-
Notifications
You must be signed in to change notification settings - Fork 0
/
SeqGAN.py
210 lines (185 loc) · 6.18 KB
/
SeqGAN.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
import tensorflow as tf
import numpy as np
from Dataloader import *
from Discriminator import *
from Generator import *
from Oracle import *
from Discriminator_DataLoader import *
############# Changes from Original Paper #######################
'''
1) I think my Keras Sequence DataLoader should be able to replace rollout.py if i add the get_reward function here
2) I'm not using the Oracle to generate samples because that's dumb, I'm just going to use real examples
'''
##################################################################
########################## Variables #############################
# Directories
data_dir = ''
oracle_checkpoint_dir = ''
generator_checkpoint_dir = ''
discriminator_checkpoint_dir = ''
# Session Settings
load_oracle_weights = False
load_generator_weights = False
load_discriminator_weights = False
train_oracle = True
save_oracle_training_weights = True
oracle_training_epochs = 3
oracle_training_callbacks = []
oracle_training_verbose = 1
pre_train_generator = True
save_generator_training_weights = True
generator_pre_training_verbose = 1
generator_pre_training_epochs = 3
generator_pre_training_callbacks = []
generator_pre_training_verbose = 1
pre_train_discriminator = True
save_discriminator_training_weights = True
discriminator_pre_training_epochs = 3
discriminator_pre_training_callbacks = []
discriminator_pre_training_verbose = 1
# Hyperparameters
batch_size = 128
sequence_length = 128
validation_split=0.1
seed=None
use_word_vectors=False
# Oracle
oracle_hidden_units = 256
oracle_leaky_relu_alpha = 0.1
oracle_layers = 1
oracle_optimizer = tf.keras.optimizers.Adam(lr=0.01)
oracle_dropout_keep_prob = 1.0
oracle_l2_regularization_lambda = 0.0
oracle_loss = 'categorical_crossentropy'
oracle_metrics = ['loss', 'val_loss']
# Generator
generator_hidden_units = 256
generator_leaky_relu_alpha = 0.1
generator_layers = 1
generator_optimizer = tf.keras.optimizers.Adam(lr=0.01)
generator_dropout_keep_prob = 1.0
generator_l2_regularization_lambda = 0.0
generator_loss = 'categorical_crossentropy'
generator_metrics = ['loss', 'val_loss']
# Discriminator
filter_sizes_by_layer = []
number_of_filters_by_layer = []
discriminator_leaky_relu_alpha = 0.1
discriminator_dropout_keep_prob = 1.0
discriminator_l2_regularization_lambda = 0.0
discriminator_learning_rate = 1e-4
carry_bias = 0
discriminator_metrics = ['loss', 'val_loss']
discriminator_pre_training_fake_batch_size = batch_size
##################################################################
# Initialize DataLoaders
train_dl = DataLoader(
data_filename = data_dir,
batch_size = batch_size,
sequence_length = sequence_length,
validation_split = validation_split,
seed = seed,
data_type = "train",
use_word_vectors=use_word_vectors
)
val_dl = DataLoader(
data_filename = data_dir,
batch_size = batch_size,
sequence_length = sequence_length,
validation_split = validation_split,
seed = seed,
data_type = "val",
use_word_vectors=use_word_vectors
)
disc_dl = Discriminator_DataLoader(
data_filename = data_dir,
batch_size = batch_size,
sequence_length = sequence_length,
validation_split = validation_split,
fake_batch_size = discriminator_pre_training_fake_batch_size,
seed = seed,
data_type = "val",
use_word_vectors=use_word_vectors
)
# Initialize Models
oracle = Oracle(
train_data_loader = train_dl,
validation_data_loader = val_dl,
units = oracle_hidden_units,
leaky_relu_alpha = oracle_leaky_relu_alpha,
num_layers = oracle_layers,
opt = oracle_optimizer,
dropout_keep_prob = oracle_dropout_keep_prob,
l2_reg_lambda = oracle_l2_regularization_lambda,
sequence_length = sequence_length,
loss = oracle_loss,
metrics = oracle_metrics
)
gen = Generator(
train_data_loader = train_dl,
validation_data_loader = val_dl,
units = generator_hidden_units,
leaky_relu_alpha=generator_leaky_relu_alpha,
num_layers = generator_layers,
opt = generator_optimizer,
dropout_keep_prob = generator_dropout_keep_prob,
l2_reg_lambda = generator_l2_regularization_lambda,
sequence_length = sequence_length,
loss = generator_loss,
metrics = generator_metrics
)
disc = Discriminator(
table_len = len(train_dl.st.table),
filter_sizes = filter_sizes_by_layer,
num_filters = number_of_filters_by_layer,
sequence_len = sequence_length,
l2_reg_lambda = discriminator_l2_regularization_lambda,
dropout_keep_prob = discriminator_dropout_keep_prob,
learning_rate = discriminator_learning_rate,
leaky_relu_alpha = discriminator_leaky_relu_alpha,
carry_bias=carry_bias,
metrics = discriminator_metrics
)
# Load Model Weights if you have them
if load_oracle_weights:
oracle.load_weights(oracle_checkpoint_dir)
if load_generator_weights:
gen.load_weights(generator_checkpoint_dir)
if load_discriminator_weights:
disc.load_weights(discriminator_checkpoint_dir)
# Train Oracle
if train_oracle:
oracle_history = oracle.train(
epochs = oracle_training_epochs,
verbose = oracle_training_verbose,
callbacks = oracle_training_callbacks,
save_weights = save_oracle_training_weights,
filepath = oracle_checkpoint_dir
)
# Pretrain Generator
if pre_train_generator:
generator_history = gen.pretrain(
epochs = generator_pre_training_epochs,
verbose = generator_pre_training_verbose,
callbacks = generator_pre_training_callbacks,
save_weights = save_generator_training_weights,
filepath = generator_checkpoint_dir
)
# Pretrain Discriminator
if pre_train_discriminator:
for i in range(discriminator_pre_training_epochs):
print("AYYYYYY starting epoch "+str(i+1) + '/' + str(discriminator_pre_training_epochs))
print("Training History: " + str(history))
epoch_finished = False
batch_num = 1
while not epoch_finished:
epoch_finished, x, y = disc_dl.get_batch(gen)
history = disc.model.train_on_batch(
x = x,
y = y,
reset_metrics = False,
return_dict = True
)
if batch_num % 20 == 0:
print("Finished batch " + str(batch_num))
batch_num += 1