-
Notifications
You must be signed in to change notification settings - Fork 27
/
utils.py
executable file
·282 lines (224 loc) · 8.88 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
import os
import re
import torch
import torch.nn as nn
from collections import OrderedDict
import numpy as np
from torch.nn.modules.loss import _Loss
import PIL
from PIL import Image
from torchvision import transforms, datasets
import torch.nn.functional as F
import cv2
from torch.utils.data import Sampler
import random
import math
from SGD import SGD
class AvgrageMeter(object):
def __init__(self):
self.reset()
def reset(self):
self.avg = 0
self.sum = 0
self.cnt = 0
self.val = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.cnt += n
self.avg = self.sum / self.cnt
def accuracy(output, target, topk=(1,)):
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].view(-1).float().sum(0)
res.append(correct_k.mul_(100.0/batch_size))
return res
def save_checkpoint(model, iters, path, optimizer=None, scheduler=None):
if not os.path.exists(path):
os.makedirs(path)
print("Saving checkpoint to file {}".format(path))
state_dict = {}
new_state_dict = OrderedDict()
for k, v in model.state_dict().items():
key = k
if k.split('.')[0] == 'module':
key = k[7:]
new_state_dict[key] = v
state_dict['model'] = new_state_dict
state_dict['iteration'] = iters
if optimizer is not None:
state_dict['optimizer'] = optimizer.state_dict()
if scheduler is not None:
state_dict['scheduler'] = scheduler.state_dict()
filename = os.path.join("{}/checkpoint.pth".format(path))
try:
torch.save(state_dict, filename)
except:
print('save {} failed, continue training'.format(path))
def sgd_optimizer(model, base_lr, momentum, weight_decay):
params = []
for key, value in model.named_parameters():
params.append(value)
param_group = [{'params': params,
'weight_decay': weight_decay}]
optimizer = SGD(param_group, lr = base_lr, momentum=momentum)
return optimizer
## data augmentation functions
class OpencvResize(object):
def __init__(self, size=256):
self.size = size
def __call__(self, img):
assert isinstance(img, PIL.Image.Image)
img = np.asarray(img) # (H,W,3) RGB
img = img[:,:,::-1] # 2 BGR
img = np.ascontiguousarray(img)
H, W, _ = img.shape
target_size = (int(self.size/H * W + 0.5), self.size) if H < W else (self.size, int(self.size/W * H + 0.5))
img = cv2.resize(img, target_size, interpolation=cv2.INTER_LINEAR)
img = img[:,:,::-1] # 2 RGB
img = np.ascontiguousarray(img)
img = Image.fromarray(img)
return img
class ToBGRTensor(object):
def __call__(self, img):
assert isinstance(img, (np.ndarray, PIL.Image.Image))
if isinstance(img, PIL.Image.Image):
img = np.asarray(img)
img = img[:,:,::-1] # 2 BGR
img = np.transpose(img, [2, 0, 1]) # 2 (3, H, W)
img = np.ascontiguousarray(img)
img = torch.from_numpy(img).float()
return img
class RandomResizedCrop(object):
def __init__(self, scale=(0.08, 1.0), target_size:int=224, max_attempts:int=10):
assert scale[0] <= scale[1]
self.scale = scale
assert target_size > 0
self.target_size = target_size
assert max_attempts >0
self.max_attempts = max_attempts
def __call__(self, img):
assert isinstance(img, PIL.Image.Image)
img = np.asarray(img, dtype=np.uint8)
H, W, C = img.shape
well_cropped = False
for _ in range(self.max_attempts):
crop_area = (H*W) * random.uniform(self.scale[0], self.scale[1])
crop_edge = round(math.sqrt(crop_area))
dH = H - crop_edge
dW = W - crop_edge
crop_left = random.randint(min(dW, 0), max(dW, 0))
crop_top = random.randint(min(dH, 0), max(dH, 0))
if dH >= 0 and dW >= 0:
well_cropped = True
break
crop_bottom = crop_top + crop_edge
crop_right = crop_left + crop_edge
if well_cropped:
crop_image = img[crop_top:crop_bottom,:,:][:,crop_left:crop_right,:]
else:
roi_top = max(crop_top, 0)
padding_top = roi_top - crop_top
roi_bottom = min(crop_bottom, H)
padding_bottom = crop_bottom - roi_bottom
roi_left = max(crop_left, 0)
padding_left = roi_left - crop_left
roi_right = min(crop_right, W)
padding_right = crop_right - roi_right
roi_image = img[roi_top:roi_bottom,:,:][:,roi_left:roi_right,:]
crop_image = cv2.copyMakeBorder(roi_image, padding_top, padding_bottom, padding_left, padding_right,
borderType=cv2.BORDER_CONSTANT, value=0)
target_image = cv2.resize(crop_image, (self.target_size, self.target_size), interpolation=cv2.INTER_LINEAR)
target_image = PIL.Image.fromarray(target_image.astype('uint8'))
return target_image
class LighteningJitter(object):
def __init__(self, eigen_vecs, eigen_values, max_eigen_jitter=0.1):
self.eigen_vecs = np.array(eigen_vecs, dtype=np.float32)
self.eigen_values = np.array(eigen_values, dtype=np.float32)
self.max_eigen_jitter = max_eigen_jitter
def __call__(self, img):
assert isinstance(img, PIL.Image.Image)
img = np.asarray(img, dtype=np.float32)
img = np.ascontiguousarray(img/255)
cur_eigen_jitter = np.random.normal(scale=self.max_eigen_jitter, size=self.eigen_values.shape)
color_purb = (self.eigen_vecs @ (self.eigen_values * cur_eigen_jitter)).reshape([1, 1, -1])
img += color_purb
img = np.ascontiguousarray(img*255)
img.clip(0, 255, out=img)
img = PIL.Image.fromarray(np.uint8(img))
return img
class Random_Batch_Sampler(Sampler):
def __init__(self, dataset, batch_size, total_iters, rank=None):
if rank is None:
if not dist.is_available():
raise RuntimeError("Requires distributed package to be available")
self.dataset_num = dataset.__len__()
self.rank = rank
self.batch_size = batch_size
self.total_iters = total_iters
def __iter__(self):
random.seed(self.rank)
for i in range(self.total_iters):
batch_iter = []
for _ in range(self.batch_size):
batch_iter.append(random.randint(0, self.dataset_num-1))
yield batch_iter
def __len__(self):
return self.total_iters
class LabelSmoothCrossEntropyLoss(_Loss):
def __init__(self, eps=0.1, class_num=1000):
super(LabelSmoothCrossEntropyLoss, self).__init__()
self.min_value = eps / class_num
self.eps = eps
def __call__(self, pred:torch.Tensor, target:torch.Tensor):
epses = self.min_value * torch.ones_like(pred)
log_probs = F.log_softmax(pred, dim=1)
if target.ndimension() == 1:
target = target.expand(1, *target.shape)
target = target.transpose(1, 0)
target = torch.zeros_like(log_probs).scatter_(1, target, 1)
target = target.type(torch.float)
target = target * (1 - self.eps) + epses
element_wise_mul = log_probs * target * -1.0
loss = torch.sum(element_wise_mul, 1)
loss = torch.mean(loss)
return loss
def get_train_dataloader(train_dir, batch_size, total_epochs, local_rank):
eigvec = np.array([
[-0.5836, -0.6948, 0.4203],
[-0.5808, -0.0045, -0.8140],
[-0.5675, 0.7192, 0.4009]
])
eigval = np.array([0.2175, 0.0188, 0.0045])
train_dataset = datasets.ImageFolder(train_dir,
transforms.Compose([
RandomResizedCrop(target_size=224, scale=(0.08, 1.0)),
LighteningJitter(eigen_vecs=eigvec[::-1,:], eigen_values=eigval, max_eigen_jitter=0.1),
transforms.RandomHorizontalFlip(0.5),
ToBGRTensor(),
])
)
datasampler = Random_Batch_Sampler(
train_dataset, batch_size=batch_size,
total_iters=total_epochs*5000, rank=local_rank)
train_loader = torch.utils.data.DataLoader(
train_dataset, num_workers=8,
pin_memory=True, batch_sampler=datasampler)
return train_loader
def get_val_dataloader(val_dir):
val_dataset = datasets.ImageFolder(val_dir,
transforms.Compose([
OpencvResize(256),
transforms.CenterCrop(224),
ToBGRTensor(),
]))
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=200, shuffle=False,
num_workers=8, pin_memory=True
)
return val_loader