Newer
Older
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
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
from models_665 import Generator
from models_665 import Discriminator
import methods_665
CUDA = True
VOCAB_SIZE = 500 ## adjust hyperparameters as necessary
MAX_SEQ_LEN = 500
START_LETTER = 0
BATCH_SIZE = 32
MLE_TRAIN_EPOCHS = 30
ADV_TRAIN_EPOCHS = 20
POS_NEG_SAMPLES = 704
GEN_EMBEDDING_DIM = 32
GEN_HIDDEN_DIM = 32
DIS_EMBEDDING_DIM = 64
DIS_HIDDEN_DIM = 32
#oracle_samples_path = './oracle_samples.trc'
oracle_samples_path = './zorkdata_long_500.pt' ## INPUT DATA ## set dataset and model parameter paths
oracle_targets_path = './zorkdata_long_500_target.pt' ## LABEL DATA
oracle_lengths_path = './zorkdata_lengths_500.pt' ## INPUT LENGTHS
oracle_state_dict_path = './oracle_ZORK20_60_v2.trc' ## can ignore
pretrained_gen_path = './gen_ZORK500_MLEtrain_v2.trc' ## GENERATOR SAVE FILE
pretrained_dis_path = './dis_ZORK500_pretrain_v2.trc' ## DISCRIMINATOR SAVE FILE
vocab_path = 'zork_vocab.txt' ## VOCAB LIST
output_path = 'generated_inputs/gen_input' ## output generatred samples
def train_generator_MLE(gen, gen_opt, oracle, real_data_samples, real_data_lengths, 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, inp_lengths, target = methods_665.prepare_generator_batch(real_data_samples[i:i + BATCH_SIZE], real_data_lengths[i:i + BATCH_SIZE], start_letter=START_LETTER,
gpu=CUDA)
# if((epoch == 0) and (i == 0)):
# print(inp, inp_lengths, target)
gen_opt.zero_grad()
loss = gen.batchNLLLoss(inp, inp_lengths, 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 = methods_665.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.
"""
seq_len = 20
for batch in range(num_batches):
s = gen.sample(BATCH_SIZE*2) # 64 works best
inp, inp_lengths, target = methods_665.prepare_generator_batch(s, seq_len, 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 = 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, real_data_targets, gan_training=False):
"""
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 = real_data_samples[-64:]
seq_len = len(pos_val[0])
neg_val = generator.sample(64)
pos_targ = real_data_targets[-64:]
# print(pos_val, pos_targ)
val_inp, val_target = methods_665.prepare_discriminator_data(pos_val, neg_val, gpu=CUDA, real_target=pos_targ)
for d_step in range(d_steps):
s = batchwise_sample(generator, POS_NEG_SAMPLES, BATCH_SIZE)
dis_inp, dis_target = methods_665.prepare_discriminator_data(real_data_samples, torch.empty(0, seq_len).cuda(), gpu=CUDA, real_target=real_data_targets)
dis_ninp, dis_ntarget = methods_665.prepare_discriminator_data(torch.empty(0, seq_len).cuda(), 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, POS_NEG_SAMPLES, BATCH_SIZE):
inp, target = dis_inp[i:i + BATCH_SIZE], dis_target[i:i + BATCH_SIZE]
ninp, ntarget = dis_ninp[i:i + BATCH_SIZE], dis_ntarget[i:i + BATCH_SIZE]
dis_opt.zero_grad()
out = discriminator.batchClassify(inp)
nout = discriminator.batchClassify(ninp)
loss_fn = nn.MSELoss() ## IS USING MSE HERE OK!?!??!?!
# if(gan_training): #bandiad due to lack of effective oracle
# for x in target:
# if x[0] == 0:
# target[1:] = out[1:]
t_out = torch.stack((out[0], out[1], out[2], nout[0]))
t_target = torch.stack((target[0], target[1], target[2], ntarget[0]))
loss = loss_fn(t_out, t_target)
# loss += loss_fn(nout[0], ntarget[0])
loss.backward()
dis_opt.step()
total_loss += loss.data.item()
total_acc += torch.sum((out>0.5)==(target>0.5)).data.item() # ZACH TODO: only use first column here?
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)
pred_pos = val_pred[0:64]
pred_neg = val_pred[-64:]
targ_pos = val_target[0:64]
targ_neg = val_target[-64:]
val_pos_class, val_pos_valid, val_pos_cov, val_neg_class, val_neg_valid, val_neg_cov = 0, 0, 0, 0, 0, 0
for x in range(len(pred_pos)):
val_pos_class += abs(pred_pos[x][0] - targ_pos[x][0])
val_pos_valid += abs(pred_pos[x][1] - targ_pos[x][1])
val_pos_cov += abs(pred_pos[x][2] - targ_pos[x][2])
val_neg_class += abs(pred_neg[x][0] - targ_neg[x][0])
val_neg_valid += abs(pred_neg[x][1] - targ_neg[x][1])
val_neg_cov += abs(pred_neg[x][2] - targ_neg[x][2])
print(' average_loss = %.4f, train_acc = %.4f' % (
total_loss, total_acc))
print(' val_pos_class = %.4f, val_pos_valid = %.4f, val_pos_cov = %.4f' % (val_pos_class / 64., val_pos_valid / 64., val_pos_cov / 64.))
print(' val_neg_class = %.4f, val_neg_valid = %.4f, val_neg_cov = %.4f' % (val_neg_class / 64., val_neg_valid / 64., val_neg_cov / 64.))
# MAIN
if __name__ == '__main__':
oracle = Generator(GEN_EMBEDDING_DIM, GEN_HIDDEN_DIM, VOCAB_SIZE, MAX_SEQ_LEN, gpu=CUDA, oracle_init=True)
#oracle.load_state_dict(torch.load(oracle_state_dict_path))
oracle_samples = torch.load(oracle_samples_path)
oracle_targets = torch.load(oracle_targets_path)
oracle_lengths = torch.load(oracle_lengths_path)
# 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(GEN_EMBEDDING_DIM, GEN_HIDDEN_DIM, VOCAB_SIZE, MAX_SEQ_LEN, gpu=CUDA)
dis = 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, oracle_lengths, MLE_TRAIN_EPOCHS) ## comment out if loading gen
torch.save(gen.state_dict(), pretrained_gen_path) ## comment out if loading gen
# gen.load_state_dict(torch.load(pretrained_gen_path)) ## uncomment if loading pre-trained gen
# PRETRAIN DISCRIMINATOR
print('\nStarting Discriminator Training...')
dis_optimizer = optim.Adagrad(dis.parameters())
train_discriminator(dis, dis_optimizer, oracle_samples, gen, oracle, 30, 3, oracle_targets, gan_training=True) ## comment out if loading dis
torch.save(dis.state_dict(), pretrained_dis_path) ## comment out if loading dis
# dis.load_state_dict(torch.load(pretrained_dis_path)) ## uncomment if loading pre-trained dis
# ADVERSARIAL TRAINING
print('\nStarting Adversarial Training...') ## adv training, comment out if you're loading an
oracle_loss = methods_665.batchwise_oracle_nll(gen, oracle, POS_NEG_SAMPLES, BATCH_SIZE, MAX_SEQ_LEN, ## already adv-trained model
start_letter=START_LETTER, gpu=CUDA) ## oracle loss is a metric i was not using and can be ignored
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, 2, 2, oracle_targets)
torch.save(dis.state_dict(), pretrained_dis_path) ## saving post-adv training if desired
torch.save(gen.state_dict(), pretrained_gen_path)
# gen.load_state_dict(torch.load(pretrained_gen_path)) ## load adv trained models if desired
# dis.load_state_dict(torch.load(pretrained_dis_path))
#generate and save examples as desired
# LOAD VOCAB
vocab_list = []
file1 = open(vocab_path, 'r')
Lines = file1.readlines()
count = 0
# Strips the newline character
for line in Lines:
vocab_list.append(line.strip())
#print(vocab_list)
# TEST SAMPLE QUALITY ## predicted labels should have decent coverage/validity
test_inp = gen.sample(20)
#h = dis.init_hidden(test_inp.size()[0])
test_out = dis.batchClassify(test_inp)
print(test_out)
# SAVE SAMPLES TO FILES
test_samples = gen.sample(128) ## choose number of samples to save
test_out = dis.batchClassify(test_samples)
test_samples = test_samples.cpu()
samples_list = test_samples.numpy()
## OUTPUT GENERATED SAMPLES
for i in range(len(samples_list)):
#print(test_out[i][2])
#if test_out[i][2] > 0.1: ## can assign test to only save good (predicted) files
with open(output_path+str(i)+'.txt', 'w') as vocab_file: ## name output sequence files
for j in samples_list[i]:
if(j == 1):
break
if(j > len(vocab_list)):
break
vocab_file.write('%s\n' % vocab_list[j])