-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_masks.py
223 lines (191 loc) · 6.12 KB
/
train_masks.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
import os
from glob import glob
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from loaddata import Loaders
from networks import MaskCore
from easydict import EasyDict
from train_models import fix_seeds
from flags import args
def epoch_mask_train(
model,
loader,
optimizer,
criterion,
mask_decay,
schedule,
device
):
model.train()
model.core.eval()
logs = EasyDict(
invariance=torch.zeros(len(loader)),
sparsity=torch.zeros(len(loader))
)
for i, (inputs, targets, m_outs) in enumerate(loader):
inputs = inputs.to(device)
outs = model(inputs)
invariance = criterion(outs, targets.to(device))
invariance -= criterion(m_outs.to(device), targets.to(device))
invariance = invariance**2
logs.invariance[i] = invariance.item()
invariance = torch.exp(invariance)
norm = torch.norm(model.mask.weight, p=1)
logs.sparsity[i] = norm
sparsity = mask_decay * norm
loss = invariance + sparsity
loss.backward()
optimizer.step()
optimizer.zero_grad()
if schedule:
schedule.step()
w = model.mask.weight.data
w = w.clamp(0., 1.)
model.mask.weight.data = w
logs.invariance = logs.invariance.mean().item()
logs.sparsity = logs.sparsity.mean().item()
return logs
def epoch_mask_test(
model,
loader,
device
):
model.eval()
correct = 0
with torch.no_grad():
for inputs, targets, _ in loader:
inputs = inputs.to(device)
outputs = model(inputs)
predicted = torch.argmax(outputs, dim=1).detach().cpu()
correct += predicted.eq(targets.view_as(predicted)).sum()
accuracy = 100 * correct.numpy()/len(loader.dataset)
return accuracy
def train_mask(
architecture,
pretrained_model,
epochs,
class_portion,
lr,
schedule,
mask_size,
mask_decay,
patience,
seed,
save_name
):
# GPU configs
device = 'cuda' if torch.cuda.is_available() else 'cpu'
torch.cuda.empty_cache()
fix_seeds(seed)
# load the model architecture
model = MaskCore(
mask_size=mask_size,
num_classes=5,
architecture=architecture,
pretrained_model=pretrained_model
)
# freeze the weights of the pre-trained network
for p in model.core.parameters():
p.requires_grad = False
model = model.to(device)
cudnn.benchmark = True
# init lambda for sparsity loss term
init_sparsity = torch.norm(model.mask.weight.data, p=1)
mask_decay = mask_decay / init_sparsity
# optimization configs
lossfunction = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
loaders = Loaders(
batch_size=1,
class_portion=class_portion,
shuffle_train=True,
shuffle_test=True,
num_workers=3,
transform='n',
pretrained_model=pretrained_model,
architecture=architecture
)
loader = loaders.testloader()
init_accuracy = epoch_mask_test(model, loader, device)
# learning rate scheduling
if schedule:
sched = torch.optim.lr_scheduler.OneCycleLR(
optimizer,
lr,
epochs=epochs,
steps_per_epoch=len(loader),
pct_start=schedule
)
else: sched = False
# prep logs
logs = EasyDict(invariance=[], sparsity=[])
# Train
iterbar = tqdm(range(1, epochs + 1), total=epochs)
cnt = 0
for epoch in iterbar:
losses = epoch_mask_train(
model,
loader,
optimizer,
lossfunction,
mask_decay,
sched,
device
)
# get losses
logs.invariance.append(losses.invariance)
logs.sparsity.append(losses.sparsity/init_sparsity)
accuracy = epoch_mask_test(model, loader, device)
description = f'Invariance: {logs.invariance[epoch-1]:.3f} - '\
f'Sparsity: {logs.sparsity[epoch-1]*100:.2f}% - '\
f'Init Accuracy: {init_accuracy:.2f}% - '\
f'Masked Accuracy: {accuracy:.2f}% - '\
f'Count: {cnt}'
iterbar.set_description(desc=description)
# stop at convergence
if epoch > 1 and accuracy/init_accuracy >= 0.99:
torch.save(model.mask.weight.detach().cpu(), save_name)
e = epoch-1
if abs(logs.sparsity[e] - logs.sparsity[e-1]) <= 1e-5:
cnt = cnt + 1
else: cnt = 0
if cnt==patience:
break
del model
def main(
architecture,
pretrained_model_folder,
mask_save_folder
):
if not os.path.exists(mask_save_folder):
os.makedirs(mask_save_folder)
model_paths = glob(pretrained_model_folder + f'/{architecture}*.pt')
if len(model_paths)==0: print('base models have not been trained yet')
else:
for p in model_paths:
pretrained_model = p
p = os.path.basename(p).rsplit('_')[-1]
save_path = mask_save_folder + f'/{architecture}_mask_' + p
if not os.path.exists(save_path):
train_mask(
architecture,
pretrained_model,
epochs=args.mask_epochs,
class_portion=args.class_portion,
lr=args.mask_lr,
schedule=args.mask_schedule,
mask_size=args.img_size,
mask_decay=args.mask_decay,
patience=args.mask_patience,
seed=args.seed,
save_name=save_path
)
else: print(f'{save_path} is already trained')
if __name__ == "__main__":
main(
architecture=args.architecture,
pretrained_model_folder=args.model_save_folder,
mask_save_folder=args.mask_save_folder
)