-
Notifications
You must be signed in to change notification settings - Fork 26
/
utils.py
253 lines (202 loc) · 8.89 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
import numpy as np
from collections import namedtuple
import torch
from torch import nn
import torchvision
from torch.optim.optimizer import Optimizer, required
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
################################################################
## Components from https://github.com/davidcpage/cifar10-fast ##
################################################################
#####################
## data preprocessing
#####################
cifar10_mean = (0.4914, 0.4822, 0.4465) # equals np.mean(train_set.train_data, axis=(0,1,2))/255
cifar10_std = (0.2471, 0.2435, 0.2616) # equals np.std(train_set.train_data, axis=(0,1,2))/255
def pad(x, border=4):
return np.pad(x, [(0, 0), (border, border), (border, border), (0, 0)], mode='reflect')
def transpose(x, source='NHWC', target='NCHW'):
return x.transpose([source.index(d) for d in target])
#####################
## data augmentation
#####################
class Crop(namedtuple('Crop', ('h', 'w'))):
def __call__(self, x, x0, y0):
return x[:,y0:y0+self.h,x0:x0+self.w]
def options(self, x_shape):
C, H, W = x_shape
return {'x0': range(W+1-self.w), 'y0': range(H+1-self.h)}
def output_shape(self, x_shape):
C, H, W = x_shape
return (C, self.h, self.w)
class FlipLR(namedtuple('FlipLR', ())):
def __call__(self, x, choice):
return x[:, :, ::-1].copy() if choice else x
def options(self, x_shape):
return {'choice': [True, False]}
class Cutout(namedtuple('Cutout', ('h', 'w'))):
def __call__(self, x, x0, y0):
x = x.copy()
x[:,y0:y0+self.h,x0:x0+self.w].fill(0.0)
return x
def options(self, x_shape):
C, H, W = x_shape
return {'x0': range(W+1-self.w), 'y0': range(H+1-self.h)}
class Transform():
def __init__(self, dataset, transforms):
self.dataset, self.transforms = dataset, transforms
self.choices = None
def __len__(self):
return len(self.dataset)
def __getitem__(self, index):
data, labels = self.dataset[index]
for choices, f in zip(self.choices, self.transforms):
args = {k: v[index] for (k,v) in choices.items()}
data = f(data, **args)
return data, labels
def set_random_choices(self):
self.choices = []
x_shape = self.dataset[0][0].shape
N = len(self)
for t in self.transforms:
options = t.options(x_shape)
x_shape = t.output_shape(x_shape) if hasattr(t, 'output_shape') else x_shape
self.choices.append({k:np.random.choice(v, size=N) for (k,v) in options.items()})
#####################
## dataset
#####################
def cifar10(root):
train_set = torchvision.datasets.CIFAR10(root=root, train=True, download=True)
test_set = torchvision.datasets.CIFAR10(root=root, train=False, download=True)
return {
'train': {'data': train_set.data, 'labels': train_set.targets},
'test': {'data': test_set.data, 'labels': test_set.targets}
}
#####################
## data loading
#####################
class Batches():
def __init__(self, dataset, batch_size, shuffle, set_random_choices=False, num_workers=0, drop_last=False):
self.dataset = dataset
self.batch_size = batch_size
self.set_random_choices = set_random_choices
self.dataloader = torch.utils.data.DataLoader(
dataset, batch_size=batch_size, num_workers=num_workers, pin_memory=True, shuffle=shuffle, drop_last=drop_last
)
def __iter__(self):
if self.set_random_choices:
self.dataset.set_random_choices()
return ({'input': x.to(device).half(), 'target': y.to(device).long()} for (x,y) in self.dataloader)
def __len__(self):
return len(self.dataloader)
#####################
## new optimizer
#####################
class SGD_GCC(Optimizer):
def __init__(self, params, lr=required, momentum=0, dampening=0,
weight_decay=0, nesterov=False):
if lr is not required and lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if momentum < 0.0:
raise ValueError("Invalid momentum value: {}".format(momentum))
if weight_decay < 0.0:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, dampening=dampening,
weight_decay=weight_decay, nesterov=nesterov)
if nesterov and (momentum <= 0 or dampening != 0):
raise ValueError("Nesterov momentum requires a momentum and zero dampening")
super(SGD_GCC, self).__init__(params, defaults)
def __setstate__(self, state):
super(SGD_GCC, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('nesterov', False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
for p in group['params']:
if p.grad is None:
continue
d_p = p.grad.data
if weight_decay != 0:
d_p.add_(weight_decay, p.data)
#GC operation for Conv layers
if len(list(d_p.size()))>3:
d_p.add_(-d_p.mean(dim = tuple(range(1,len(list(d_p.size())))), keepdim = True))
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(1 - dampening, d_p)
if nesterov:
d_p = d_p.add(momentum, buf)
else:
d_p = buf
p.data.add_(-group['lr'], d_p)
return loss
class SGD_GC(Optimizer):
def __init__(self, params, lr=required, momentum=0, dampening=0,
weight_decay=0, nesterov=False):
if lr is not required and lr < 0.0:
raise ValueError("Invalid learning rate: {}".format(lr))
if momentum < 0.0:
raise ValueError("Invalid momentum value: {}".format(momentum))
if weight_decay < 0.0:
raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
defaults = dict(lr=lr, momentum=momentum, dampening=dampening,
weight_decay=weight_decay, nesterov=nesterov)
if nesterov and (momentum <= 0 or dampening != 0):
raise ValueError("Nesterov momentum requires a momentum and zero dampening")
super(SGD_GC, self).__init__(params, defaults)
def __setstate__(self, state):
super(SGD_GC, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('nesterov', False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
weight_decay = group['weight_decay']
momentum = group['momentum']
dampening = group['dampening']
nesterov = group['nesterov']
for p in group['params']:
if p.grad is None:
continue
d_p = p.grad.data
if weight_decay != 0:
d_p.add_(weight_decay, p.data)
#GC operation for Conv layers and FC layers
if len(list(d_p.size()))>1:
d_p.add_(-d_p.mean(dim = tuple(range(1,len(list(d_p.size())))), keepdim = True))
if momentum != 0:
param_state = self.state[p]
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
else:
buf = param_state['momentum_buffer']
buf.mul_(momentum).add_(1 - dampening, d_p)
if nesterov:
d_p = d_p.add(momentum, buf)
else:
d_p = buf
p.data.add_(-group['lr'], d_p)
return loss