-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
178 lines (137 loc) · 4.59 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
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
import torch as th
import torch.optim as optim
import torch.nn as nn
import torchnet as tnt
from timeit import default_timer as timer
import numpy as np
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
from pprint import pprint
import os, pdb, sys, json, subprocess, \
argparse, math, copy, glob
from copy import deepcopy
from scipy import interpolate
from blowtorch import models, loss, exptutils, loader, viz
opt = exptutils.add_args([
['-o', '/data/results', 'output'],
['-m', 'lenet', 'lenet | mnistfc | allcnn | wrn* | resnet*'],
['--frac', 1., 'dataset fraction'],
['--dataset', 'mnist', 'mnist | cifar10 | cifar100 | imagenet | svhn*'],
['-b', 128, 'bsz'],
['-B', 100, '#epochs'],
['-j', 1, '#gpus'],
['--lr', 0.1, 'lr'],
['--lrs', '[[0,0.1]]', 'lr schedule'],
['-s', 42, 'seed'],
['--env', '', 'visdom env'],
['-v', False, 'verbose'],
['--freq', 1, 'val/save freq.'],
['--save', False, 'save'],
['-r', '', 'reload model']
])
opt['augment'] = True
def train(e, model, criterion, optimizer):
model.train()
lr = exptutils.schedule(e, opt, 'lr')
exptutils.set_lr(optimizer, lr)
g = opt['g']
ds = loader.get_loader(opt, is_train=True)
maxb = len(ds)
ts, ts2 = timer(), timer()
s = dict(lr=lr, e=e, f=[], top1=[])
for bi, (x,y) in enumerate(ds):
x, y = x.to(g), y.to(g)
model.zero_grad()
yh = model(x)
f = criterion(yh, y)
f.backward()
optimizer.step()
s['f'].append(f.item())
s['top1'].append(exptutils.error(yh, y))
if timer() - ts2 > 5:
print((exptutils.color('blue', '[%2d][%4d/%4d] %2.4f %.2f%%'))%(e,bi,maxb,
np.mean(s['f']), np.mean(s['top1'])))
ts2 = timer()
print((exptutils.color('blue', '+[%2d] %2.4f %2.2f%% [%.2fs]'))% (e,
np.mean(s['f']), np.mean(s['top1']), timer()-ts))
return s
def val(e, model, criterion):
model.eval()
g = opt['g']
ds = loader.get_loader(opt, is_train=False)
maxb = len(ds)
ts, ts2 = timer(), timer()
s = dict(e=e, f=[], top1=[])
with th.no_grad():
for bi, (x,y) in enumerate(ds):
x, y = x.to(g), y.to(g)
yh = model(x)
f = criterion(yh, y)
s['f'].append(f.item())
s['top1'].append(exptutils.error(yh, y))
if timer() - ts2 > 5:
print((exptutils.color('red', '[%2d][%4d/%4d] %2.4f %.2f%%'))%(e,bi,maxb,
np.mean(s['f']), np.mean(s['top1'])))
ts2 = timer()
print((exptutils.color('red', '*[%2d] %2.4f %2.2f%% [%.2fs]'))% (e,
np.mean(s['f']), np.mean(s['top1']), timer()-ts))
return s
def reload(model):
if opt['r'] == '':
return 0, [], []
assert os.path.exists(opt['r']), 'Could not find: %s'%opt['r']
def check_opt(o):
global opt
print('Old opt: ')
pprint(opt)
print('New opt: ')
pprint(o)
r = input('Press y[yes] to continue: ')
if r == 'y' or r == 'yes':
opt = deepcopy(o)
d = th.load(opt['r'])
model.load_state_dict(d['state_dict'])
check_opt(d['opt'])
return d['e']+1, d['train_stats'], d['val_stats']
def save(d):
if not opt['save']:
return
fn = os.path.join(opt['o'], opt['fname'] + '.pt')
th.save(d, fn)
def setup():
# setup rand and gpus
exptutils.setup(opt)
# logging/saving
blklist = ['augment', 'b', 'd', 'frac', 'g', 'gs', 'env', 'freq', 'meta',
'j', 'lr', 'lrs', 's', 'save', 'v', 'l2','depth','widen']
if not 'fname' in opt:
exptutils.build_filename(opt, blklist)
# git status
opt['meta'] = exptutils.gitrev(opt)
# visdom
s = opt['fname']
opt['title'] = s[s.find("{")+1:s.find("}")]
def main():
model = getattr(models, opt['m'])(opt)
criterion = loss.wrap(nn.CrossEntropyLoss(), loss.ell2(opt, model))
lr = exptutils.schedule(0, opt, 'lr')
optimizer = th.optim.SGD(model.parameters(), lr=lr,
nesterov=True, momentum=0.9)
start_e, sts, svs = reload(model)
model, criterion = exptutils.cudafy(opt, model, criterion)
pprint(opt)
for e in range(start_e, opt['B']):
print('')
st, sv = None, None
st = train(e, model, criterion, optimizer)
sts.append(st)
if e % opt['freq'] == 0 or e == opt['B']-1:
sv = val(e, model, criterion)
svs.append(sv)
save(dict(
opt=opt,
e=e, train_stats=sts, val_stats=svs,
state_dict=model.state_dict())
)
setup()
main()