-
Notifications
You must be signed in to change notification settings - Fork 17
/
train_rl.py
executable file
·399 lines (334 loc) · 14.7 KB
/
train_rl.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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
from __future__ import print_function
import argparse
import torch
import torch.optim as optim
from torch.autograd import Variable
import math
import ranker
import random
import time
import sys
import sim_user
from model import NetSynUser
import torch.nn as nn
parser = argparse.ArgumentParser(description='PyTorch Example')
parser.add_argument('--batch-size', type=int, default=8,
help='input batch size for training')
parser.add_argument('--test-batch-size', type=int, default=128,
help='input batch size for testing')
parser.add_argument('--epochs', type=int, default=15,
help='number of epochs to train')
# learning
parser.add_argument('--lr', type=float, default=0.0001,
help='learning rate')
parser.add_argument('--tau', type=float, default=1,
help='softmax temperature')
parser.add_argument('--seed', type=int, default=7771,
help='random seed')
parser.add_argument('--log-interval', type=int, default=10,
help='how many batches to wait before logging training status')
parser.add_argument('--neg-num', type=int, default=5,
help='number of negative candidates in the denominator')
parser.add_argument('--model-folder', type=str, default="models/",
help='triplet loss margin ')
parser.add_argument('--top-k', type=int, default=4,
help='top k candidate for policy and nearest neighbors')
parser.add_argument('--pretrained-model', type=str, default="models/sl-12.pt",
help='path to pretrained sl model')
parser.add_argument('--triplet-margin', type=float, default=0.1, metavar='EV',
help='triplet loss margin ')
# exp. control
parser.add_argument('--train-turns', type=int, default=5,
help='dialog turns for training')
parser.add_argument('--test-turns', type=int, default=5,
help='dialog turns for testing')
args = parser.parse_args()
class TripletLossIP(nn.Module):
def __init__(self, margin):
super(TripletLossIP, self).__init__()
self.margin = margin
def forward(self, anchor, positive, negative, average=True):
dist = torch.sum(
(anchor - positive) ** 2 - (anchor - negative) ** 2 ,
dim=1) + self.margin
dist_hinge = torch.clamp(dist, min=0.0)
if average:
return torch.mean(dist_hinge)
else:
return dist_hinge
# experiment monitor
class ExpMonitor():
def __init__(self, train_mode):
self.train_mode = train_mode
if train_mode:
num_turns = args.train_turns
num_act = user.train_fc_input.size(0)
else:
num_turns = args.test_turns
num_act = user.test_fc_input.size(0)
self.loss = torch.Tensor(num_turns).zero_()
self.all_loss = torch.Tensor(num_turns).zero_()
self.rank = torch.Tensor(num_turns).zero_()
self.all_rank = torch.Tensor(num_turns).zero_()
self.count = 0.0
self.all_count = 0.0
self.start_time = time.time()
self.pos_idx = torch.Tensor(num_act).zero_()
self.act_idx = torch.Tensor(num_act).zero_()
return
def log_step(self, ranking, loss, user_img_idx, act_img_idx, k):
tmp_rank = ranking.float().mean()
self.rank[k] += tmp_rank
self.all_rank[k] += tmp_rank
self.loss[k] += loss[0]
self.all_loss[k] += loss[0]
for i in range(user_img_idx.size(0)):
self.pos_idx[user_img_idx[i]] += 1
self.act_idx[act_img_idx[i]] += 1
self.count += 1
self.all_count += 1
return
def print_interval(self, epoch, batch_idx, num_epoch):
if self.train_mode:
output_string = 'Train Epoch:'
num_input = user.train_fc_input.size(0)
else:
output_string = 'Eval Epoch:'
num_input = user.test_fc_input.size(0)
output_string += '{} [{}/{} ({:.0f}%)]\tTime:{:.2f}\tNumAct:{}\n'.format(
epoch, batch_idx, num_epoch, 100. * batch_idx / num_epoch, time.time() - self.start_time, self.pos_idx.sum()
)
output_string += 'pos:({:.0f}, {:.0f}) \tact:({:.0f}, {:.0f})\n'.format(
self.pos_idx.max(), self.pos_idx.min(), self.act_idx.max(), self.act_idx.min()
)
if self.train_mode:
dialog_turns = args.train_turns
else:
dialog_turns = args.test_turns
self.rank.mul_(dialog_turns / self.count)
self.loss.mul_(1.0 / self.count)
output_string += 'rank:'
for i in range(dialog_turns):
output_string += '{:.4f}\t '.format(self.rank[i] / num_input)
output_string += '\nloss:'
for i in range(dialog_turns):
output_string += '{:.4f}\t '.format(self.loss[i])
print(output_string)
self.loss.zero_()
self.rank.zero_()
self.count = 0.0
sys.stdout.flush()
return
def print_all(self, epoch):
if self.train_mode:
num_input = user.train_fc_input.size(0)
else:
num_input = user.test_fc_input.size(0)
if self.train_mode:
dialog_turns = args.train_turns
else:
dialog_turns = args.test_turns
self.all_rank.mul_(dialog_turns / self.all_count)
self.all_loss.mul_(1.0 / self.all_count)
output_string = '{} #rank:'.format(epoch)
for i in range(dialog_turns):
output_string += '{:.4f}\t '.format(self.all_rank[i] / num_input)
output_string += '\n{} #loss:'.format(epoch)
for i in range(dialog_turns):
output_string += '{:.4f}\t '.format(self.all_loss[i])
print(output_string)
self.all_loss.zero_()
self.all_rank.zero_()
self.all_count = 0.0
self.loss.zero_()
self.rank.zero_()
self.count = 0.0
sys.stdout.flush()
return
random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
user = sim_user.SynUser()
ranker = ranker.Ranker()
behavior_model = NetSynUser(user.vocabSize + 1)
target_model = NetSynUser(user.vocabSize + 1)
triplet_loss = TripletLossIP(margin=args.triplet_margin)
# load pre-trained model
behavior_model.load_state_dict(torch.load(args.pretrained_model, map_location=lambda storage, loc: storage))
# load pre-trained model
target_model.load_state_dict(torch.load(args.pretrained_model, map_location=lambda storage, loc: storage))
if torch.cuda.is_available():
behavior_model.cuda()
target_model.cuda()
triplet_loss.cuda()
def rollout_search(behavior_state, target_state, cur_turn, max_turn, user_img_idx, all_input):
# 1. compute the top-k nearest neighbor for current state
top_k_act_img_idx = ranker.k_nearest_neighbors(target_state.data, K=args.top_k)
# 2. rollout for each candidate in top k
target_hx_bk = target_model.hx
rollout_values = []
for i in range(args.top_k):
target_model.init_hid(args.batch_size)
if torch.cuda.is_available():
target_model.hx = target_model.hx.cuda()
target_model.hx.data.copy_(target_hx_bk.data)
act_img_idx = top_k_act_img_idx[:, i]
score = 0
for j in range(max_turn - cur_turn):
txt_input = user.get_feedback(act_idx=act_img_idx, user_idx=user_img_idx, train_mode=True)
if torch.cuda.is_available():
txt_input = txt_input.cuda()
txt_input = Variable(txt_input, volatile=True)
if torch.cuda.is_available():
act_img_idx = act_img_idx.cuda()
act_emb = ranker.feat[act_img_idx]
action = target_model.merge_forward(Variable(act_emb, volatile=True), txt_input)
act_img_idx = ranker.nearest_neighbor(action.data)
ranking_candidate = ranker.compute_rank(action.data, user_img_idx)
score = score + ranking_candidate
rollout_values.append(score)
rollout_values = torch.stack(rollout_values, dim=1)
# compute greedy actions
_, greedy_idx = rollout_values.min(dim=1)
# recover target_state
target_model.hx = target_hx_bk
if torch.cuda.is_available():
greedy_idx = greedy_idx.cuda()
act_opt = torch.gather(top_k_act_img_idx, 1, greedy_idx.cpu().unsqueeze(1)).view(-1)
# 3. compute loss
# compute the log prob for candidates
dist_action = []
act_input = all_input[act_opt]
if torch.cuda.is_available():
act_input = act_input.cuda()
act_emb = behavior_model.forward_image(Variable(act_input))
dist = -torch.sum((behavior_state - act_emb) ** 2, dim=1) / args.tau
dist_action.append(dist)
for i in range(args.neg_num):
neg_img_idx = torch.LongTensor(args.batch_size)
user.sample_idx(neg_img_idx, train_mode=True)
neg_input = all_input[neg_img_idx]
if torch.cuda.is_available():
neg_input = neg_input.cuda()
neg_emb = behavior_model.forward_image(Variable(neg_input))
dist = -torch.sum((behavior_state - neg_emb) ** 2, dim=1) / args.tau
dist_action.append(dist)
dist_action = torch.stack(dist_action, dim=1)
label_idx = torch.LongTensor(args.batch_size).fill_(0)
if torch.cuda.is_available():
label_idx = label_idx.cuda()
loss = torch.nn.functional.cross_entropy(input=dist_action, target=Variable(label_idx))
# compute the reg following the pre-training loss
if torch.cuda.is_available():
user_img_idx = user_img_idx.cuda()
target_emb = ranker.feat[user_img_idx]
reg = torch.sum((behavior_state - Variable(target_emb)) ** 2, dim=1).mean()
return act_opt, reg + loss
user_img_idx_ = torch.LongTensor(args.batch_size)
act_img_idx_ = torch.LongTensor(args.batch_size)
user.sample_idx(user_img_idx_, train_mode=True)
user.sample_idx(act_img_idx_, train_mode=True)
def train_rl(epoch, optimizer):
behavior_model.set_rl_mode()
target_model.eval()
triplet_loss.train()
exp_monitor_candidate = ExpMonitor(train_mode=True)
# train / test
all_input = user.train_feature
dialog_turns = args.train_turns
#
user_img_idx = torch.LongTensor(args.batch_size)
act_img_idx = torch.LongTensor(args.batch_size)
# update ranker
ranker.update_rep(target_model, all_input)
num_epoch = math.ceil(all_input.size(0) / args.batch_size)
for batch_idx in range(1, num_epoch + 1):
# sample data index
user.sample_idx(user_img_idx, train_mode=True)
user.sample_idx(act_img_idx, train_mode=True)
target_model.init_hid(args.batch_size)
behavior_model.init_hid(args.batch_size)
if torch.cuda.is_available():
target_model.hx = target_model.hx.cuda()
behavior_model.hx = behavior_model.hx.cuda()
loss_sum = 0
for k in range(dialog_turns):
# construct data
txt_input = user.get_feedback(act_idx=act_img_idx.cpu(), user_idx=user_img_idx.cpu(), train_mode=True)
if torch.cuda.is_available():
txt_input = txt_input.cuda()
# update model part
if torch.cuda.is_available():
act_img_idx = act_img_idx.cuda()
act_emb = ranker.feat[act_img_idx]
behavior_state = behavior_model.merge_forward(Variable(act_emb), Variable(txt_input))
# update base model part
target_state = target_model.merge_forward(Variable(act_emb, volatile=True),
Variable(txt_input, volatile=True))
ranking_candidate = ranker.compute_rank(behavior_state.data, user_img_idx)
act_img_idx_mc, loss = rollout_search(behavior_state, target_state, k, dialog_turns, user_img_idx, all_input)
loss_sum = loss + loss_sum
act_img_idx.copy_(act_img_idx_mc)
exp_monitor_candidate.log_step(ranking_candidate, loss.data, user_img_idx, act_img_idx, k)
optimizer.zero_grad()
loss_sum.backward()
optimizer.step()
if batch_idx % args.log_interval == 0:
print('# candidate ranking #')
exp_monitor_candidate.print_interval(epoch, batch_idx, num_epoch)
print('# candidate ranking #')
exp_monitor_candidate.print_all(epoch)
return
def eval(epoch):
# train_mode = True
print('eval epoch #{}'.format(epoch))
behavior_model.eval()
triplet_loss.eval()
train_mode = False
all_input = user.test_feature
dialog_turns = args.test_turns
exp_monitor_candidate = ExpMonitor(train_mode=train_mode)
user_img_idx = torch.LongTensor(args.batch_size)
act_img_idx = torch.LongTensor(args.batch_size)
neg_img_idx = torch.LongTensor(args.batch_size)
num_epoch = math.ceil(all_input.size(0) / args.batch_size)
ranker.update_rep(behavior_model, all_input)
for batch_idx in range(1, num_epoch + 1):
# sample data index
user.sample_idx(user_img_idx, train_mode=train_mode)
user.sample_idx(act_img_idx, train_mode=train_mode)
behavior_model.init_hid(args.batch_size)
if torch.cuda.is_available():
behavior_model.hx = behavior_model.hx.cuda()
if torch.cuda.is_available():
act_img_idx = act_img_idx.cuda()
act_emb = ranker.feat[act_img_idx]
for k in range(dialog_turns):
txt_input = user.get_feedback(act_idx=act_img_idx.cpu(), user_idx=user_img_idx.cpu(), train_mode=train_mode)
if torch.cuda.is_available():
txt_input = txt_input.cuda()
txt_input = Variable(txt_input, volatile=True)
action = behavior_model.merge_forward(Variable(act_emb, volatile=True), txt_input)
act_img_idx = ranker.nearest_neighbor(action.data)
user.sample_idx(neg_img_idx, train_mode=train_mode)
if torch.cuda.is_available():
user_img_idx = user_img_idx.cuda()
neg_img_idx = neg_img_idx.cuda()
act_img_idx = act_img_idx.cuda()
user_emb = ranker.feat[user_img_idx]
neg_emb = ranker.feat[neg_img_idx]
new_act_emb = ranker.feat[act_img_idx]
ranking_candidate = ranker.compute_rank(action.data, user_img_idx)
loss = triplet_loss.forward(action, Variable(user_emb), Variable(neg_emb))
act_emb = new_act_emb
# log
exp_monitor_candidate.log_step(ranking_candidate, loss.data, user_img_idx, act_img_idx, k)
exp_monitor_candidate.print_all(epoch)
return
optimizer = optim.Adam(behavior_model.parameters(), lr=args.lr, betas=(0.9, 0.999), eps=1e-8)
for i in range(20):
eval(i)
train_rl(i, optimizer)
torch.save(behavior_model.state_dict(), (args.model_folder+'rl-{}.pt').format(i))
eval(20)