/
utils.py
435 lines (326 loc) · 16.2 KB
/
utils.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
import os
import shutil
import gc
import torch
from torch import nn
from torch.nn import functional as F
import yaml
from tqdm import tqdm
import torch.distributed as dist
from torch.utils.data.distributed import DistributedSampler
from torch.utils.data import DataLoader
import numpy as np
import random
from math import floor, ceil
import pickle
import contextlib
@contextlib.contextmanager
def _disable_tracking_bn_stats(model):
def switch_attr(m):
if hasattr(m, 'track_running_stats'):
m.track_running_stats ^= True
model.apply(switch_attr)
yield
model.apply(switch_attr)
class VATLoss(nn.Module):
def __init__(self, xi=0.1, eps=0.1, ip=1):
"""VAT loss
:param xi: hyperparameter of VAT (default: 10.0)
:param eps: size of adversarial perturbation
:param ip: iterations of projected gradient descent: 1)
"""
super(VATLoss, self).__init__()
self.xi = xi
self.eps = eps
self.ip = ip
def forward(self, model, x):
with torch.no_grad():
pred = F.softmax(model(x), dim=1)
# prepare random unit tensor
d = torch.nn.functional.normalize(torch.rand(x.shape)).to(x.device)
with _disable_tracking_bn_stats(model):
# calc adversarial direction
for _ in range(self.ip):
d.requires_grad_()
pred_hat = model(x + self.xi * d)
logp_hat = F.log_softmax(pred_hat, dim=1)
adv_distance = F.kl_div(logp_hat, pred, reduction='batchmean')
adv_distance.backward()
d = torch.nn.functional.normalize(d.grad.detach().clone())
model.zero_grad()
# calc LDS
r_adv = d * self.eps
pred_hat = model(x + r_adv)
logp_hat = F.log_softmax(pred_hat, dim=1)
lds = F.kl_div(logp_hat, pred, reduction='batchmean')
return lds
def get_embeddings(model, loader, val_loader):
global_embeds = []
global_labels = []
gc.collect()
with torch.no_grad():
for i, (images, labels) in enumerate(tqdm(loader)):
local_embeds = model(images)
local_embeds = torch.nan_to_num(local_embeds)
labels = labels.to(int(os.environ["RANK"]) % torch.cuda.device_count())
world_embeds = [local_embeds for i in range(int(os.environ["WORLD_SIZE"]))]
world_labels = [labels for i in range(int(os.environ["WORLD_SIZE"]))]
dist.all_gather(world_embeds, local_embeds)
dist.all_gather(world_labels, labels)
world_embeds = [w.cpu() for w in world_embeds]
world_labels = [l.cpu() for l in world_labels]
global_embeds += world_embeds
global_labels += world_labels
global_embeds_val = []
global_labels_val = []
gc.collect()
with torch.no_grad():
for images, labels in tqdm(val_loader):
local_embeds = model(images)
labels = labels.to(int(os.environ["RANK"]) % torch.cuda.device_count())
world_embeds = [local_embeds for i in range(int(os.environ["WORLD_SIZE"]))]
world_labels = [labels for i in range(int(os.environ["WORLD_SIZE"]))]
dist.all_gather(world_embeds, local_embeds)
dist.all_gather(world_labels, labels)
world_embeds = [w.cpu() for w in world_embeds]
world_labels = [l.cpu() for l in world_labels]
global_embeds_val += world_embeds
global_labels_val += world_labels
return (torch.concat(global_embeds, 0), torch.concat(global_labels, 0), torch.concat(global_embeds_val, 0), torch.concat(global_labels_val, 0))
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
torch.save(state, filename)
if is_best:
shutil.copyfile(filename, 'model_best.pth.tar')
def save_config_file(model_checkpoints_folder, args):
if not os.path.exists(model_checkpoints_folder):
os.makedirs(model_checkpoints_folder)
with open(os.path.join(model_checkpoints_folder, 'config.yml'), 'w') as outfile:
yaml.dump(args, outfile, default_flow_style=False)
def accuracy(output, target, topk=(1,)):
output = output.to(torch.device('cpu'))
target = target.to(torch.device('cpu'))
maxk = max(topk)
batch_size = target.shape[0]
_, idx = output.sort(dim=1, descending=True)
pred = idx.narrow(1, 0, maxk).t()
correct = pred.eq(target.reshape(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].reshape(-1).float().sum(dim=0, keepdim=True)
res.append(correct_k.mul_(100.0 / batch_size))
return res
class RepeatLoader:
def __init__(self, loader):
self.loader = loader
self.it = iter(loader)
def __iter__(self):
return self
def __next__(self):
try:
return next(self.it)
except Exception as e:
self.it = iter(self.loader)
return next(self.it)
def create_imagefolder(data, samples, path, transform, new_path=None):
imgfolder = datasets.ImageFolder(path, transform=transform)
imgfolder.class_to_idx = data['class_map']
imgfolder.classes = list(data['class_map'].keys())
imgfolder.samples = samples
if new_path is not None:
imgfolder.root = new_path
return imgfolder
def create_loader(data, cuda_kwargs, shuffle=False):
loader_kwargs = {'batch_size': BATCH_SIZE, 'shuffle': shuffle}
loader_kwargs.update(cuda_kwargs)
loader = DataLoader(data, **loader_kwargs)
return loader
def epoch_accuracy(loader_s, loader_t, student, teacher):
student.eval()
teacher.eval()
out_epoch_s = [accuracy(student(L.to(0)), y)[0].cpu().item() for L, y in loader_s]
out_epoch_t = [accuracy(teacher(L.to(0)), y)[0].cpu().item() for L, y in loader_t]
student.train()
teacher.train()
return sum(out_epoch_s) / len(out_epoch_s), sum(out_epoch_t) / len(out_epoch_t)
def add_to_imagefolder(paths, labels, dataset):
"""
Adds the paths with the labels to an image classification dataset
:list paths: a list of absolute image paths to add to the dataset
:list labels: a list of labels for each path
:Dataset dataset: the dataset to add the samples to
"""
new_samples = list(zip(paths, labels))
dataset.samples += new_samples
return dataset.samples
# splits the datasets of the two views so that
# the instances inside are still aligned by index
def train_test_split_samples(samples0, samples1, test_size, random_state=None):
if random_state is not None:
random.seed(random_state)
assert test_size > 0 and test_size < 1, \
'test_size should be a float between (0, 1)'
assert len(samples0[0]) == len(samples1[0]), \
f'number of samples in samples0 ({len(samples0)}), samples1 {len(samples1)} are not equal'
idx_samples = list(range(len(samples0[0])))
idx_test = random.sample(idx_samples, floor(test_size * len(samples0[0])))
idx_train = list(set(idx_samples) - set(idx_test))
# convert to np array for convenient array indexing
#samples0_np = np.stack([np.array(a) for a in samples0])
#samples1_np = np.stack([np.array(a) for a in samples1])
samples_train0 = samples0[0][idx_train], samples0[1][idx_train]
samples_test0 = samples0[0][idx_test], samples0[1][idx_test]
samples_train1 = samples1[0][idx_train], samples1[1][idx_train]
samples_test1 = samples1[0][idx_test], samples1[1][idx_test]
assert len(samples_train0[0]) == len(samples_train1[0]), 'sample sizes not equal after split'
assert len(samples_test0[0]) == len(samples_test1[0]), 'sample sizes not equal after split'
return samples_train0, samples_test0, samples_train1, samples_test1
class EarlyStopper:
def __init__(self, stopping_metric, patience):
self.stopping_metric = stopping_metric
self.patience = patience
self.epochs_since_improvement = 0
self.stop = False
self.best_val_loss = float("inf")
self.best_val_acc = 0
# TODO perhaps there is a more elegant way to write this
def is_new_best_metric(self, val_acc, val_loss):
self.epochs_since_improvement += 1
if self.stopping_metric == 'loss' and val_loss < self.best_val_loss - 1e-4:
self.best_val_loss = val_loss
self.best_val_acc = val_acc
self.epochs_since_improvement = 0
return True
elif self.stopping_metric == 'accuracy' and val_acc > self.best_val_acc + 1e-4:
self.best_val_loss = val_loss
self.best_val_acc = val_acc
self.epochs_since_improvement = 0
return True
return False
def early_stop(self):
if self.epochs_since_improvement > self.patience:
return True
return False
def get_dataset_tensors(from_embed, dataset='IN1K'):
"""
from_embed - the model from which the embeddings were derived
dataset - the name of the dataset that was embedded
"""
with open(f'./{from_embed}_{dataset}_train.ds', 'rb') as fp:
x_train_embeds, x_train_labels = pickle.load(fp)
with open(f'./{from_embed}_{dataset}_val.ds', 'rb') as fp:
x_val_embeds, x_val_labels = pickle.load(fp)
x_val_embeds = torch.tensor(x_val_embeds)
x_val_labels = torch.tensor(x_val_labels)
x_train_embeds = torch.tensor(x_train_embeds)
x_train_labels = torch.tensor(x_train_labels)
return (x_train_embeds, x_train_labels, x_val_embeds, x_val_labels), np.unique(x_train_labels).shape[0]
def cascade_round(arr):
s = 0.0
arr_cp = np.zeros_like(arr)
for i, a in enumerate(arr):
s += a
if s - (s // 1) > .5:
arr_cp[i] = ceil(a)
else:
arr_cp[i] = floor(a)
return arr_cp.astype(np.int32)
def subset_npercent(tensors, dataset='IN1K', percent=1, balanced=False):
percent /= 100.0
print(percent)
np.random.seed(13)
x_train_embeds, x_train_labels, x_val_embeds, x_val_labels = tensors
if dataset == 'iNat2017':
unique0, counts0 = np.unique(x_train_labels, return_counts=True)
keep = unique0[-1010:]
train_keep = np.hstack([np.where(x_train_labels == l)[0] for l in keep])
val_keep = np.hstack([np.where(x_val_labels == l)[0] for l in keep])
x_train_embeds = x_train_embeds[train_keep]
x_train_labels = x_train_labels[train_keep]
x_val_embeds = x_val_embeds[val_keep]
x_val_labels = x_val_labels[val_keep]
if percent == 0.01 and dataset == 'IN1K':
with open('./1percent_idx.pkl', 'rb') as fp:
idx = pickle.load(fp)
elif percent == 0.1 and dataset == 'IN1K':
with open('./10percent_idx.pkl', 'rb') as fp:
idx = pickle.load(fp)
elif balanced: # create a custom balanced split of percent
# for count_per_class
unique, counts = np.unique(x_train_labels, return_counts=True)
counts_per_class = np.maximum(np.minimum((np.ones_like(counts) * ((x_train_labels.shape[0] / counts.shape[0]) * percent)) - 1.0, counts), 1.0)
counts_per_class_rounded = np.minimum(cascade_round(counts_per_class), counts)
while sum(counts_per_class_rounded) < (x_train_labels.shape[0] * percent): # add a single example in the cascade round
counts_per_class = np.minimum(counts_per_class + (1 / counts.shape[0]), counts)
counts_per_class_rounded = np.minimum(cascade_round(counts_per_class), counts)
mask = np.hstack([np.random.choice(np.where(x_train_labels == unique[l])[0], int(counts_per_class_rounded[l]), replace=False)
for l in range(unique.shape[0])])
idx = mask
else: # create a custom stratified split
unique, counts = np.unique(x_train_labels, return_counts=True)
count_per_class = percent * counts
# ok, but this is not exactly n% we will have some rounding to do here
count_per_class = cascade_round(count_per_class)
mask = np.hstack([np.random.choice(np.where(x_train_labels == unique[l])[0], count_per_class[l], replace=False)
for l in range(unique.shape[0])])
idx = mask
unlb_idx = list(set(list(range(x_train_labels.shape[0]))) - set(idx))
unique0, counts0 = np.unique(x_train_labels, return_counts=True)
x_val_embeds = torch.tensor(x_val_embeds)
x_val_labels = torch.tensor(x_val_labels)
x_unlbl_embeds = torch.tensor(x_train_embeds[unlb_idx])
x_unlbl_labels = torch.tensor(x_train_labels[unlb_idx])
x_train_embeds = torch.tensor(x_train_embeds[idx])
x_train_labels = torch.tensor(x_train_labels[idx])
unique, counts = np.unique(x_train_labels, return_counts=True)
print(counts)
print(sum(counts), sum(counts0))
print(sum(counts) / sum(counts0))
return x_train_embeds, x_train_labels, x_unlbl_embeds, x_unlbl_labels, x_val_embeds, x_val_labels
def make_dataset(from_embed, dataset='IN1K', percent=1, balanced=False):
tensors, num_classes = get_dataset_tensors(from_embed, dataset=dataset)
if percent < 100:
x_train_embeds, x_train_labels, x_unlbl_embeds, x_unlbl_labels, x_val_embeds, x_val_labels = subset_npercent(tensors, dataset=dataset, percent=percent, balanced=balanced)
train_dataset = torch.utils.data.TensorDataset(x_train_embeds, x_train_labels)
unlbl_dataset = torch.utils.data.TensorDataset(x_unlbl_embeds, x_unlbl_labels)
val_dataset = torch.utils.data.TensorDataset(x_val_embeds, x_val_labels)
else:
x_train_embeds, x_train_labels, x_val_embeds, x_val_labels = tensors
train_dataset = torch.utils.data.TensorDataset(x_train_embeds, x_train_labels)
unlbl_dataset = None
val_dataset = torch.utils.data.TensorDataset(x_val_embeds, x_val_labels)
return train_dataset, unlbl_dataset, val_dataset, num_classes
def make_concat_dataset(from_embed, to_embed, dataset='IN1K', percent=1, balanced=False):
tensors0, num_classes = get_dataset_tensors(from_embed, dataset=dataset)
tensors1, num_classes = get_dataset_tensors(to_embed, dataset=dataset)
if percent < 100:
x_train_embeds, x_train_labels, x_unlbl_embeds, x_unlbl_labels, x_val_embeds, x_val_labels = subset_npercent(tensors0, dataset=dataset, percent=percent, balanced=balanced)
y_train_embeds, y_train_labels, y_unlbl_embeds, y_unlbl_labels, y_val_embeds, y_val_labels = subset_npercent(tensors1, dataset=dataset, percent=percent, balanced=balanced)
train_dataset = torch.utils.data.TensorDataset(torch.cat((x_train_embeds, y_train_embeds), -1), x_train_labels)
unlbl_dataset = torch.utils.data.TensorDataset(torch.cat((x_unlbl_embeds, y_unlbl_embeds), -1), x_unlbl_labels)
val_dataset = torch.utils.data.TensorDataset(torch.cat((x_val_embeds, y_val_embeds), -1), x_val_labels)
else:
x_train_embeds, x_train_labels, x_val_embeds, x_val_labels = tensors0
y_train_embeds, y_train_labels, y_val_embeds, y_val_labels = tensors1
train_dataset = torch.utils.data.TensorDataset(torch.cat((x_train_embeds, y_train_embeds), -1), x_train_labels)
unlbl_dataset = None
val_dataset = torch.utils.data.TensorDataset(torch.cat((x_val_embeds, y_val_embeds), -1), x_val_labels)
return train_dataset, unlbl_dataset, val_dataset, num_classes
def step_perf(loader_train0, loader_train1, loader_val0, loader_val1, model0, model1, s):
val_acc0, val_acc1 = epoch_accuracy(loader_val0, loader_val1, model0, model1)
#val_acc0x, val_acc1x = epoch_accuracy(loader_val1, loader_val0, model0, model1)
train_acc0, train_acc1 = epoch_accuracy(loader_train0, loader_train1, model0, model1)
s += 1
return {'val_acc0': val_acc0,
'val_acc1': val_acc1,
'train_acc0': train_acc0,
'train_acc1': train_acc1}, s
def create_sampler_loader(args, rank, world_size, data, cuda_kwargs={'num_workers': 12, 'pin_memory': True, 'shuffle': False}, shuffle=True):
sampler = DistributedSampler(data, rank=rank, num_replicas=world_size, shuffle=shuffle)
loader_kwargs = {'batch_size': args.batch_size, 'sampler': sampler}
loader_kwargs.update(cuda_kwargs)
loader = DataLoader(data, **loader_kwargs)
return sampler, loader
def setup(rank, world_size):
dist.init_process_group("nccl", rank=rank, world_size=world_size)
def cleanup():
dist.destroy_process_group()