-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
271 lines (211 loc) · 11.4 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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
import os
# os.environ["CUDA_VISIBLE_DEVICES"] = '0,1'
print(os.environ["CUDA_VISIBLE_DEVICES"])
import torch
import torch.utils.data as data
from torch import optim, nn
from tensorboardX import SummaryWriter
import numpy as np
import argparse
import random
from utils.configs import set_configs
from data_readers.train_data_loaders import TrainfusedEventData
from e2v.e2v_model import *
from utils.evaluate import PerceptualLoss
from pytorch_msssim import SSIM
from utils.data_io import show_whole_img
from loss import FlowReconLoss #FlowL1LossDict
from utils.flow_utils import FrameWarp
from DCEIFlow.utils.utils import setup_seed
class Train:
'''Train CISTA-Flow after getting pretrained CISTA (GT Flow) and DCEIFlow/ERAFT (GT I)'''
def __init__(self, cfgs, device):
# self.image_dim = cfgs.image_dim
self.device = device
if cfgs.model_name:
self.model_name = '{}_{}_b{}_d{}_c{}'.format(cfgs.model_name, cfgs.model_mode, \
cfgs.num_bins, cfgs.depth, cfgs.base_channels)
else:
self.model_name = '{}_b{}_d{}_c{}'.format(cfgs.model_mode, \
cfgs.num_bins, cfgs.depth, cfgs.base_channels)
self.path_to_model = os.path.join(cfgs.path_to_model, self.model_name)
if not os.path.exists(self.path_to_model):
os.makedirs(self.path_to_model)
# Loss
self.frame_warp = FrameWarp(mode=cfgs.warp_mode)
self.loss_fn = FlowReconLoss(cfgs.image_dim, self.frame_warp, ds=cfgs.ds, is_bi=False, lpips_net='vgg').to(self.device) #self.device
self.model_mode = cfgs.model_mode
if self.model_mode == 'cista-eiflow':
self.model = DCEIFlowCistaNet(cfgs)
if cfgs.distributed:
self.model = DCEIFlowCistaNet2GPU(cfgs)
elif self.model_mode == 'cista-eraft':
self.model = ERAFTCistaNet(cfgs)
else:
assert self.model_mode in ['cista-eiflow', 'cista-eraft']
if cfgs.load_epoch_for_train:
checkpoint = torch.load(os.path.join(self.path_to_model, '{}_{}.pth.tar'\
.format(self.model_name, cfgs.load_epoch_for_train)), map_location='cpu')
self.model.load_state_dict(checkpoint['state_dict'], strict=True) #True
elif cfgs.path_to_e2vflow: # if having pretrained CISTA-Flow network
checkpoint = torch.load(cfgs.path_to_e2vflow, map_location='cpu')
self.model.load_state_dict(checkpoint['state_dict'], strict=True)
print('Load path_to_e2vflow: {}'.format(cfgs.path_to_e2vflow))
else:
# load pretrained reconstruction network CISTA (GT Flow)
if cfgs.path_to_e2v:
checkpoint = torch.load(cfgs.path_to_e2v, map_location='cpu')
self.model.cista_net.load_state_dict(checkpoint['state_dict'], strict=True)
print('Load path_to_e2v: {}'.format(cfgs.path_to_e2v))
else:
assert cfgs.path_to_e2v, "Should load pretrained CISTA (GT Flow)"
# load pretrained flow network
if cfgs.path_to_flownet:
checkpoint = torch.load(cfgs.path_to_flownet, map_location='cpu')
self.model.event_flownet.load_state_dict(checkpoint['state_dict'], strict=True)
print('Load path_to_flownet: {}'.format(cfgs.path_to_flownet))
else:
assert cfgs.path_to_flownet, "Should load pretrained DCEIFlow/ERAFT (GT I)"
if not cfgs.distributed:
self.model = self.model.to(device)
self.model.train()
print(self.model)
# Load training data
path_to_train_data = cfgs.path_to_train_data
train_data = TrainfusedEventData(os.path.join(path_to_train_data, 'train_e2v_estflow.txt'), cfgs)
self.train_loader = data.DataLoader(train_data,batch_size=cfgs.batch_size, shuffle=cfgs.shuffle, num_workers=4) # num_workers=4-----------
lr = cfgs.lr*(0.9**np.floor(cfgs.load_epoch_for_train/10.))
self.optimizer = optim.Adam(self.model.parameters(),lr=lr)
self.scheduler = optim.lr_scheduler.StepLR(self.optimizer, step_size=10, gamma=0.9)
# Save training results
self.is_SummaryWriter = cfgs.is_SummaryWriter
if self.is_SummaryWriter:
self.writer = SummaryWriter('./summary/{}'\
.format(self.model_name))
def run_train(self, cfgs):
'''
1. [0, flow_epoch], train DCEIFlow/ERAFT (Rec I), fix CISTA (GT Flow)
2. [flow_epoch, flow_epoch + rec_epoch], train CISTA (Pred Flow), fix DCEIFlow/ERAFT (Rec I)
3. [flow_epoch + rec_epoch, epoch], train CISTA-Flow without GT data iteratively
'''
for epoch in range(cfgs.load_epoch_for_train, cfgs.epochs):
lr = self.scheduler.get_last_lr()[0]
if epoch <cfgs.flow_epoch:
self.model.fix_params(net_name='rec')
train_rec = False
elif epoch >=cfgs.flow_epoch and epoch < cfgs.flow_epoch + cfgs.rec_epoch:
self.model.fix_params(net_name='flow')
train_rec = True
else:
self.optimizer.param_groups[0]['lr'] = 3e-5
if (epoch-cfgs.flow_epoch - cfgs.rec_epoch)%4 >=2:
self.model.fix_params(net_name='flow')
train_rec = True
else:
self.model.fix_params(net_name='rec')
train_rec = False
print('lr:', self.optimizer.param_groups[0]['lr'])
print('train_rec: ', train_rec)
self.train_epoch(epoch, cfgs, train_rec, 'cuda:0' if cfgs.distributed else self.device)
self.scheduler.step()
if epoch == 0 or (epoch+1)==cfgs.flow_epoch+cfgs.rec_epoch or ((epoch+1)>=cfgs.flow_epoch + cfgs.rec_epoch and (epoch+1-cfgs.flow_epoch-cfgs.rec_epoch)%2 == 0) or (epoch+1) % 10 == 0: # + cfgs.rec_epoch or (epoch+1) % 10 == 0:
torch.save({'epoch': epoch, 'state_dict': self.model.state_dict()},
os.path.join(self.path_to_model, '{}_{}.pth.tar'\
.format(self.model_name, epoch+1)))
def run_train_distributed(self, cfgs):
'''
1. [0, flow_epoch], train DCEIFlow/ERAFT (Rec I), fix CISTA (GT Flow)
2. [flow_epoch, flow_epoch + rec_epoch], train CISTA (Pred Flow), fix DCEIFlow/ERAFT (Rec I)
3. [flow_epoch + rec_epoch, epoch], train CISTA-Flow without GT data iteratively
'''
for epoch in range(cfgs.load_epoch_for_train, cfgs.epochs):
lr = self.scheduler.get_last_lr()[0]
if epoch <cfgs.flow_epoch:
self.model.module.fix_params(net_name='rec')
train_rec = False
elif epoch >=cfgs.flow_epoch and epoch < cfgs.flow_epoch + cfgs.rec_epoch:
self.model.module.fix_params(net_name='flow')
train_rec = True
else:
self.optimizer.param_groups[0]['lr'] = 3e-5
if (epoch-cfgs.flow_epoch - cfgs.rec_epoch)%4 >=2:
self.model.module.fix_params(net_name='flow')
train_rec = True
else:
self.model.module.fix_params(net_name='rec')
train_rec = False
print('lr:', self.optimizer.param_groups[0]['lr'])
print('train_rec: ', train_rec)
self.train_epoch(epoch, cfgs, train_rec, 'cuda:0' if cfgs.distributed else self.device)
self.scheduler.step()
if epoch == 0 or (epoch+1)==cfgs.flow_epoch+cfgs.rec_epoch or ((epoch+1)>=cfgs.flow_epoch + cfgs.rec_epoch and (epoch+1-cfgs.flow_epoch-cfgs.rec_epoch)%2 == 0) or (epoch+1) % 10 == 0: # + cfgs.rec_epoch or (epoch+1) % 10 == 0:
torch.save({'epoch': epoch, 'state_dict': self.model.module.state_dict()},
os.path.join(self.path_to_model, '{}_{}.pth.tar'\
.format(self.model_name, epoch+1)))
def train_epoch(self, epoch, cfgs, train_rec=False, device='cuda:0'):
torch.cuda.empty_cache()
batch_num =len(self.train_loader)
loss = 0
states = None
output = None
for batch_idx, seq_data in enumerate(self.train_loader):
loss = 0
cur_gt = dict([])
for s in range(len(seq_data)):
cur_data = {key: value.to(device) for key, value in seq_data[s][0].items()}
cur_target = {key: value.to(device) for key, value in seq_data[s][1].items()}
if s == 0:
cur_data['rec_img0'] = torch.zeros_like(cur_target['gt_img1'])
states = None
else:
cur_data['rec_img0'] = output.clone()
cur_gt['gt_img1'] = cur_target['gt_img1'].clone()
# when training DCEIFlow (Rec I), provide gt_flow for CISTA (GT Flow)
if epoch < cfgs.flow_epoch:
cur_gt['gt_flow'] = cur_target['gt_flow'].clone()
output, batch_flow, states = self.model(cur_data, states, cur_gt)
if train_rec:
loss_mode = 'rec'
is_loss_consis = True if s >=2 else False
else:
loss_mode = 'flow'
is_loss_consis = False
if epoch >= cfgs.flow_epoch+cfgs.rec_epoch:
loss_mode = 'both'
# Print memory usage
# print(s, f"Memory allocated: {torch.cuda.memory_allocated() / 1e9:.2f} GB")
# print(s, f"Max memory allocated: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB")
loss += self.loss_fn(output, cur_data['rec_img0'], batch_flow, cur_target, loss_mode, is_loss_consis=is_loss_consis)
if self.is_SummaryWriter:
self.writer.add_scalar('loss', loss, batch_num*epoch+batch_idx)
self.optimizer.zero_grad()
loss.backward(retain_graph=False)
self.optimizer.step()
if batch_idx%50 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tloss: {:.6f}'.format(\
epoch+1, batch_idx*self.train_loader.batch_size, len(self.train_loader.dataset),\
100.*batch_idx/len(self.train_loader), loss.data)) # .data.cpu().numpy()
if __name__ == '__main__':
# seed = 1234
# torch.manual_seed(seed)
# torch.cuda.manual_seed(seed)
# torch.cuda.manual_seed_all(seed)
# np.random.seed(seed)
# random.seed(seed)
setup_seed(1234)
## config parameters
parser = argparse.ArgumentParser(
description='Training options')
set_configs(parser)
cfgs = parser.parse_args()
cfgs.shuffle = True
if cfgs.distributed:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
model_train = Train(cfgs, device)
if cfgs.distributed:
model_train.run_train_distributed(cfgs)
else:
model_train.run_train(cfgs)