-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_dist.py
197 lines (169 loc) · 8.47 KB
/
train_dist.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
import argparse
import torch
import pickle
import torch.nn as nn
import numpy as np
import os
import glob
import pdb
from data_loader import SeqVolumeDataset
from model import ActionNet
from torch.autograd import Variable
from torch.nn.utils.rnn import pack_padded_sequence
import torch.utils.data
import torch.utils.data.distributed
from torchvision import transforms
import logging
#from colorlog import ColoredFormatter
import horovod.torch as hvd
from loss import ContrastiveLoss
def metric_average(val, name):
tensor = torch.tensor(val)
avg_tensor = hvd.allreduce(tensor, name=name, average = False)
return avg_tensor.item()
def main(args):
#if not os.path.exists(args.model_path):
# os.makedirs(args.model_path)
hvd.init()
logging.basicConfig(level=logging.INFO,
format='%(asctime)s %(name)-12s %(levelname)-8s %(message)s',
datefmt='%m-%d %H:%M',
filename='{0}/training.log'.format(args.model_path),
filemode='a')
# define a Handler which writes INFO messages or higher to the sys.stderr
console = logging.StreamHandler()
console.setLevel(logging.INFO)
formatter = logging.Formatter( "%(asctime)s %(name)-12s %(levelname)-8s %(message)s")
console.setFormatter(formatter)
logging.getLogger('').addHandler(console)
data_loader_val = None
data_set_training = SeqVolumeDataset(args.data_dir, args.list_fn, args.seq_len, w = 61, h = 61, d= 2100//25 + 1)
data_sampler = torch.utils.data.distributed.DistributedSampler(data_set_training, num_replicas=hvd.size(), rank=hvd.rank())
data_loader = torch.utils.data.DataLoader(
data_set_training, batch_size=args.batch_size, shuffle=(data_sampler is None),
num_workers=args.num_workers, pin_memory=True, sampler=data_sampler)
if args.list_fn_val:
data_set_val = SeqVolumeDataset(args.data_dir_val, args.list_fn_val, args.seq_len, w = 61, h = 61, d= 2100//25 + 1)
data_sampler_val = torch.utils.data.distributed.DistributedSampler(data_set_val, num_replicas=hvd.size(), rank=hvd.rank())
data_loader_val = torch.utils.data.DataLoader(data_set_val, batch_size = args.batch_size, shuffle = False, num_workers = args.num_workers, pin_memory = True, sampler = data_sampler_val)
net = ActionNet(args.hidden_size, args.class_num)
print(net)
logging.info('HVD size %d, HVD rank %d', hvd.size(), hvd.rank())
if torch.cuda.is_available():
net.cuda()
net = nn.DataParallel(net)
if args.model_fn:
logging.info('Loading from %s', args.model_fn)
net.load_state_dict( torch.load(args.model_fn) )
# Loss and Optimizer
criterion = nn.CrossEntropyLoss()
criterion_gc = ContrastiveLoss(0.5)
params = net.parameters()
optimizer = torch.optim.Adam(params, lr=args.learning_rate)
optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=net.named_parameters())
hvd.broadcast_parameters(net.state_dict(), root_rank=0)
# Train the Models
total_step = len(data_loader)
for epoch in range(args.num_epochs):
data_sampler.set_epoch(epoch)
net.train()
for i, (data, lbls) in enumerate(data_loader):
# Set mini-batch dataset
data = Variable(data)
lbls = Variable(lbls)
if torch.cuda.is_available():
data = data.cuda()
lbls = lbls.cuda()
if data.size(0) != args.batch_size:
continue
net.zero_grad()
cur_batch_size = data.size(0)
outputs, hs = net(data)
lbls = lbls.expand(-1, outputs.size(1)).contiguous()
outputs = outputs.view(outputs.size(0) * outputs.size(1), outputs.size(2))
lbls = lbls.view(-1)
loss_gc = criterion_gc(hs)
loss_c = criterion(outputs, lbls)
loss = loss_gc + loss_c
loss.backward()
accuracy = (lbls.data == outputs.data.max(dim = 1)[1]).sum().item() * 1.0 / lbls.size(0)
optimizer.step()
if i % args.log_step == 0:
logging.info('Epoch [%d/%d], Step [%d/%d], Rank[%d], Loss_gc: %.4f, Loss_c: %.4f, Loss: %.4f, Accuracy: %.4f'
,epoch, args.num_epochs, i, total_step, hvd.rank(), loss_gc.data.item(), loss_c.data.item(), loss.data.item(), accuracy)
if i == 0:
os.system('nvidia-smi')
if hvd.rank() == 0:
torch.save(net.state_dict(),
os.path.join(args.model_path,
'action-net-%d.pkl' %(epoch+1)))
# Now testing.
if data_loader_val:
net.eval()
val_total = 0
val_correct = 0
with torch.no_grad():
for i, (data, lbls) in enumerate(data_loader_val):
if True:
data = Variable(data)
lbls = Variable(lbls)
if torch.cuda.is_available():
data = data.cuda()
lbls = lbls.cuda()
cur_batch_size = data.size(0)
if cur_batch_size != args.batch_size:
continue
outputs, hs = net(data)
#lbls = lbls.unsqueeze(1)
lbls = lbls.expand(-1, outputs.size(1)).contiguous()
outputs = outputs.view(outputs.size(0) * outputs.size(1), outputs.size(2))
lbls = lbls.view(-1)
loss_c = criterion(outputs, lbls)
loss_gc = criterion_gc(hs)
loss = loss_c + loss_gc
correct = (lbls.data == outputs.data.max(dim = 1)[1]).sum().item()
accuracy = correct * 1.0 / lbls.size(0)
val_correct += correct
val_total += lbls.size(0)
# Print log info
if i % args.log_step == 0:
logging.info('Testing Epoch [%d/%d], Step [%d/%d], Rank[%d], Loss_gc: %.4f, Loss_c: %.4f, Loss: %.4f, Accuracy: %.4f'
,epoch, args.num_epochs, i, len(data_loader_val), hvd.rank(), loss_gc.item(), loss_c.item(), loss.item(), accuracy)
val_t = metric_average(val_total, 'sum_total')
val_c = metric_average(val_correct, 'sum_correct')
if hvd.rank() == 0:
logging.info('Testing Epoch [%d/%d], Val Accuracy: %.4f', epoch, args.num_epochs, val_c * 1.0 / val_t)
#logging.info('Testing Epoch [%d/%d], Val Accuracy: %.4f', epoch, args.num_epochs, val_correct * 1.0 / val_total)
if hvd.rank() == 0:
torch.save(net.state_dict(),
os.path.join(args.model_path,
'action-net-%d.pkl' %(epoch+1)))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--model_path', type=str, default='./models/' ,
help='path for saving trained models')
parser.add_argument('--model_fn', type = str, default = '')
parser.add_argument('--data_dir', type=str)
parser.add_argument('--data_dir_val', type=str)
parser.add_argument('--list_fn', type=str,
help='list of the video clips')
parser.add_argument('--list_fn_val', type=str,
help='list of the video clips')
parser.add_argument('--log_step', type=int , default=100,
help='step size for prining log info')
parser.add_argument('--save_step', type=int , default=100,
help='step size for saving trained models')
parser.add_argument('--num_layers', type=int , default=1 ,
help='number of layers in lstm')
parser.add_argument('--class_num', type=int , default=17,
help='number of class')
parser.add_argument('--hidden_size', type=int , default=512,
help='number of class')
parser.add_argument('--num_epochs', type=int, default=100)
parser.add_argument('--seq_len', type=int, default=10)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--num_workers', type=int, default=10)
parser.add_argument('--learning_rate', type=float, default=0.0001)
args = parser.parse_args()
print(args)
main(args)