-
Notifications
You must be signed in to change notification settings - Fork 10
/
cifar10.py
226 lines (174 loc) · 6.47 KB
/
cifar10.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
#!/usr/bin/env python
"""
cifar10.py
"""
from __future__ import division, print_function
import sys
import json
import argparse
import numpy as np
from time import time
from PIL import Image
from basenet import BaseNet
from basenet.hp_schedule import HPSchedule
from basenet.helpers import to_numpy, set_seeds
from basenet.vision import transforms as btransforms
import torch
from torch import nn
from torch.nn import functional as F
torch.backends.cudnn.benchmark = True
from torchvision import transforms, datasets
# --
# CLI
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', type=int, default=30)
parser.add_argument('--extra', type=int, default=0)
parser.add_argument('--burnout', type=int, default=0)
parser.add_argument('--lr-schedule', type=str, default='one_cycle')
parser.add_argument('--lr-max', type=float, default=0.1)
parser.add_argument('--weight-decay', type=float, default=5e-4)
parser.add_argument('--momentum', type=float, default=0.9)
parser.add_argument('--batch-size', type=int, default=128)
parser.add_argument('--sgdr-period-length', type=int, default=10)
parser.add_argument('--sgdr-t-mult', type=int, default=2)
parser.add_argument('--seed', type=int, default=123)
parser.add_argument('--download', action="store_true")
return parser.parse_args()
args = parse_args()
set_seeds(args.seed)
# --
# IO
print('cifar10.py: making dataloaders...', file=sys.stderr)
transform_train = transforms.Compose([
btransforms.ReflectionPadding(margin=(4, 4)),
transforms.RandomCrop(32),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
btransforms.NormalizeDataset(dataset='cifar10'),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
btransforms.NormalizeDataset(dataset='cifar10'),
])
try:
trainset = datasets.CIFAR10(root='./data', train=True, download=args.download, transform=transform_train)
testset = datasets.CIFAR10(root='./data', train=False, download=args.download, transform=transform_test)
except:
raise Exception('cifar10.py: error loading data -- try rerunning w/ `--download` flag')
trainloader = torch.utils.data.DataLoader(
trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4,
pin_memory=True,
)
testloader = torch.utils.data.DataLoader(
testset,
batch_size=512,
shuffle=False,
num_workers=4,
pin_memory=True,
)
dataloaders = {
"train" : trainloader,
"test" : testloader,
}
# --
# Model definition
# Derived from models in `https://github.com/kuangliu/pytorch-cifar`
class PreActBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1):
super().__init__()
self.bn1 = nn.BatchNorm2d(in_channels)
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False)
)
def forward(self, x):
out = F.relu(self.bn1(x))
shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x
out = self.conv1(out)
out = self.conv2(F.relu(self.bn2(out)))
return out + shortcut
class ResNet18(BaseNet):
def __init__(self, num_blocks=[2, 2, 2, 2], num_classes=10):
super().__init__(loss_fn=F.cross_entropy)
self.in_channels = 64
self.prep = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(64),
nn.ReLU()
)
self.layers = nn.Sequential(
self._make_layer(64, 64, num_blocks[0], stride=1),
self._make_layer(64, 128, num_blocks[1], stride=2),
self._make_layer(128, 256, num_blocks[2], stride=2),
self._make_layer(256, 256, num_blocks[3], stride=2),
)
self.classifier = nn.Linear(512, num_classes)
def _make_layer(self, in_channels, out_channels, num_blocks, stride):
strides = [stride] + [1] * (num_blocks-1)
layers = []
for stride in strides:
layers.append(PreActBlock(in_channels=in_channels, out_channels=out_channels, stride=stride))
in_channels = out_channels
return nn.Sequential(*layers)
def forward(self, x):
x = x.half()
x = self.prep(x)
x = self.layers(x)
x_avg = F.adaptive_avg_pool2d(x, (1, 1))
x_avg = x_avg.view(x_avg.size(0), -1)
x_max = F.adaptive_max_pool2d(x, (1, 1))
x_max = x_max.view(x_max.size(0), -1)
x = torch.cat([x_avg, x_max], dim=-1)
x = self.classifier(x)
return x
# --
# Define model
print('cifar10.py: initializing model...', file=sys.stderr)
cuda = torch.device('cuda')
model = ResNet18().to(cuda).half()
model.verbose = True
print(model, file=sys.stderr)
# --
# Initialize optimizer
print('cifar10.py: initializing optimizer...', file=sys.stderr)
if args.lr_schedule == 'linear_cycle':
lr_scheduler = HPSchedule.linear_cycle(hp_max=args.lr_max, epochs=args.epochs, extra=args.extra)
elif args.lr_schedule == 'sgdr':
lr_scheduler = HPSchedule.sgdr(
hp_init=args.lr_max,
period_length=args.sgdr_period_length,
t_mult=args.sgdr_t_mult,
)
else:
lr_scheduler = getattr(HPSchedule, args.lr_schedule)(hp_max=args.lr_max, epochs=args.epochs)
model.init_optimizer(
opt=torch.optim.SGD,
params=model.parameters(),
hp_scheduler={"lr" : lr_scheduler},
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=True,
)
# --
# Train
print('cifar10.py: training...', file=sys.stderr)
t = time()
for epoch in range(args.epochs + args.extra + args.burnout):
train = model.train_epoch(dataloaders, mode='train', metric_fns=['n_correct'])
test = model.eval_epoch(dataloaders, mode='test', metric_fns=['n_correct'])
print(json.dumps({
"epoch" : int(epoch),
"lr" : model.hp['lr'],
"test_acc" : float(test['acc']),
"train_acc" : float(train['acc']),
"time" : time() - t,
}))
sys.stdout.flush()
model.save('weights')