-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain.py
211 lines (171 loc) · 8.82 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
import os
import sys
import numpy as np
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
import torch.optim as optim
from tqdm import tqdm
from data_processing import possible_segments
import dataset
from model.model import MLLC_FullAtt
from engine import Engine
from metrics import AverageMeter
import eval
from config import args, device, path_prefix
torch.manual_seed(0)
torch.backends.cudnn.enabled = True
if __name__ == '__main__':
dataset_name = args.dataset_name
input_size_0 = 4096 if args.feature_type_0 == 'rgb' else 1024
input_size_1 = 4096 if args.feature_type_1 == 'rgb' else 1024
model = MLLC_FullAtt(input_size_0=input_size_0, input_size_1=input_size_1, txt_input_size=args.lang_hidden_size, hidden_size=args.hidden_size).to(device)
optimizer = optim.Adam(model.parameters(),lr=args.lr,weight_decay=1e-8)#weight decay to 5e-7
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=0.8, patience=20, verbose=True)
train_dataset = getattr(dataset, dataset_name)('train')
val_dataset = getattr(dataset, dataset_name)('val')
test_dataset = getattr(dataset, dataset_name)('test')
def iterator(split):
if split == 'train':
dataloader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=False,
collate_fn=train_dataset.collate_fn)
elif split == 'val':
dataloader = DataLoader(val_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=False,
collate_fn=val_dataset.collate_fn)
elif split == 'test':
dataloader = DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
pin_memory=False,
collate_fn=test_dataset.collate_fn)
else:
raise ValueError
return dataloader
def network(sample):
parse_tree = sample['batch_tree']
input_0 = sample['batch_{}_feats'.format(args.feature_type_0)].to(device)
input_1 = sample['batch_{}_feats'.format(args.feature_type_1)].to(device)
visual_mask = sample['batch_mask'].to(device)
batch_gt = sample['batch_gt']
batch_context_gt = sample['batch_context_gt']
output = model(parse_tree, input_0, input_1, visual_mask)
batch_size, prop_num, _ = output.shape
def ranking_loss(output, strong_supervised=False):
loss_mask = visual_mask.view(batch_size, -1)
main_output = torch.max(output,dim=2)[0]
gt_predicted_score = torch.gather(main_output, 1, torch.LongTensor(batch_gt)[:,None].expand(-1,prop_num).to(device))
loss = F.margin_ranking_loss(main_output.view(batch_size,-1),
gt_predicted_score.view(batch_size,-1),
-torch.ones(1).to(device),
margin=0.1, reduction='none')*loss_mask
ranking_loss = torch.sum(loss)/torch.sum(loss_mask)
if strong_supervised:
assert output.dim() == 3
gt_predicted_score = []
context_output = []
loss_mask = visual_mask.view(batch_size, -1)
for i, (gt,context_gt) in enumerate(zip(batch_gt,batch_context_gt)):
row = torch.stack([output[i, gt, context_gt] for _ in range(prop_num)])
context_output.append(output[i, gt])
gt_predicted_score.append(row)
gt_predicted_score = torch.stack(gt_predicted_score).to(device)
context_output = torch.stack(context_output).to(device)
loss = F.margin_ranking_loss(context_output.view(batch_size, -1),
gt_predicted_score.view(batch_size, -1),
-torch.ones(1).to(device),
margin=0.1, reduction='none') * loss_mask
ranking_loss += args.context_weight*torch.sum(loss)/torch.sum(loss_mask)
return ranking_loss
# loss_value = 1*ranking_loss(rgb_output)+1*ranking_loss(flow_output)#+1*ranking_loss(loc_output)
loss_value = ranking_loss(output,args.strong_supervised)
assert torch.sum(torch.isnan(output)).item() == 0
score_mask = visual_mask.expand(-1,-1,prop_num)*visual_mask.transpose(1,2).expand(-1,prop_num,-1)
scores = (torch.max((10000+output)*score_mask,dim=2)[0])
# TODO: fix zero and minus score problem
return loss_value, scores
def on_start(state):
state['loss_meter'] = AverageMeter()
state['test_interval'] = int(len(train_dataset)/args.batch_size*args.test_interval)
state['t'] = 1
model.train()
if args.verbose:
state['progress_bar'] = tqdm(total=state['test_interval'])
def on_forward(state):
torch.nn.utils.clip_grad_norm_(model.parameters(), 10)
state['loss_meter'].update(state['loss'].item(), 1)
def on_update(state):# Save All
if args.verbose:
state['progress_bar'].update(1)
if state['t'] % state['test_interval'] == 0:
model.eval()
if args.verbose:
state['progress_bar'].close()
val_state = engine.test(network, iterator('val'))
test_state = engine.test(network, iterator('test'))
state['scheduler'].step(-val_state['loss_meter'].avg)
saved_model_filename = path_prefix + 'checkpoint/{}/{}-{}-H{}/iter{:06d}-{:.4f}-{:.4f}.pkl'.format(
dataset_name, args.feature_type_0, args.feature_type_1, args.hidden_size,
state['t'], val_state['rank1'], val_state['miou'])
rootfolder1 = os.path.dirname(saved_model_filename)
rootfolder2 = os.path.dirname(rootfolder1)
rootfolder3 = os.path.dirname(rootfolder2)
if not os.path.exists(rootfolder3):
print('Make directory %s ...' % rootfolder3)
os.mkdir(rootfolder3)
if not os.path.exists(rootfolder2):
print('Make directory %s ...' % rootfolder2)
os.mkdir(rootfolder2)
if not os.path.exists(rootfolder1):
print('Make directory %s ...' % rootfolder1)
os.mkdir(rootfolder1)
torch.save(model.state_dict(), saved_model_filename)
print('iter: {} train loss {:.4f} '
'val loss {:.4f} rank@1: {:.4f} rank@5: {:.4f} miou: {:.4f} '
'test loss {:.4f} rank@1: {:.4f} rank@5: {:.4f} miou: {:.4f}'.format(
state['t'], state['loss_meter'].avg,
val_state['loss_meter'].avg, val_state['rank1'], val_state['rank5'], val_state['miou'],
test_state['loss_meter'].avg, test_state['rank1'], test_state['rank5'], test_state['miou'],))
sys.stdout.flush()
if args.verbose:
state['progress_bar'] = tqdm(total=state['test_interval'])
model.train()
val_state['loss_meter'].reset()
test_state['loss_meter'].reset()
state['loss_meter'].reset()
def on_end(state):
if args.verbose:
state['progress_bar'].close()
def on_test_start(state):
state['loss_meter'] = AverageMeter()
state['sorted_segments_list'] = []
def on_test_forward(state):
state['loss_meter'].update(state['loss'].item(), 1)
scores = state['output'].cpu().data.numpy().squeeze()
sorted_index = np.argsort(scores)[::-1]
sorted_segments = [possible_segments[i] for i in sorted_index]
state['sorted_segments_list'].append(sorted_segments)
def on_test_end(state):
data = state['iterator'].dataset.data
state['rank1'], state['rank5'], state['miou'] = eval.eval_predictions(state['sorted_segments_list'], data, verbose=False)
engine = Engine()
engine.hooks['on_start'] = on_start
engine.hooks['on_forward'] = on_forward
engine.hooks['on_update'] = on_update
engine.hooks['on_end'] = on_end
engine.hooks['on_test_start'] = on_test_start
engine.hooks['on_test_forward'] = on_test_forward
engine.hooks['on_test_end'] = on_test_end
engine.train(network,
iterator('train'),
maxepoch=args.max_epoch,
optimizer=optimizer,
scheduler=scheduler)