-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathdo_segm.py
186 lines (150 loc) · 7.66 KB
/
do_segm.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
#
# SPDX-FileCopyrightText: 2021 Idiap Research Institute
#
# Written by Prabhu Teja <[email protected]>,
#
# SPDX-License-Identifier: MIT
import os
import shutil
import time
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from datasets import (CityscapesDataset, CrossCityDataset, get_test_transforms,
get_train_transforms)
from generate_pseudo_labels import validate_model
from network import DeeplabMulti as DeepLab
from network import JointSegAuxDecoderModel, NoisyDecoders
from utils import (ScoreUpdater, adjust_learning_rate, cleanup,
get_arguments, label_selection, parse_split_list,
savelst_tgt, seed_torch, self_training_regularized_infomax,
self_training_regularized_infomax_cct, set_logger)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
osp = os.path
args = get_arguments()
if not os.path.exists(args.save):
os.makedirs(args.save)
logger = set_logger(args.save, 'training_logger', False)
def make_network(args):
model = DeepLab(13, False)
model = torch.nn.DataParallel(model)
sd = torch.load('pretrained/Cityscapes_source_class13.pth', map_location=device)['state_dict']
model.load_state_dict(sd)
model = model.module
if args.unc_noise:
aux_decoders = NoisyDecoders(args.decoders, args.dropout)
model = JointSegAuxDecoderModel(model, aux_decoders)
return model
def test(model, round_idx):
transforms = get_test_transforms()
ds = CrossCityDataset(root=args.data_tgt_dir, list_path=args.data_tgt_test_list.format(args.city), transforms=transforms)
loader = torch.utils.data.DataLoader(ds, batch_size=6, pin_memory=torch.cuda.is_available(), num_workers=6)
scorer = ScoreUpdater(13, len(loader))
logger.info('###### Start evaluating in target domain test set in round {}! ######'.format(round_idx))
start_eval = time.time()
model.eval()
with torch.no_grad():
for batch in loader:
img, label, _ = batch
pred = model(img.to(device)).argmax(1).cpu()
scorer.update(pred.view(-1), label.view(-1))
model.train()
logger.info('###### Finish evaluating in target domain test set in round {}! Time cost: {:.2f} seconds. ######'.format(
round_idx, time.time()-start_eval))
def train(mix_trainloader, model, interp, optimizer, args):
"""Create the model and start the training."""
tot_iter = len(mix_trainloader)
for i_iter, batch in enumerate(mix_trainloader):
images, labels, name = batch
labels = labels.long()
optimizer.zero_grad()
adjust_learning_rate(optimizer, i_iter, tot_iter, args)
if args.info_max_loss:
pred = model(images.to(device), training=True)
loss = self_training_regularized_infomax(pred, labels.to(device), args)
elif args.unc_noise:
pred, noise_pred = model(images.to(device), training=True)
loss = self_training_regularized_infomax_cct(pred, labels.to(device), noise_pred, args)
else:
pred = model(images.to(device))
loss = F.cross_entropy(pred, labels.to(device), ignore_index=255)
loss.backward()
optimizer.step()
logger.info('iter = {} of {} completed, loss = {:.4f}'.format(i_iter+1, tot_iter, loss.item()))
def main():
seed_torch(args.randseed)
logger.info('Starting training with arguments')
logger.info(vars(args))
save_path = args.save
save_pseudo_label_path = osp.join(save_path, 'pseudo_label') # in 'save_path'. Save labelIDs, not trainIDs.
save_stats_path = osp.join(save_path, 'stats') # in 'save_path'
save_lst_path = osp.join(save_path, 'list')
if not os.path.exists(save_path):
os.makedirs(save_path)
if not os.path.exists(save_pseudo_label_path):
os.makedirs(save_pseudo_label_path)
if not os.path.exists(save_stats_path):
os.makedirs(save_stats_path)
if not os.path.exists(save_lst_path):
os.makedirs(save_lst_path)
tgt_portion = args.init_tgt_port
image_tgt_list, image_name_tgt_list, _, _ = parse_split_list(args.data_tgt_train_list.format(args.city))
model = make_network(args).to(device)
test(model, -1)
for round_idx in range(args.num_rounds):
save_round_eval_path = osp.join(args.save, str(round_idx))
save_pseudo_label_color_path = osp.join(
save_round_eval_path, 'pseudo_label_color') # in every 'save_round_eval_path'
if not os.path.exists(save_round_eval_path):
os.makedirs(save_round_eval_path)
if not os.path.exists(save_pseudo_label_color_path):
os.makedirs(save_pseudo_label_color_path)
src_portion = args.init_src_port
########## pseudo-label generation
conf_dict, pred_cls_num, save_prob_path, save_pred_path = validate_model(model,
save_round_eval_path,
round_idx, args)
cls_thresh = label_selection.kc_parameters(
conf_dict, pred_cls_num, tgt_portion, round_idx, save_stats_path, args)
label_selection.label_selection(cls_thresh, round_idx, save_prob_path, save_pred_path,
save_pseudo_label_path, save_pseudo_label_color_path, save_round_eval_path, args)
tgt_portion = min(tgt_portion + args.tgt_port_step, args.max_tgt_port)
tgt_train_lst = savelst_tgt(image_tgt_list, image_name_tgt_list, save_lst_path, save_pseudo_label_path)
rare_id = np.load(save_stats_path + '/rare_id_round' + str(round_idx) + '.npy')
mine_id = np.load(save_stats_path + '/mine_id_round' + str(round_idx) + '.npy')
# mine_chance = args.mine_chance
src_transforms, tgt_transforms = get_train_transforms(args, mine_id)
srcds = CityscapesDataset(transforms=src_transforms)
tgtds = CrossCityDataset(args.data_tgt_dir.format(args.city), tgt_train_lst,
pseudo_root=save_pseudo_label_path, transforms=tgt_transforms)
if args.no_src_data:
mixtrainset = tgtds
else:
mixtrainset = torch.utils.data.ConcatDataset([srcds, tgtds])
mix_loader = DataLoader(mixtrainset, batch_size=args.batch_size, shuffle=True,
num_workers=args.batch_size, pin_memory=torch.cuda.is_available())
src_portion = min(src_portion + args.src_port_step, args.max_src_port)
optimizer = optim.SGD(model.optim_parameters(args), lr=args.learning_rate,
momentum=args.momentum, weight_decay=args.weight_decay)
interp = nn.Upsample(size=args.input_size[::-1], mode='bilinear', align_corners=True)
torch.backends.cudnn.enabled = True # enable cudnn
torch.backends.cudnn.benchmark = True
start = time.time()
for epoch in range(args.epr):
train(mix_loader, model, interp, optimizer, args)
print('taking snapshot ...')
torch.save(model.state_dict(), osp.join(args.save,
'2nthy_round' + str(round_idx) + '_epoch' + str(epoch) + '.pth'))
end = time.time()
logger.info('###### Finish model retraining dataset in round {}! Time cost: {:.2f} seconds. ######'.format(
round_idx, end - start))
test(model, round_idx)
cleanup(args.save)
cleanup(args.save)
shutil.rmtree(save_pseudo_label_path)
test(model, args.num_rounds - 1)
if __name__ == "__main__":
main()