-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
223 lines (188 loc) · 9.62 KB
/
trainer.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
import time
import math
import numpy as np
import torch
from utils import AverageMeter, load_datasets
class PanoSeq2SeqTrainer():
"""Trainer for training and validation process"""
def __init__(self, opts, agent, optimizer, train_iters_epoch=100):
self.opts = opts
self.agent = agent
self.optimizer = optimizer
self.train_iters_epoch = train_iters_epoch
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def train(self, epoch, train_env, tb_logger=None):
batch_time = AverageMeter()
losses = AverageMeter()
dists = AverageMeter()
movements = AverageMeter()
val_losses = AverageMeter()
val_acces = AverageMeter()
print('Training on {} env ...'.format(train_env.splits[0]))
# switch to train mode
self.agent.env = train_env
self.agent.encoder.train()
self.agent.model.train()
if self.opts.second_training:
self.agent.model.first_stage_model.training = False
self.agent.feedback = self.opts.feedback_training
self.agent.value_loss = None
self.agent.val_acc = None
# load dataset path for computing ground truth distance
self.agent.gt = {}
for item in load_datasets(train_env.splits, self.opts):
self.agent.gt[item['path_id']] = item
end = time.time()
for iter in range(1, self.train_iters_epoch + 1):
# rollout the agent
if self.opts.arch == 'self-monitoring':
loss, traj = self.agent.rollout_monitor()
elif self.opts.arch == 'speaker-baseline':
loss, traj = self.agent.rollout()
else:
raise NotImplementedError()
dist_from_goal = np.mean(self.agent.dist_from_goal)
movement = np.mean(self.agent.traj_length)
losses.update(loss.item(), self.opts.batch_size)
dists.update(dist_from_goal, self.opts.batch_size)
movements.update(movement, self.opts.batch_size)
if self.agent.value_loss is not None:
val_losses.update(self.agent.value_loss.item(), self.opts.batch_size)
if self.agent.val_acc is not None:
val_acces.update(np.mean(self.agent.val_acc), self.opts.batch_size)
# zero the gradients before backward pass
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if tb_logger and iter % 10 == 0:
current_iter = iter + (epoch - 1) * self.train_iters_epoch
tb_logger.add_scalar('train/loss_train', loss, current_iter)
tb_logger.add_scalar('train/dist_from_goal', dist_from_goal, current_iter)
tb_logger.add_scalar('train/movements', movement, current_iter)
if self.agent.value_loss is not None:
tb_logger.add_scalar('train/value_loss', self.agent.value_loss, current_iter)
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
epoch, iter, self.train_iters_epoch, batch_time=batch_time,
loss=losses))
if tb_logger:
tb_logger.add_scalar('epoch/learning_rate', self.optimizer.param_groups[0]['lr'], epoch)
tb_logger.add_scalar('epoch/train/loss', losses.avg, epoch)
tb_logger.add_scalar('epoch/train/dist_from_goal', dists.avg, epoch)
tb_logger.add_scalar('epoch/train/movements', movements.avg, epoch)
if self.agent.value_loss is not None:
tb_logger.add_scalar('epoch/train/val_loss', val_losses.avg, epoch)
if self.agent.val_acc is not None:
tb_logger.add_scalar('epoch/train/val_acc', val_acces.avg, epoch)
def eval(self, epoch, val_env, tb_logger=None):
batch_time = AverageMeter()
losses = AverageMeter()
dists = AverageMeter()
movements = AverageMeter()
val_losses = AverageMeter()
val_acces = AverageMeter()
env_name, (env, evaluator) = val_env
print('Evaluating on {} env ...'.format(env_name))
self.agent.env = env
self.agent.env.reset_epoch()
self.agent.model.eval()
self.agent.encoder.eval()
self.agent.feedback = self.opts.feedback
self.agent.value_loss = None
self.agent.val_acc = None
# load dataset path for computing ground truth distance
self.agent.gt = {}
for item in load_datasets([env_name]):
self.agent.gt[item['path_id']] = item
val_iters_epoch = math.ceil(len(env.data) / self.opts.batch_size)
self.agent.results = {}
looped = False
iter = 1
with torch.no_grad():
end = time.time()
while True:
if self.opts.progress_inference:
traj = self.agent.sample_progress_inference(self.opts.beam_size)
elif self.opts.eval_beam:
traj = self.agent.sample_beam(self.opts.beam_size)
else:
# rollout the agent
if self.opts.arch == 'self-monitoring':
loss, traj = self.agent.rollout_monitor()
elif self.opts.arch == 'speaker-baseline':
loss, traj = self.agent.rollout()
else:
raise NotImplementedError()
dist_from_goal = np.mean(self.agent.dist_from_goal)
movement = np.mean(self.agent.traj_length)
losses.update(loss.item(), self.opts.batch_size)
dists.update(dist_from_goal, self.opts.batch_size)
movements.update(movement, self.opts.batch_size)
if self.agent.value_loss is not None:
val_losses.update(self.agent.value_loss.item(), self.opts.batch_size)
if self.agent.val_acc is not None:
val_acces.update(np.mean(self.agent.val_acc), self.opts.batch_size)
if tb_logger and iter % 10 == 0:
current_iter = iter + (epoch - 1) * val_iters_epoch
tb_logger.add_scalar('{}/loss'.format(env_name), loss, current_iter)
tb_logger.add_scalar('{}/dist_from_goal'.format(env_name), dist_from_goal, current_iter)
tb_logger.add_scalar('{}/movements'.format(env_name), movement, current_iter)
if self.agent.value_loss is not None:
tb_logger.add_scalar('{}/val_loss'.format(env_name), self.agent.value_loss, current_iter)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'.format(
epoch, iter, val_iters_epoch, batch_time=batch_time,
loss=losses))
# write into results
for traj_ in traj:
if traj_['instr_id'] in self.agent.results:
looped = True
else:
result = {
'path': traj_['path'],
'distance': traj_['distance'],
'img_attn': traj_['img_attn'],
'ctx_attn': traj_['ctx_attn'],
'value': traj_['value'],
'viewpoint_idx': traj_['viewpoint_idx'],
'navigable_idx': traj_['navigable_idx']
}
self.agent.results[traj_['instr_id']] = result
if looped:
break
iter += 1
if tb_logger:
tb_logger.add_scalar('epoch/{}/loss'.format(env_name), losses.avg, epoch)
tb_logger.add_scalar('epoch/{}/dist_from_goal'.format(env_name), dists.avg, epoch)
tb_logger.add_scalar('epoch/{}/movements'.format(env_name), movements.avg, epoch)
if self.agent.value_loss is not None:
tb_logger.add_scalar('epoch/{}/val_loss'.format(env_name), val_losses.avg, epoch)
if self.agent.val_acc is not None:
tb_logger.add_scalar('epoch/{}/val_acc'.format(env_name), val_acces.avg, epoch)
# dump into JSON file
if self.opts.eval_beam:
self.agent.results_path = '{}{}-beam_{}_{}_epoch_{}.json'.format(self.opts.results_dir, self.opts.exp_name,
self.opts.beam_size, env_name, epoch)
else:
self.agent.results_path = '{}{}_{}_epoch_{}.json'.format(self.opts.results_dir, self.opts.exp_name,
env_name, epoch)
self.agent.write_results()
score_summary, _ = evaluator.score(self.agent.results_path)
result_str = ''
success_rate = 0.0
for metric, val in score_summary.items():
result_str += '| {}: {} '.format(metric, val)
if metric in ['success_rate']:
success_rate = val
if tb_logger:
tb_logger.add_scalar('score/{}/{}'.format(env_name, metric), val, epoch)
print(result_str)
return success_rate