs} epochs')
    args.total_epochs = total_epochs

    # finetune
    step_total = args.start_step
    epoch = 0
    while step_total < args.steps:
        step_total = train_one_epoch(
            step_total,
            model=model,
            model_orig=model_orig,
            dataloader=dataloader,
            dataloader_eval=dataloader_eval,
            optimizer=optimizer,
            scheduler=scheduler,
            embedding_text_labels_norm=embedding_text_labels_norm,
            normalize=normalize,
            args=args,
            epoch=epoch
        )
        print(f'Epoch {epoch} done.')
        epoch += 1

    # save final model
    torch.save(unwrap_model(model).model.state_dict(), f'{args.output_dir}/checkpoints/final.pt')
    torch.save(optimizer.state_dict(), f'{args.output_dir}/checkpoints/final_opt.pt')

    if args.output_dir.endswith('_temp'):
        # rename temp dir to final dir
        os.rename(args.output_dir, args.output_dir[:-5])

class ClipVisionModel(torch.nn.Module):
    def __init__(self, model, args, normalize):
        super().__init__()
        self.model = model
        self.args = args
        self.normalize = normalize

    def forward(self, vision, output_normalize):
        vision = self.normalize(vision)
        embedding = self.model(vision)
        if output_normalize:
            embedding = F.normalize(embedding, dim=-1)
        return embedding


class ComputeLossWrapper:
    def __init__(self, embedding_orig, embedding_text_labels_norm, reduction='mean', loss=None,
                 logit_scale=100.):
        self.embedding_orig = embedding_orig
        self.embedding_text_labels_norm = embedding_text_labels_norm
        self.reduction = reduction
        self.loss_str = loss
        self.logit_scale = logit_scale

    def __call__(self, embedding, targets):
        return compute_loss(
            loss_str=self.loss_str, embedding=embedding, targets=targets,
            embedding_orig=self.embedding_orig, logit_scale=self.logit_scale,
            embedding_text_labels_norm=self.embedding_text_labels_norm, reduction=self.reduction
            )

def train_one_epoch(
        step_total, model, model_orig, dataloader, optimizer, scheduler, normalize,
        embedding_text_labels_norm, args, epoch, dataloader_eval=None
):
    model_orig.eval()
    model.train()

    loss_meter = AverageMeter('loss')
    cos_sim_meter = AverageMeter('cos-sim')
    acc_meter = AverageMeter('acc')
    racc_meter = AverageMeter('racc')

    epoch_start_time = time.time()
    for i, (data, targets) in enumerate(dataloader):
        is_classification = isinstance(targets, torch.Tensor)
        data = data.cuda()
        n_samples = data.shape[0]
        if is_classification:
            targets = targets.cuda()

        with torch.no_grad():
            embedding_orig = model_orig(vision=data, output_normalize=args.output_normalize)

        # loss for the attack
        loss_inner_wrapper = ComputeLossWrapper(
            embedding_orig, embedding_text_labels_norm,
            reduction='none' if args.attack == 'apgd' else 'mean', loss=args.inner_loss,
            logit_scale=100.
            )
        model.eval()

        if args.attack == 'pgd':
            data_adv = pgd(
                forward=model,
                loss_fn=loss_inner_wrapper,
                data_clean=data,
                targets=targets,
                norm=args.norm,
                eps=args.eps,
                iterations=args.iterations_adv,
                stepsize=args.stepsize_adv,
                output_normalize=args.output_normalize,
                perturbation=torch.zeros_like(data).uniform_(-args.eps, args.eps).requires_grad_(True),
                mode='max',
                verbose=False
            )
        elif args.attack == 'apgd':
            # apgd currently always applies output normalization
            data_adv = apgd(
                model=model,
                loss_fn=loss_inner_wrapper,
                x=data,
                y=targets,
                norm=args.norm,
                eps=args.eps,
                n_iter=args.iterations_adv,
                verbose=True
            )
        elif args.attack == 'none':
            data_adv = data

        del loss_inner_wrapper
        model.train()

        embedding_clean = model(data, output_normalize=args.output_normalize)
        if args.clean_weight > 0.:
            loss_clean = compute_loss(
                loss_str=args.loss_clean, embedding=embedding_clean, targets=targets,
                embedding_orig=embedding_orig, logit_scale=100., embedding_text_labels_norm=None
                )
        else:
            loss_clean = 0.

        embedding_adv = model(data_adv, output_normalize=args.output_normalize)
        del data, data_adv

        if args.trades:
            embedding_clean_no_grad = embedding_clean.detach().clone()
            embedding_orig.cpu()

        loss = compute_loss(
            loss_str=args.loss, embedding=embedding_adv, targets=targets,
            embedding_orig=embedding_orig if not args.trades else embedding_clean_no_grad,
            logit_scale=100., embedding_text_labels_norm=embedding_text_labels_norm
            )
        loss_total = args.clean_weight * loss_clean + (1 - args.clean_weight) * loss
        loss_total.backward()
        optimizer.step()
        optimizer.zero_grad()
        step_total += 1
        scheduler(step_total)

        with torch.no_grad():
            # only for logging
            embedding_orig.cuda()
            cos_sim_clean = F.cosine_similarity(embedding_clean, embedding_orig, dim=1).mean()
            cos_sim = F.cosine_similarity(embedding_adv, embedding_orig, dim=1).mean()
            if is_classification:
                logits_adv = embedding_adv @ embedding_text_labels_norm
                racc = compute_acc(logits_adv, targets)
                embedding_clean_norm = F.normalize(embedding_clean, dim=1)
                logits_clean = embedding_clean_norm @ embedding_text_labels_norm
                acc = compute_acc(logits_clean, targets)
                acc_meter.update(acc, n_samples)
                racc_meter.update(racc, n_samples)
                del embedding_clean_norm, embedding_clean
            else:
                acc = None
                racc = None

        loss_meter.update(loss.item(), n_samples)
        cos_sim_meter.update(cos_sim.item(), n_samples)

        eval_logs = dict()
        if (step_total-1) % args.eval_freq == 0:
            # we compute acc and racc (against supervised apgd) on validation data
            model.eval()
            data_eval, targets_eval = next(iter(dataloader_eval))
            data_eval, targets_eval = data_eval.cuda(), targets_eval.cuda()
            loss_eval_wrapper = ComputeLossWrapper(
                embedding_orig=None, embedding_text_labels_norm=embedding_text_labels_norm,
                reduction='none', loss='ce', logit_scale=100.
                )
            data_eval_adv = apgd(
                model=model,
                loss_fn=loss_eval_wrapper,
                x=data_eval,
                y=targets_eval,
                norm=args.norm,
                eps=args.eps,
                n_iter=50,
                initial_stepsize=0.05 * args.eps if args.clean_weight > 0 else None,
                verbose=False
            )
            with torch.no_grad():
                embedding_adv_eval_norm = model(data_eval_adv, output_normalize=True)  # we set output_normalize to True
                logits_eval_adv = embedding_adv_eval_norm @ embedding_text_labels_norm
                racc_eval = compute_acc(logits_eval_adv, targets_eval)
                embedding_eval_norm = model(data_eval, output_normalize=True)
                logits_eval = embedding_eval_norm @ embedding_text_labels_norm
                acc_eval = compute_acc(logits_eval, targets_eval)
                # note we compute the cosine sim between clean and adv embedding,
                # not between orig and adv embedding as for training
                cos_sim_eval = F.cosine_similarity(embedding_adv_eval_norm, embedding_eval_norm, dim=1).mean()
            eval_logs['eval/racc'] = racc_eval
            eval_logs['eval/acc'] = acc_eval
            eval_logs['eval/cos-sim'] = cos_sim_eval
            print(f'[eval-acc] {acc_eval:.2f} [eval-racc] {racc_eval:.2f} [eval-cos-sim] {cos_sim_eval:.3f}')
            model.train()
            del data_eval_adv, data_eval, targets_eval, embedding_adv_eval_norm, logits_eval_adv, embedding_eval_norm, logits_eval

        lr_ = optimizer.param_groups[0].get('lr')
        if (step_total-1) % args.log_freq == 0:
            log_str = f'[step] {step_total} [lr] {lr_:.6f} [loss] {loss.item():.6f} [cos-sim] {cos_sim.item():.3f}'
            if is_classification:
                log_str += f' [acc] {acc:.2f} [racc] {racc:.2f}'
            print(log_str)
            log_data = {
                'step': step_total,
                'lr': lr_,
                'loss': loss.item(),
                'loss-total': loss_total.item(),
                'cos-sim-clean': cos_sim_clean.item(),
                'cos-sim': cos_sim.item(),
                'acc': acc,
                'racc': racc,
                'avg/loss': loss_meter.avg,
                'avg/cos-sim': cos_sim_meter.avg,
                'avg/acc': acc_meter.avg,
                'avg/racc': racc_meter.avg,
            }
            log_data.update(eval_logs)
            if (step_total-1) % (args.log_freq * 10) == 0:
                # compute expected average epoch time in hours
                batch_average_time = (time.time() - epoch_start_time) / (i + 1) / (60**2)
                epoch_average_time = batch_average_time * len(dataloader)
                this_epoch_remaining = epoch_average_time - \
                                       (time.time() - epoch_start_time) / 60**2
                total_remaining = epoch_average_time * (args.total_epochs - epoch - i / len(dataloader))
                print(f'[epoch average time] {epoch_average_time:.2f} [this epoch remaining] '
                      f'{this_epoch_remaining:.2f} [total remaining] {total_remaining:.2f}')

                log_data.update({
                    'time/total-remaining': total_remaining,
                    'time/this-epoch-remaining': this_epoch_remaining,
                    'time/epoch-average-time': epoch_average_time,
                    'time/batch-average-time': batch_average_time,
                    'other/epoch': epoch + i / len(dataloader),
                })
            wandb.log(log_data)

        # save 10 models over the course of training
        if args.save_checkpoints and (step_total % (args.steps // 10) == 0):
            # save model and optimizer state_dict
            torch.save(unwrap_model(model).model.state_dict(), f'{args.output_dir}/checkpoints/step_{step_total}.pt')
            torch.save(optimizer.state_dict(), f'{args.output_dir}/checkpoints/step_{step_total}_opt.pt')
        # every 200 steps, save a fallback model, which gets overwritten
        if step_total % 200 == 0:
            torch.save(unwrap_model(model).model.state_dict(), f'{args.output_dir}/checkpoints/fallback_{step_total}.pt')
            torch.save(optimizer.state_dict(), f'{args.output_dir}/checkpoints/fallback_{step_total}_opt.pt')
            # remove old fallback models
            for file in os.listdir(f'{args.output_dir}/checkpoints'):
                if file.startswith('fallback') and not str(step_total) in file:
                    os.remove(f'{args.output_dir}/checkpoints/{file}')

        if step_total >= args.steps:
            break

        torch.cuda.empty_cache()
    return step_total


@torch.no_grad()
def compute_acc(logits, targets):
    preds_clean = logits.max(dim=1)[1].detach()
    acc = (preds_clean.eq(targets).sum() / targets.shape[0]).item() * 100
    return acc


def compute_loss(loss_str, embedding, targets, embedding_orig, logit_scale,
                 embedding_text_labels_norm=None, reduction='mean'):
    if loss_str == 'l2':
        loss = l2(out=embedding, targets=embedding_orig, reduction=reduction)
    elif loss_str == 'ce':
        loss = ce(
            out=embedding @ (logit_scale * embedding_text_labels_norm),
            targets=targets,
            reduction=reduction
        )
    else:
        raise ValueError(f'loss {loss_str} not supported')
    return loss

def l2(out, targets, reduction='none'):
    # squared l2 - it does not divide by the latent dimension
    # should have shape (batch_size, embedding_size)
    assert out.shape == targets.shape, f'{out.shape} != {targets.shape}'
    assert out.shape[0] > 1
    # Compute the element-wise squared error
    squared_error_batch = F.mse_loss(out, targets, reduction='none')
    if reduction == 'mean':
        squared_error_batch = torch.mean(squared_error_batch.sum(dim=1))
    else:
        squared_error_batch = squared_error_batch.sum(dim=1)
        assert squared_error_batch.shape == (out.shape[0],), f'{squared_error_batch.shape} != {(out.shape[0],)}'
    return squared_error_batch

def ce(out, targets, reduction='mean'):
    # out = logits
    assert out.shape[0] == targets.shape[0], (out.shape, targets.shape)
    assert out.shape[0] > 1

    return F.cross_entropy(out, targets, reduction=reduction)

if __name__ == '__main__':
    # set seeds
    torch.manual_seed(0)
    np.random.seed(0)

    # Parse command-line arguments
    args = parser.parse_args()
    args.eps /= 255
    args.stepsize_adv /= 255
    # make sure there is no string in args that should be a bool
    assert not any([isinstance(x, str) and x in ['True', 'False'] for x in args.__dict__.values()]), f'args contains a string that should be a bool: {args}'
    assert args.eval_freq % args.log_freq == 0, 'eval_freq must be a multiple of log_freq'

    if args.devices != '':
        # set cuda visible devices
        os.environ['CUDA_VISIBLE_DEVICES'] = args.devices

    num_gpus = torch.cuda.device_count()
    if num_gpus > 1:
        print(f'Number of GPUs available: {num_gpus}')
    else:
        print('No multiple GPUs available.')

    # set model name and output dir
    random_str = ''.join(random.choices(string.ascii_letters + string.digits, k=5))
    args.finetuned_model_name = f'{args.clip_model_name}_{args.pretrained}_{args.dataset}_{args.loss}_{args.dataset}_{args.experiment_name}_{random_str}'
    args.finetuned_model_name = args.finetuned_model_name.replace('/', '_')
    args.output_dir = os.path.join(args.output_dir, args.finetuned_model_name)
    # run
    main(args)