-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
193 lines (155 loc) · 6.24 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
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import sys
import os
import time
import numpy as np
import argparse
import torch
import torchvision
import torch.nn as nn
import torch.optim as optim
from torch.utils.tensorboard import SummaryWriter
from apex import amp, optimizers
from apex.parallel import DistributedDataParallel
import torch.multiprocessing
import logging
from utils import logging_utils
logging_utils.config_logger()
from utils.YParams import YParams
from utils.data_loader import get_data_loader_distributed
from utils.plotting import generate_images, meanL1
from networks import UNet
def adjust_LR(optimizer, params, iternum):
"""Piecewise constant rate decay"""
if params.distributed and iternum<5000:
lr = params.ngpu*params.lr*(iternum/5000.) #warmup for distributed training
elif iternum<40000:
lr = params.ngpu*params.lr
elif iternum>80000:
lr = params.ngpu*params.lr/4.
else:
lr = params.ngpu*params.lr/2.
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def train(params, args, world_rank):
logging.info('rank %d, begin data loader init'%world_rank)
train_data_loader = get_data_loader_distributed(params, world_rank)
logging.info('rank %d, data loader initialized'%world_rank)
model = UNet.UNet(params).cuda()
if not args.resuming:
model.apply(model.get_weights_function(params.weight_init))
optimizer = optimizers.FusedAdam(model.parameters(), lr = params.lr)
#model, optimizer = amp.initialize(model, optimizer, opt_level="O1") # for automatic mixed precision
if params.distributed:
model = DistributedDataParallel(model)
iters = 0
startEpoch = 0
checkpoint = None
if args.resuming:
if world_rank==0:
logging.info("Loading checkpoint %s"%params.checkpoint_path)
checkpoint = torch.load(params.checkpoint_path, map_location='cuda:{}'.format(args.local_rank))
model.load_state_dict(checkpoint['model_state'])
iters = checkpoint['iters']
startEpoch = checkpoint['epoch'] + 1
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
if world_rank==0:
logging.info(model)
logging.info("Starting Training Loop...")
device = torch.cuda.current_device()
for epoch in range(startEpoch, startEpoch+params.num_epochs):
start = time.time()
tr_time = 0.
log_time = 0.
for i, data in enumerate(train_data_loader, 0):
iters += 1
adjust_LR(optimizer, params, iters)
inp, tar = map(lambda x: x.to(device), data)
tr_start = time.time()
b_size = inp.size(0)
model.zero_grad()
gen = model(inp)
loss = UNet.loss_func(gen, tar, params)
loss.backward() # fixed precision
# automatic mixed precision:
#with amp.scale_loss(loss, optimizer) as scaled_loss:
# scaled_loss.backward()
optimizer.step()
tr_end = time.time()
tr_time += tr_end - tr_start
# Output training stats
if world_rank==0:
log_start = time.time()
gens = []
tars = []
with torch.no_grad():
for i, data in enumerate(train_data_loader, 0):
if i>=16:
break
inp, tar = map(lambda x: x.to(device), data)
gen = model(inp)
gens.append(gen.detach().cpu().numpy())
tars.append(tar.detach().cpu().numpy())
gens = np.concatenate(gens, axis=0)
tars = np.concatenate(tars, axis=0)
# Scalars
args.tboard_writer.add_scalar('G_loss', loss.item(), iters)
# Plots
fig, chi, L1score = meanL1(gens, tars)
args.tboard_writer.add_figure('pixhist', fig, iters, close=True)
args.tboard_writer.add_scalar('Metrics/chi', chi, iters)
args.tboard_writer.add_scalar('Metrics/rhoL1', L1score[0], iters)
args.tboard_writer.add_scalar('Metrics/vxL1', L1score[1], iters)
args.tboard_writer.add_scalar('Metrics/vyL1', L1score[2], iters)
args.tboard_writer.add_scalar('Metrics/vzL1', L1score[3], iters)
args.tboard_writer.add_scalar('Metrics/TL1', L1score[4], iters)
fig = generate_images(inp.detach().cpu().numpy()[0], gens[-1], tars[-1])
args.tboard_writer.add_figure('genimg', fig, iters, close=True)
log_end = time.time()
log_time += log_end - log_start
# Save checkpoint
torch.save({'iters': iters, 'epoch':epoch, 'model_state': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()}, params.checkpoint_path)
end = time.time()
if world_rank==0:
logging.info('Time taken for epoch {} is {} sec'.format(epoch + 1, end-start))
logging.info('train step time={}, logging time={}'.format(tr_time, log_time))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", default=0, type=int)
parser.add_argument("--run_num", default='00', type=str)
parser.add_argument("--yaml_config", default='./config/UNet.yaml', type=str)
parser.add_argument("--config", default='default', type=str)
args = parser.parse_args()
run_num = args.run_num
params = YParams(os.path.abspath(args.yaml_config), args.config)
params.distributed = False
if 'WORLD_SIZE' in os.environ:
params.distributed = int(os.environ['WORLD_SIZE']) > 1
world_rank = 0
if params.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.gpu = args.local_rank
world_rank = torch.distributed.get_rank()
torch.backends.cudnn.benchmark = True
args.resuming = False
# Set up directory
baseDir = './expts/'
expDir = os.path.join(baseDir, args.config+'/'+str(run_num)+'/')
if world_rank==0:
if not os.path.isdir(expDir):
os.mkdir(expDir)
os.mkdir(expDir+'training_checkpoints/')
logging_utils.log_to_file(logger_name=None, log_filename=os.path.join(expDir, 'out.log'))
params.log()
args.tboard_writer = SummaryWriter(log_dir=os.path.join(expDir, 'logs/'))
params.experiment_dir = os.path.abspath(expDir)
params.checkpoint_path = os.path.join(params.experiment_dir, 'training_checkpoints/ckpt.tar')
if os.path.isfile(params.checkpoint_path):
args.resuming=True
train(params, args, world_rank)
if world_rank == 0:
args.tboard_writer.flush()
args.tboard_writer.close()
logging.info('DONE ---- rank %d'%world_rank)