-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathvtrain.py
189 lines (154 loc) · 7.2 KB
/
vtrain.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
import time
import pickle as pkl
import torch
import torch.nn as nn
import torchvision
from torchvision import models, transforms
from torch.optim import SGD, Adam, RMSprop, AdamW
from vmodel import Activs_prober, Conv_prober
from aggmo import AggMo
from adabelief import AdaBelief
# from kfac import KFAC
def get_dataloader(use_data, download, bsize):
######### Dataloaders #########
transform = transforms.Compose(
[transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
transform_test = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
d_path = "./"
if (use_data=="CIFAR-10"):
n_classes = 10
trainset = torchvision.datasets.CIFAR10(root=d_path+'datasets/cifar10/', train=True, download=(download), transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=bsize, shuffle=True, num_workers=2)
testset = torchvision.datasets.CIFAR10(root=d_path+'datasets/cifar10/', train=False, download=(download), transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=bsize, shuffle=False, num_workers=2)
elif (use_data=="CIFAR-100"):
n_classes = 100
trainset = torchvision.datasets.CIFAR100(root=d_path+'datasets/cifar100/', train=True, download=(download), transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=bsize, shuffle=True, num_workers=2)
testset = torchvision.datasets.CIFAR100(root=d_path+'datasets/cifar100/', train=False, download=(download), transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=bsize, shuffle=False, num_workers=2)
elif (use_data=="MNIST"):
n_classes = 10
trainset = torchvision.datasets.MNIST(root=d_path+'datasets/mnist/', train=True, download=(download), transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=bsize, shuffle=True, num_workers=2)
testset = torchvision.datasets.MNIST(root=d_path+'datasets/mnist/', train=False, download=(download), transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=bsize, shuffle=False, num_workers=2)
else:
raise Exception("Not CIFAR-10/CIFAR-100")
return trainloader, testloader
def get_optimizer(net, lr, wd, mom, opt_type="SGD"):
if opt_type == "SGD":
optimizer = SGD(net.parameters(), lr=lr, momentum=mom, weight_decay=wd)
elif opt_type == "RMSprop":
optimizer = RMSprop(net.parameters(), lr=lr, momentum=mom, weight_decay=wd)
elif opt_type == "Adam":
optimizer = Adam(net.parameters(), lr=lr, weight_decay=wd)
# elif opt_type == "KFAC":
# optimizer = KFAC(net, damping=False, learning_rate=lr, weight_decay=wd)
elif opt_type == "AggMo":
optimizer = AggMo(net.parameters(), lr=lr, weight_decay=wd)
elif opt_type == "AdaBelief":
optimizer = AdaBelief(net.parameters(), lr=lr, weight_decay=wd)
return optimizer
def rescale(net, net_base):
param_norms = {}
for idx, (mod, mod_base) in enumerate(zip(net.modules(), net_base.modules())):
if(isinstance(mod, nn.Conv2d)):
#print( 'before='+ str( torch.norm(torch.norm(mod.weight, dim=(2,3), keepdim=True), dim=1, keepdim=True)[1]) )
mod.weight.data = (mod.weight.data / torch.linalg.norm(mod.weight, dim=(1,2,3), keepdim=True)) * torch.linalg.norm(mod_base.weight, dim=(1,2,3), keepdim=True)
#print( 'after='+ str( torch.norm(torch.norm(mod.weight, dim=(2,3), keepdim=True), dim=1, keepdim=True)[1]) )
# saved_weights = torch.clone(mod.weight).detach()
# param_norms[idx] = torch.norm(torch.norm(saved_weights, dim=(2,3), keepdim=True), dim=1, keepdim=True)[1]
#print('end!')
return net
def train(net, net_base, dataloader, optimizer, criterion, device, batch_size, epoch, ablate=False):
start = time.time()
net.train()
correct = 0.0
total = 0.0
for batch_index, (images, labels) in enumerate(dataloader):
labels = labels.to(device)
images = images.to(device)
optimizer.zero_grad()
outputs = net(images)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
_, preds = outputs.max(1)
correct += preds.eq(labels).sum()
total += batch_size
if ablate:
net = rescale(net, net_base)
n_iter = (epoch - 1) * len(dataloader) + batch_index + 1
print('Training Epoch: {epoch} [{trained_samples}/{total_samples}]\tAccuracy: {:0.4f}\tLoss: {:0.4f}\tLR: {:0.6f}'.format(
correct.float() / total,
loss.item(),
optimizer.param_groups[0]['lr'],
epoch=epoch,
trained_samples=batch_index * batch_size + len(images),
total_samples=len(dataloader.dataset)
), end="\r")
finish = time.time()
print('epoch {} training time consumed: {:.2f}s'.format(epoch, finish - start))
return correct.float() / len(dataloader.dataset), loss.item()
@torch.no_grad()
def eval(net, dataloader, criterion, device, epoch=0, tb=True):
start = time.time()
net.eval()
test_loss = 0.0 # cost function error
correct = 0.0
for (images, labels) in dataloader:
images = images.to(device)
labels = labels.to(device)
outputs = net(images)
loss = criterion(outputs, labels)
test_loss += loss.item()
_, preds = outputs.max(1)
correct += preds.eq(labels).sum()
finish = time.time()
print('Evaluating Network.....')
print('Test set: Epoch: {}, Average loss: {:.4f}, Accuracy: {:.4f}, Time consumed:{:.2f}s'.format(
epoch,
test_loss / len(dataloader.dataset),
correct.float() / len(dataloader.dataset),
finish - start
))
print()
return correct.float() / len(dataloader.dataset), test_loss / len(dataloader.dataset)
def net_save(net, accs_dict, trained_root, suffix, ablate=False):
print('Saving Model...')
state = {
'net': net.state_dict(),
'Train_acc': accs_dict['Train'],
'Test_acc': accs_dict['Test'],
'lr': accs_dict['lr'],
'train_loss': accs_dict['train_loss'],
'test_loss': accs_dict['test_loss']
}
torch.save(state, trained_root + 'VGG_model_{}.pth'.format(suffix))
props_dict = {
"params_list": net.params_list,
"grads_list": net.grads_list,
"activs_norms": [],
"activs_corr": [],
# "activs_ranks": [],
"std_list": [],
"grads_norms": [],
}
for mod in net.modules():
if(isinstance(mod, Activs_prober)):
props_dict["activs_norms"].append(mod.activs_norms)
props_dict["activs_corr"].append(mod.activs_corr)
# props_dict["activs_ranks"].append(mod.activs_ranks)
if(isinstance(mod, Conv_prober)):
props_dict["std_list"].append(mod.std_list)
props_dict["grads_norms"].append(mod.grads_norms)
print('Saving properties...')
with open(trained_root + "properties_{}.pkl".format(suffix), 'wb') as f:
pkl.dump(props_dict, f)