-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
386 lines (340 loc) · 20.2 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
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
# coding=utf-8
from __future__ import division, print_function, unicode_literals
import argparse
import json
import random
import datetime
from io import open
import os
import shutil
import numpy as np
import torch
from torch.optim import Adam
import torch.nn as nn
from utils import util, multiwoz_dataloader
from models.model import Model
from utils.util import detected_device, PAD_token, pp_mkdir
from multiwoz.Evaluators import *
# from tqdm import tqdm
# SOS_token = 0
# EOS_token = 1
# UNK_token = 2
# PAD_token = 3
# pp added: print out env
util.get_env_info()
all_start_time = datetime.datetime.now()
print('Start time={}'.format(all_start_time.strftime("%Y-%m-%d %H:%M:%S")))
parser = argparse.ArgumentParser(description='multiwoz1-bsl-tr')
# Group args
# 1. Data & Dirs
data_arg = parser.add_argument_group(title='Data')
data_arg.add_argument('--data_dir', type=str, default='data/multi-woz', help='the root directory of data')
data_arg.add_argument('--log_dir', type=str, default='logs')
data_arg.add_argument('--result_dir', type=str, default='results/bsl')
data_arg.add_argument('--pre_model_dir', type=str, default='results/moe4_gru-27062/model')
data_arg.add_argument('--model_name', type=str, default='translate.ckpt')
# 2.Network
net_arg = parser.add_argument_group(title='Network')
net_arg.add_argument('--cell_type', type=str, default='lstm')
net_arg.add_argument('--attention_type', type=str, default='bahdanau')
net_arg.add_argument('--depth', type=int, default=1, help='depth of rnn')
net_arg.add_argument('--emb_size', type=int, default=50)
net_arg.add_argument('--hid_size_enc', type=int, default=150)
net_arg.add_argument('--hid_size_dec', type=int, default=150)
net_arg.add_argument('--hid_size_pol', type=int, default=150)
net_arg.add_argument('--max_len', type=int, default=50)
net_arg.add_argument('--vocab_size', type=int, default=400, metavar='V')
net_arg.add_argument('--use_attn', type=util.str2bool, nargs='?', const=True, default=True) # F
net_arg.add_argument('--use_emb', type=util.str2bool, nargs='?', const=True, default=False)
# 3.Train
train_arg = parser.add_argument_group(title='Train')
train_arg.add_argument('--mode', type=str, default='train', help='training or testing: test, train, RL')
train_arg.add_argument('--optim', type=str, default='adam')
train_arg.add_argument('--max_epochs', type=int, default=20) # 15
train_arg.add_argument('--lr_rate', type=float, default=0.005)
train_arg.add_argument('--lr_decay', type=float, default=0.0)
train_arg.add_argument('--l2_norm', type=float, default=0.00001)
train_arg.add_argument('--clip', type=float, default=5.0, help='clip the gradient by norm')
train_arg.add_argument('--teacher_ratio', type=float, default=1.0, help='probability of using targets for learning')
train_arg.add_argument('--dropout', type=float, default=0.0)
train_arg.add_argument('--early_stop_count', type=int, default=2)
train_arg.add_argument('--epoch_load', type=int, default=0)
train_arg.add_argument('--load_param', type=util.str2bool, nargs='?', const=True, default=False)
train_arg.add_argument('--start_epoch', type=int, default=0) # when to use SentMoE
# 4. MISC
misc_arg = parser.add_argument_group('MISC')
misc_arg.add_argument('--seed', type=int, default=0, metavar='S', help='random seed (default: 1)')
misc_arg.add_argument('--batch_size', type=int, default=64, metavar='N', help='input batch size for training (default: 64)')
misc_arg.add_argument('--db_size', type=int, default=30)
misc_arg.add_argument('--bs_size', type=int, default=94)
misc_arg.add_argument('--beam_width', type=int, default=10, help='Beam width used in beamsearch')
#
# 5. Here add new args
new_arg = parser.add_argument_group('New')
new_arg.add_argument('--intent_type', type=str, default=None, help='separate experts by intents: None, domain, sysact or domain_act') # pp added
# different implementation of moe
# 1. only weight loss & hyper weights
# --use_moe_loss=True --learn_loss_weight=False --use_moe_model=False
# 2. only weight loss & learn weights
# --use_moe_loss=True --learn_loss_weight=True --use_moe_model=False
# 3. only split models
# --use_moe_loss=False --learn_loss_weight=False --use_moe_model=True
# 4. both & hyper weights
# --use_moe_loss=True --learn_loss_weight=False --use_moe_model=True
# 5. both & learn weights
# --use_moe_loss=True --learn_loss_weight=True --use_moe_model=True
new_arg.add_argument('--use_moe_loss', type=util.str2bool, nargs='?', const=True, default=False, help='inner models weighting loss')
new_arg.add_argument('--learn_loss_weight', type=util.str2bool, nargs='?', const=True, default=False, help='learn weight of moe loss')
new_arg.add_argument('--use_moe_model', type=util.str2bool, nargs='?', const=True, default=False, help='inner models structure partition')
new_arg.add_argument('--debug', type=util.str2bool, nargs='?', const=True, default=False, help='if True use small data for debugging')
new_arg.add_argument('--train_valid', type=util.str2bool, nargs='?', const=True, default=False, help='if True add valid data for training')
new_arg.add_argument('--train_ratio', type=float, default=1.0) # use xx percent of training data
new_arg.add_argument('--lambda_expert', type=float, default=0.5) # use xx percent of training data
new_arg.add_argument('--mu_expert', type=float, default=0.5) # use xx percent of training data
new_arg.add_argument('--gamma_expert', type=float, default=0.5) # use xx percent of training data
new_arg.add_argument('--SentMoE', type=util.str2bool, nargs='?', const=True, default=False, help='if True use sentence info')
new_arg.add_argument('--if_detach', type=util.str2bool, nargs='?', const=True, default=False) # if detach expert parts
new_arg.add_argument('--rp_share_rnn', type=util.str2bool, nargs='?', const=True, default=True) # if detach expert parts
new_arg.add_argument('--future_info', type=str, default='proba') # use hidd or proba
args = parser.parse_args()
args.device = detected_device.type
print('args.device={}'.format(args.device))
print('args.intent_type={}'.format(args.intent_type))
# construct dirs
args.model_dir = '%s/model' % args.result_dir
args.train_output = '%s/data/train_dials' % args.result_dir
args.valid_output = '%s/data/valid_dials' % args.result_dir
args.decode_output = '%s/data/test_dials' % args.result_dir
args.delex_path = '%s/delex.json' % args.data_dir
print(args)
# pp added: init seed
util.init_seed(args.seed)
def trainOne(print_loss_total,print_act_total, print_grad_total, input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor=None, name=None):
loss, loss_acts, grad = model.model_train(input_tensor, input_lengths, target_tensor, target_lengths, db_tensor, bs_tensor, mask_tensor, name)
# pp added: experts' loss
# print('@'*20, '\n', target_tensor)
'''
if args.use_moe_loss and False: # data separate by intents
gen_loss_list = []
if mask_tensor is not None: # data separate by intents
# print(mask_tensor)
for mask in mask_tensor: # each intent has a mask [Batch, 1]
target_tensor_i = target_tensor.clone()
target_tensor_i = target_tensor_i.masked_fill_(mask, value=PAD_token)
# print(mask)
# print(target_tensor_i)
# print('*'*50)
loss_i, loss_acts_i, grad_i = model.model_train(input_tensor, input_lengths, target_tensor_i, target_lengths, db_tensor, bs_tensor, mask_tensor, name)
gen_loss_list.append(loss_i)
# print('loss', loss, '; mean_experts_loss', torch.mean(torch.tensor(gen_loss_list)), '\ngen_loss_list', ['%.4f' % s if s!=0 else '0' for s in gen_loss_list])
# mu_expert = 0.5
mu_expert = args.mu_expert
loss = (1 - mu_expert) * loss + mu_expert * torch.mean(torch.tensor(gen_loss_list))
'''
#print(loss, loss_acts)
print_loss_total += loss
print_act_total += loss_acts
print_grad_total += grad
model.global_step += 1
model.sup_loss = torch.zeros(1)
return print_loss_total, print_act_total, print_grad_total
def trainIters(model, intent2index, n_epochs=10, args=args):
prev_min_loss, early_stop_count = 1 << 30, args.early_stop_count
start = datetime.datetime.now()
# Valid_Scores, Test_Scores = [], []
Scores = []
val_dials_gens, test_dials_gens = [], []
origin = args.SentMoE # original flag
for epoch in range(1, n_epochs + 1):
# pp added
if origin:
if epoch > args.start_epoch:
args.SentMoE = True
print('BeginSentMOE', '-'*50)
else:
args.SentMoE = False
print('%s\nEpoch=%s (%s %%)' % ('~'*50, epoch, epoch / n_epochs * 100))
print_loss_total = 0; print_grad_total = 0; print_act_total = 0 # Reset every print_every
start_time = datetime.datetime.now()
# watch out where do you put it
model.optimizer = Adam(lr=args.lr_rate, params=filter(lambda x: x.requires_grad, model.parameters()), weight_decay=args.l2_norm)
model.optimizer_policy = Adam(lr=args.lr_rate, params=filter(lambda x: x.requires_grad, model.policy.parameters()), weight_decay=args.l2_norm)
# Training
model.train()
step = 0
for data in train_loader: # each element of data tuple has [batch_size] samples
step += 1
model.optimizer.zero_grad()
model.optimizer_policy.zero_grad()
# Transfer to GPU
if torch.cuda.is_available():
data = [data[i].cuda() if isinstance(data[i], torch.Tensor) else data[i] for i in range(len(data))]
input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor = data
print_loss_total, print_act_total, print_grad_total = trainOne(print_loss_total, print_act_total, print_grad_total, input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor)
if step > 1 and args.debug:
break # for debug
if args.train_ratio!=1.0 and step > args.train_ratio * len(train_loader):
break # only train of
train_len = len(train_loader) # 886 data # len(train_loader.dataset.datasets) # 8423 dialogues
print_loss_avg = print_loss_total / train_len
print_act_total_avg = print_act_total / train_len
print_grad_avg = print_grad_total / train_len
print('Train Time:%.4f' % (datetime.datetime.now() - start_time).seconds)
print('Train Loss: %.6f\nTrain Grad: %.6f' % (print_loss_avg, print_grad_avg))
if not args.debug:
step = 0
# VALIDATION
if args.train_valid: # if add valid data for training
model.train()
valid_loss = 0
for name, val_file in list(val_dials.items())[-step:]:
loader = multiwoz_dataloader.get_loader_by_dialogue(val_file, name,
input_lang_word2index, output_lang_word2index,
args.intent_type, intent2index)
data = iter(loader).next()
# Transfer to GPU
if torch.cuda.is_available():
data = [data[i].cuda() if isinstance(data[i], torch.Tensor) else data[i] for i in range(len(data))]
input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor = data
proba, _, _ = model.forward(input_tensor, input_lengths, target_tensor, target_lengths, db_tensor,
bs_tensor, mask_tensor) # pp added: mask_tensor
proba = proba.view(-1, model.vocab_size) # flatten all predictions
loss = model.gen_criterion(proba, target_tensor.view(-1))
valid_loss += loss.item()
valid_len = len(val_dials) # 1000
valid_loss /= valid_len
# pp added: evaluate valid
print('Train Valid Loss: %.6f' % valid_loss)
# pp added
with torch.no_grad():
model.eval()
val_dials_gen = {}
valid_loss = 0
for name, val_file in list(val_dials.items())[-step:]: # for py3
loader = multiwoz_dataloader.get_loader_by_dialogue(val_file, name,
input_lang_word2index, output_lang_word2index,
args.intent_type, intent2index)
data = iter(loader).next()
# Transfer to GPU
if torch.cuda.is_available():
data = [data[i].cuda() if isinstance(data[i], torch.Tensor) else data[i] for i in range(len(data))]
input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor = data
proba, _, _ = model.forward(input_tensor, input_lengths, target_tensor, target_lengths, db_tensor,
bs_tensor, mask_tensor) # pp added: mask_tensor
proba = proba.view(-1, model.vocab_size) # flatten all predictions
loss = model.gen_criterion(proba, target_tensor.view(-1))
valid_loss += loss.item()
# pp added: evaluation - Plan A
# models.eval()
output_words, loss_sentence = model.predict(input_tensor, input_lengths, target_tensor, target_lengths,
db_tensor, bs_tensor, mask_tensor)
# models.train()
val_dials_gen[name] = output_words
valid_len = len(val_dials) # 1000
valid_loss /= valid_len
# pp added: evaluate valid
print('Valid Loss: %.6f' % valid_loss)
# BLEU, MATCHES, SUCCESS, SCORE, P, R, F1
Valid_Score = evaluator.summarize_report(val_dials_gen, mode='Valid')
# Valid_Score = evaluateModel(val_dials_gen, val_dials, delex_path, mode='Valid')
val_dials_gens.append(val_dials_gen) # save generated output for each epoch
# Testing
# pp added
model.eval()
test_dials_gen ={}
test_loss = 0
for name, test_file in list(test_dials.items())[-step:]:
loader = multiwoz_dataloader.get_loader_by_dialogue(test_file, name,
input_lang_word2index, output_lang_word2index,
args.intent_type, intent2index)
data = iter(loader).next()
# Transfer to GPU
if torch.cuda.is_available():
data = [data[i].cuda() if isinstance(data[i], torch.Tensor) else data[i] for i in range(len(data))]
input_tensor, input_lengths, target_tensor, target_lengths, bs_tensor, db_tensor, mask_tensor = data
proba, _, _ = model.forward(input_tensor, input_lengths, target_tensor, target_lengths, db_tensor,
bs_tensor, mask_tensor) # pp added: mask_tensor
proba = proba.view(-1, model.vocab_size) # flatten all predictions
loss = model.gen_criterion(proba, target_tensor.view(-1))
test_loss += loss.item()
output_words, loss_sentence = model.predict(input_tensor, input_lengths, target_tensor, target_lengths,
db_tensor, bs_tensor, mask_tensor)
test_dials_gen[name] = output_words
# pp added: evaluate test
test_len = len(test_dials) # 1000
test_loss /= test_len
# pp added: evaluate valid
print('Test Loss: %.6f' % valid_loss)
Test_Score = evaluator.summarize_report(test_dials_gen, mode='Test')
# Test_Score = evaluateModel(test_dials_gen, test_dials, delex_path, mode='Test')
test_dials_gens.append(test_dials_gen)
try:
with open(args.decode_output + '/test_dials_gen_%s.json' % epoch, 'w') as outfile:
json.dump(test_dials_gen, outfile, indent=4)
except:
print('json.dump.err.test')
model.train()
# pp added: evaluation - Plan B
# print(50 * '=' + 'Evaluating start...')
# # eval_with_train(models)
# eval_with_train3(models, val_dials, mode='valid')
# eval_with_train3(models, test_dials, mode='test')
# print(50 * '=' + 'Evaluating end...')
model.saveModel(epoch)
# BLEU, MATCHES, SUCCESS, SCORE, TOTAL
Scores.append(tuple([epoch]) + Valid_Score + tuple(['%.2f'%np.exp(valid_loss)]) + Test_Score + tuple(['%.2f'%np.exp(test_loss)])) # combine the tuples; 11 elements
# summary of evaluation metrics
import pandas as pd
# BLEU, MATCHES, SUCCESS, SCORE, P, R, F1
fields = ['Epoch',
'Valid BLEU', 'Valid Matches', 'Valid Success', 'Valid Score', 'Valid P', 'Valid R', 'Valid F1', 'Valid PPL',
'Test BLEU', 'Test Matches', 'Test Success', 'Test Score', 'Test P', 'Test R', 'Test F1', 'Test PPL']
df = pd.DataFrame(Scores, columns=fields)
sdf = df.sort_values(by=['Valid Score'], ascending=False)
print('Top3:', '=' * 60)
print(sdf.head(3).transpose())
print('Best:', '=' * 60) # selected by valid score
best_df = sdf.head(1)[['Epoch', 'Test PPL', 'Test BLEU', 'Test Matches', 'Test Success', 'Test Score', 'Test P', 'Test R', 'Test F1']]
print(best_df.transpose())
# save best prediction to json, evaluated on valid set
best_model_id = np.int(best_df['Epoch']) - 1 # epoch start with 1
try:
with open(args.valid_output + '/val_dials_gen.json', 'w') as outfile:
json.dump(val_dials_gens[best_model_id], outfile, indent=4)
except:
print('json.dump.err.valid')
try:
with open(args.decode_output + '/test_dials_gen.json', 'w') as outfile:
json.dump(test_dials_gens[best_model_id], outfile, indent=4)
except:
print('json.dump.err.test')
return best_df
if __name__ == '__main__':
input_lang_index2word, output_lang_index2word, input_lang_word2index, output_lang_word2index = util.loadDictionaries(mdir=args.data_dir)
# pp added: load intents
intent2index, index2intent = util.loadIntentDictionaries(intent_type=args.intent_type, intent_file='{}/intents.json'.format(args.data_dir)) if args.intent_type else (None, None)
# pp added: data loaders
train_loader = multiwoz_dataloader.get_loader('{}/train_dials.json'.format(args.data_dir), input_lang_word2index, output_lang_word2index, args.intent_type, intent2index, batch_size=args.batch_size)
# valid_loader_list = multiwoz_dataloader.get_loader_by_full_dialogue('{}/val_dials.json'.format(args.data_dir), input_lang_word2index, output_lang_word2index, args.intent_type, intent2index)
# test_loader_list = multiwoz_dataloader.get_loader_by_full_dialogue('{}/test_dials.json'.format(args.data_dir), input_lang_word2index, output_lang_word2index, args.intent_type, intent2index)
# Load validation file list:
with open('{}/val_dials.json'.format(args.data_dir)) as outfile:
val_dials = json.load(outfile)
# Load test file list:
with open('{}/test_dials.json'.format(args.data_dir)) as outfile:
test_dials = json.load(outfile)
# delex_path = '%s/delex.json' % args.data_dir
# create dir for generated outputs of valid and test set
pp_mkdir(args.valid_output)
pp_mkdir(args.decode_output)
model = Model(args, input_lang_index2word, output_lang_index2word, input_lang_word2index, output_lang_word2index, intent2index, index2intent)
# models = nn.DataParallel(models, device_ids=[0,1]) # latter for parallel
model = model.to(detected_device)
if args.load_param:
model.loadModel(args.epoch_load)
evaluator = MultiWozEvaluator('MultiWozEvaluator', delex_path=args.delex_path)
# Test_Score = evaluator.summarize_report(test_dials_gen, mode='Test')
trainIters(model, intent2index, n_epochs=args.max_epochs, args=args)
all_end_time = datetime.datetime.now()
print('End time={}'.format(all_end_time.strftime("%Y-%m-%d %H:%M:%S")))
print('Use time={} seconds'.format((all_end_time-all_start_time).seconds))