-
Notifications
You must be signed in to change notification settings - Fork 42
/
train.py
129 lines (113 loc) · 5.1 KB
/
train.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
from model import *
from plot import *
import torch
from torch.autograd import Variable
import ipdb
class classifier():
def __init__(self, args, data):
self.train_loader = data.train_loader
self.test_loader = data.test_loader
self.batch_size = args.batch_size
self.num_train = data.num_train
self.num_classes = data.num_classes
assert args.model_type == 'mlp' or args.model_type == 'cnn'
if args.model_type == 'mlp':
self.net = mlp(args.conditioned, data.input_dims, data.num_classes, hidden_size=256)
elif args.model_type == 'cnn':
self.net = cnn(data.in_channel, args.conditioned, data.num_classes)
if args.use_gpu:
self.net.cuda()
self.classificationCriterion = nn.CrossEntropyLoss()
self.syntheticCriterion = nn.MSELoss()
self.plot = args.plot
self.num_epochs = args.num_epochs
self.model_name = args.model_name
self.conditioned = args.conditioned
self.best_perf = 0.
self.stats = dict(grad_loss=[], classify_loss=[])
print "[%] model name will be", self.model_name
def optimizer_module(self, optimizer, forward, out, label_onehot=None):
optimizer.zero_grad()
out, grad = forward(out, label_onehot)
out.backward(grad.detach().data)
optimizer.step()
out = out.detach()
return out
def save_grad(self, name):
def hook(grad):
self.backprop_grads[name] = grad
self.backprop_grads[name].volatile = False
return hook
def optimizer_dni_module(self, images, labels, label_onehot, grad_optimizer, optimizer, forward):
# synthetic model
# Forward + Backward + Optimize
grad_optimizer.zero_grad()
optimizer.zero_grad()
outs, grads = forward(images, label_onehot)
self.backprop_grads = {}
handles = {}
keys = []
for i, (out, grad) in enumerate(zip(outs, grads)):
handles[str(i)] = out.register_hook(self.save_grad(str(i)))
keys.append(str(i))
outputs = outs[-1]
loss = self.classificationCriterion(outputs, labels)
loss.backward(retain_variables=True)
for (k, v) in handles.items():
v.remove()
grad_loss = 0.
for k in keys:
grad_loss += self.syntheticCriterion(grads[int(k)], self.backprop_grads[k].detach())
grad_loss.backward()
grad_optimizer.step()
self.stats['grad_loss'].append(grad_loss.data[0])
self.stats['classify_loss'].append(loss.data[0])
return loss, grad_loss
def train_model(self):
for epoch in range(self.num_epochs):
for i, (images, labels) in enumerate(self.train_loader):
# Convert torch tensor to Variable
labels_onehot = torch.zeros([labels.size(0), self.num_classes])
labels_onehot.scatter_(1, labels.unsqueeze(1), 1)
images = Variable(images).cuda()
labels = Variable(labels).cuda()
labels_onehot = Variable(labels_onehot).cuda()
out = images
# Forward + Backward + Optimize
for (optimizer, forward) in zip(self.net.optimizers, self.net.forwards):
if self.conditioned:
out = self.optimizer_module(optimizer, forward, out, labels_onehot)
else:
out = self.optimizer_module(optimizer, forward, out)
# synthetic model
# Forward + Backward + Optimize
loss, grad_loss = self.optimizer_dni_module(images, labels, labels_onehot,
self.net.grad_optimizer, self.net.optimizer, self.net)
if (i+1) % 100 == 0:
print ('Epoch [%d/%d], Step [%d/%d], Loss: %.4f, Grad Loss: %.4f'
%(epoch+1, self.num_epochs, i+1, self.num_train//self.batch_size, loss.data[0], grad_loss.data[0]))
if (epoch+1) % 10 == 0:
perf = self.test_model(epoch+1)
if perf > self.best_perf:
torch.save(self.net.state_dict(), self.model_name+'_model_best.pkl')
self.net.train()
# Save the Model ans Stats
pkl.dump(self.stats, open(self.model_name+'_stats.pkl', 'wb'))
torch.save(self.net.state_dict(), self.model_name+'_model.pkl')
if self.plot:
plot(self.stats, name=self.model_name)
def test_model(self, epoch):
# Test the Model
self.net.eval()
correct = 0
total = 0
for images, labels in self.test_loader:
images = Variable(images).cuda()
outputs = self.net(images)
outputs = outputs[-1]
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted.cpu() == labels).sum()
perf = 100 * correct / total
print('Epoch %d: Accuracy of the network on the 10000 test images: %d %%' % (epoch, perf))
return perf