-
Notifications
You must be signed in to change notification settings - Fork 149
/
main.py
185 lines (146 loc) · 7.23 KB
/
main.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
from __future__ import print_function
from math import ceil
import numpy as np
import sys
import pdb
import torch
import torch.optim as optim
import torch.nn as nn
import generator
import discriminator
import helpers
CUDA = False
VOCAB_SIZE = 5000
MAX_SEQ_LEN = 20
START_LETTER = 0
BATCH_SIZE = 32
MLE_TRAIN_EPOCHS = 100
ADV_TRAIN_EPOCHS = 50
POS_NEG_SAMPLES = 10000
GEN_EMBEDDING_DIM = 32
GEN_HIDDEN_DIM = 32
DIS_EMBEDDING_DIM = 64
DIS_HIDDEN_DIM = 64
oracle_samples_path = './oracle_samples.trc'
oracle_state_dict_path = './oracle_EMBDIM32_HIDDENDIM32_VOCAB5000_MAXSEQLEN20.trc'
pretrained_gen_path = './gen_MLEtrain_EMBDIM32_HIDDENDIM32_VOCAB5000_MAXSEQLEN20.trc'
pretrained_dis_path = './dis_pretrain_EMBDIM_64_HIDDENDIM64_VOCAB5000_MAXSEQLEN20.trc'
def train_generator_MLE(gen, gen_opt, oracle, real_data_samples, epochs):
"""
Max Likelihood Pretraining for the generator
"""
for epoch in range(epochs):
print('epoch %d : ' % (epoch + 1), end='')
sys.stdout.flush()
total_loss = 0
for i in range(0, POS_NEG_SAMPLES, BATCH_SIZE):
inp, target = helpers.prepare_generator_batch(real_data_samples[i:i + BATCH_SIZE], start_letter=START_LETTER,
gpu=CUDA)
gen_opt.zero_grad()
loss = gen.batchNLLLoss(inp, target)
loss.backward()
gen_opt.step()
total_loss += loss.data.item()
if (i / BATCH_SIZE) % ceil(
ceil(POS_NEG_SAMPLES / float(BATCH_SIZE)) / 10.) == 0: # roughly every 10% of an epoch
print('.', end='')
sys.stdout.flush()
# each loss in a batch is loss per sample
total_loss = total_loss / ceil(POS_NEG_SAMPLES / float(BATCH_SIZE)) / MAX_SEQ_LEN
# sample from generator and compute oracle NLL
oracle_loss = helpers.batchwise_oracle_nll(gen, oracle, POS_NEG_SAMPLES, BATCH_SIZE, MAX_SEQ_LEN,
start_letter=START_LETTER, gpu=CUDA)
print(' average_train_NLL = %.4f, oracle_sample_NLL = %.4f' % (total_loss, oracle_loss))
def train_generator_PG(gen, gen_opt, oracle, dis, num_batches):
"""
The generator is trained using policy gradients, using the reward from the discriminator.
Training is done for num_batches batches.
"""
for batch in range(num_batches):
s = gen.sample(BATCH_SIZE*2) # 64 works best
inp, target = helpers.prepare_generator_batch(s, start_letter=START_LETTER, gpu=CUDA)
rewards = dis.batchClassify(target)
gen_opt.zero_grad()
pg_loss = gen.batchPGLoss(inp, target, rewards)
pg_loss.backward()
gen_opt.step()
# sample from generator and compute oracle NLL
oracle_loss = helpers.batchwise_oracle_nll(gen, oracle, POS_NEG_SAMPLES, BATCH_SIZE, MAX_SEQ_LEN,
start_letter=START_LETTER, gpu=CUDA)
print(' oracle_sample_NLL = %.4f' % oracle_loss)
def train_discriminator(discriminator, dis_opt, real_data_samples, generator, oracle, d_steps, epochs):
"""
Training the discriminator on real_data_samples (positive) and generated samples from generator (negative).
Samples are drawn d_steps times, and the discriminator is trained for epochs epochs.
"""
# generating a small validation set before training (using oracle and generator)
pos_val = oracle.sample(100)
neg_val = generator.sample(100)
val_inp, val_target = helpers.prepare_discriminator_data(pos_val, neg_val, gpu=CUDA)
for d_step in range(d_steps):
s = helpers.batchwise_sample(generator, POS_NEG_SAMPLES, BATCH_SIZE)
dis_inp, dis_target = helpers.prepare_discriminator_data(real_data_samples, s, gpu=CUDA)
for epoch in range(epochs):
print('d-step %d epoch %d : ' % (d_step + 1, epoch + 1), end='')
sys.stdout.flush()
total_loss = 0
total_acc = 0
for i in range(0, 2 * POS_NEG_SAMPLES, BATCH_SIZE):
inp, target = dis_inp[i:i + BATCH_SIZE], dis_target[i:i + BATCH_SIZE]
dis_opt.zero_grad()
out = discriminator.batchClassify(inp)
loss_fn = nn.BCELoss()
loss = loss_fn(out, target)
loss.backward()
dis_opt.step()
total_loss += loss.data.item()
total_acc += torch.sum((out>0.5)==(target>0.5)).data.item()
if (i / BATCH_SIZE) % ceil(ceil(2 * POS_NEG_SAMPLES / float(
BATCH_SIZE)) / 10.) == 0: # roughly every 10% of an epoch
print('.', end='')
sys.stdout.flush()
total_loss /= ceil(2 * POS_NEG_SAMPLES / float(BATCH_SIZE))
total_acc /= float(2 * POS_NEG_SAMPLES)
val_pred = discriminator.batchClassify(val_inp)
print(' average_loss = %.4f, train_acc = %.4f, val_acc = %.4f' % (
total_loss, total_acc, torch.sum((val_pred>0.5)==(val_target>0.5)).data.item()/200.))
# MAIN
if __name__ == '__main__':
oracle = generator.Generator(GEN_EMBEDDING_DIM, GEN_HIDDEN_DIM, VOCAB_SIZE, MAX_SEQ_LEN, gpu=CUDA)
oracle.load_state_dict(torch.load(oracle_state_dict_path))
oracle_samples = torch.load(oracle_samples_path).type(torch.LongTensor)
# a new oracle can be generated by passing oracle_init=True in the generator constructor
# samples for the new oracle can be generated using helpers.batchwise_sample()
gen = generator.Generator(GEN_EMBEDDING_DIM, GEN_HIDDEN_DIM, VOCAB_SIZE, MAX_SEQ_LEN, gpu=CUDA)
dis = discriminator.Discriminator(DIS_EMBEDDING_DIM, DIS_HIDDEN_DIM, VOCAB_SIZE, MAX_SEQ_LEN, gpu=CUDA)
if CUDA:
oracle = oracle.cuda()
gen = gen.cuda()
dis = dis.cuda()
oracle_samples = oracle_samples.cuda()
# GENERATOR MLE TRAINING
print('Starting Generator MLE Training...')
gen_optimizer = optim.Adam(gen.parameters(), lr=1e-2)
train_generator_MLE(gen, gen_optimizer, oracle, oracle_samples, MLE_TRAIN_EPOCHS)
# torch.save(gen.state_dict(), pretrained_gen_path)
# gen.load_state_dict(torch.load(pretrained_gen_path))
# PRETRAIN DISCRIMINATOR
print('\nStarting Discriminator Training...')
dis_optimizer = optim.Adagrad(dis.parameters())
train_discriminator(dis, dis_optimizer, oracle_samples, gen, oracle, 50, 3)
# torch.save(dis.state_dict(), pretrained_dis_path)
# dis.load_state_dict(torch.load(pretrained_dis_path))
# ADVERSARIAL TRAINING
print('\nStarting Adversarial Training...')
oracle_loss = helpers.batchwise_oracle_nll(gen, oracle, POS_NEG_SAMPLES, BATCH_SIZE, MAX_SEQ_LEN,
start_letter=START_LETTER, gpu=CUDA)
print('\nInitial Oracle Sample Loss : %.4f' % oracle_loss)
for epoch in range(ADV_TRAIN_EPOCHS):
print('\n--------\nEPOCH %d\n--------' % (epoch+1))
# TRAIN GENERATOR
print('\nAdversarial Training Generator : ', end='')
sys.stdout.flush()
train_generator_PG(gen, gen_optimizer, oracle, dis, 1)
# TRAIN DISCRIMINATOR
print('\nAdversarial Training Discriminator : ')
train_discriminator(dis, dis_optimizer, oracle_samples, gen, oracle, 5, 3)