-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathground_relation.py
270 lines (193 loc) · 10.2 KB
/
ground_relation.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
# ====================================================
# @Author : Xiao Junbin
# @Email : [email protected]
# @File : ground_relation.py
# ====================================================
# from networks.basic import *
from networks.relation2relation import *
from utils import *
from torch.optim.lr_scheduler import ReduceLROnPlateau
import os.path as osp
import time
from torch.nn.utils import clip_grad_value_
class GroundRelation():
def __init__(self, vocab, train_loader, val_loader, checkpoint_path, model_prefix, vis_step, save_step, visual_dim, lr, batch_size, epoch_num, cuda):
self.vocab = vocab
self.train_loader = train_loader
self.val_loader = val_loader
self.model_dir = checkpoint_path
self.model_name = model_prefix
self.vis_step = vis_step
self.save_step = save_step
self.lr = lr
self.grad_clip = 10
self.batch_size = batch_size
self.epoch_num = epoch_num
self.cuda = cuda
self.input_size = 512
self.hidden_size = 512
self.visual_dim = visual_dim
self.word_dim = 300
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu") # sets device for model and PyTorch tensors
def build_model(self):
self.relation_ground = AttHierarchicalGround(self.input_size, self.hidden_size, self.visual_dim, self.word_dim)
self.relation_reconstruction = DecoderRNN(self.input_size, self.hidden_size, len(self.vocab), 1, 10)
params = [{'params':self.relation_reconstruction.parameters()}, {'params':self.relation_ground.parameters()}]
self.optimizer = torch.optim.Adam(params=params,
lr=self.lr)
self.scheduler = ReduceLROnPlateau(self.optimizer, 'min', factor=0.1, patience=2, verbose=True)
self.relation_ground.to(self.device)
self.relation_reconstruction.to(self.device)
self.criterion = nn.CrossEntropyLoss().to(self.device)
def save_model(self, epoch):
torch.save(self.relation_reconstruction.state_dict(),
osp.join(self.model_dir, '{}-reconstruct-{}.ckpt'.format(self.model_name, epoch + 1)))
torch.save(self.relation_ground.state_dict(),
osp.join(self.model_dir, '{}-ground-{}.ckpt'.format(self.model_name, epoch + 1)))
def resume(self):
"""
Initialize with the basic model obtained at the 1st stage.
:param epoch:
:return:
"""
ground_model_file = osp.join(self.model_dir,'vRGV-basic-ground-3.ckpt')
reconstruct_model_file = osp.join(self.model_dir,'vRGV-basic-reconstruct-3.ckpt')
ground_dict = torch.load(ground_model_file)
reconstruct_dict = torch.load(reconstruct_model_file)
# new_ground_dict = {}
# for k, v in self.relation_ground.state_dict().items():
# if k in ground_dict:
# v = ground_dict[k]
# new_ground_dict[k] = v
# self.relation_ground.load_state_dict(new_ground_dict) #only reconstruction part is better
new_reconstruct_dict = {}
for k, v in self.relation_reconstruction.state_dict().items():
if k in reconstruct_dict:
v = reconstruct_dict[k]
new_reconstruct_dict[k] = v
self.relation_reconstruction.load_state_dict(new_reconstruct_dict)
def run(self, pretrain=False):
self.build_model()
if pretrain:
self.resume()
save_loss = np.inf
for epoch in range(0, self.epoch_num):
train_loss = self.train(epoch)
val_loss = self.val(epoch)
print('==> Epoch:[{}/{}] [Training loss: {:.4f} {:.4f} Val loss: {:.4f} {:.4f}]'.
format(epoch, self.epoch_num, train_loss, np.exp(train_loss), val_loss, np.exp(val_loss)))
self.scheduler.step(val_loss)
if val_loss < save_loss:
save_loss = val_loss
self.save_model(epoch)
def train(self, epoch):
print('==> Epoch:[{}/{}] [training stage Encode_lr: {} Decode_lr: {}]'.
format(epoch, self.epoch_num, self.optimizer.param_groups[1]['lr'], self.optimizer.param_groups[0]['lr']))
self.relation_ground.train()
self.relation_reconstruction.train()
total_step = len(self.train_loader)
epoch_loss = 0
for iter, (relation_text, videos, relations, valid_lengths, video_names) in enumerate(self.train_loader):
videos = videos.to(self.device)
relations = relations.to(self.device)
targets = pack_padded_sequence(relations, valid_lengths, batch_first=True)[0]
video_out, video_hidden = self.relation_ground(videos, relation_text)
relation_decode = self.relation_reconstruction(video_out, video_hidden, relations, valid_lengths)
loss = self.criterion(relation_decode , targets)
self.relation_ground.zero_grad()
self.relation_reconstruction.zero_grad()
loss.backward()
# clip_gradient(self.optimizer, self.grad_clip)
# clip_grad_value_()
self.optimizer.step()
cur_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
if iter % self.vis_step == 0:
print(' [{}/{}]-{}-{:.4f}-{:5.4f}'.
format(iter, total_step, cur_time, loss.item(), np.exp(loss.item())))
epoch_loss += loss.item()
return epoch_loss / total_step
def val(self, epoch):
print('==> Epoch:[{}/{}][validation stage]'.format(epoch, self.epoch_num))
self.relation_ground.eval()
self.relation_reconstruction.eval()
total_step = len(self.val_loader)
epoch_loss = 0
with torch.no_grad():
for iter, (relation_text, videos, relations, valid_lengths, video_names) in enumerate(self.val_loader):
videos = videos.to(self.device)
relations = relations.to(self.device)
targets = pack_padded_sequence(relations, valid_lengths, batch_first=True)[0]
video_out, video_hidden = self.relation_ground(videos, relation_text)
relation_decode = self.relation_reconstruction(video_out, video_hidden, relations, valid_lengths)
loss = self.criterion(relation_decode, targets)
cur_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
if iter % self.vis_step == 0:
print(' [{}/{}]-{}-{:.4f}-{:5.4f}'.
format(iter, total_step, cur_time, loss.item(), np.exp(loss.item())))
epoch_loss += loss.item()
return epoch_loss / total_step
def predict(self, ep):
self.build_model()
ground_model_path = osp.join(self.model_dir, '{}-ground-{}.ckpt'.format(self.model_name, ep))
reconstruction_path = osp.join(self.model_dir, '{}-reconstruct-{}.ckpt'.format(self.model_name, ep))
self.relation_reconstruction.eval()
self.relation_ground.eval()
self.relation_ground.load_state_dict(torch.load(ground_model_path))
self.relation_reconstruction.load_state_dict(torch.load(reconstruction_path))
total = len(self.val_loader)
pos_num = 0
fout = open('results/prediction.txt', 'w')
with torch.no_grad():
for iter, (relation_text, videos, relations, valid_lengths, video_names) in enumerate(self.val_loader):
videos = videos.to(self.device)
video_out, video_hidden = self.relation_ground(videos, relation_text)
sample_ids = self.relation_reconstruction.sample(video_out, video_hidden)
sample_ids = sample_ids[0].cpu().numpy()
predict_relation = []
for id in sample_ids:
word = self.vocab.idx2word[id]
predict_relation.append(word)
if word == '<end>': break
predict_relation = ' '.join(predict_relation[1:-1])
# print(relation_text[0], predict_relation)
table = str.maketrans('-_', ' ')
relation = relation_text[0].translate(table)
output_str = "{}:{}".format(relation, predict_relation)
fout.writelines(output_str+'\n')
if relation == predict_relation:
pos_num += 1
if iter%self.vis_step == 0:
print("{}:{}".format(iter, output_str))
print("Reconstrution Rate: ", pos_num / total)
fout.close()
def ground_attention(self, ep, save_name):
"""output the spatial temporal attention as grounding results"""
self.build_model()
ground_model_path = osp.join(self.model_dir, '{}-ground-{}.ckpt'.format(self.model_name, ep))
self.relation_ground.eval()
self.relation_ground.load_state_dict(torch.load(ground_model_path))
video_res = {}
total = len(self.val_loader)
with torch.no_grad():
for iter, (relation_text, videos, relations, valid_lengths, video_names) in enumerate(self.val_loader):
videos = videos.to(self.device)
sub_atts, obj_atts, beta1, beta2 = self.relation_ground(videos, relation_text, mode='val')
# print(sub_atts.shape, beta1.shape) #(32, 120, 40), (32, 120)
data_sub_atts = sub_atts.data.cpu().numpy()
data_obj_atts = obj_atts.data.cpu().numpy()
data_beta2 = beta2.data.cpu().numpy()
data_beta1 = beta1.data.cpu().numpy()
real_bs = data_sub_atts.shape[0]
for bs in range(real_bs):
data = {}
data['sub'] = data_sub_atts[bs].tolist()
data['obj'] = data_obj_atts[bs].tolist()
data['beta2'] = data_beta2[bs].tolist()
data['beta1'] = data_beta1[bs].tolist()
vname = video_names[bs]
if vname not in video_res: video_res[vname] = {}
video_res[vname][relation_text[bs]] = data
if iter % self.vis_step == 0:
print('Finished: {}-{}'.format(iter, total))
save_results(save_name, video_res)