-
Notifications
You must be signed in to change notification settings - Fork 0
/
Discriminator.py
147 lines (122 loc) · 5.56 KB
/
Discriminator.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
import tensorflow as tf
import numpy as np
############# Changes from Original Paper #######################
'''
1) Only included L2 regularization to kernel, not bias
2) Used Leaky-Relu instead of normal relu
3) Made this weird custom layer to handle all the different convolutional stuff
4) Didn't get predictions, only got scores, added get_predictions function that takes the logits
5) Changed from 2 class softmax to sigmoid output binary_crossentropy loss cuz it seemed pointless
6) This one was a bit confusing, so there's probably tons of errors in there
'''
class Highway(tf.keras.layers.Layer):
def __init__(self, size, leaky_relu_alpha = 0.1, carry_bias=-2.0):
super(Highway, self).__init__()
self.leaky_relu_alpha = leaky_relu_alpha
self.w_t = tf.Variable(tf.truncated_normal([size, size], stddev=0.1), name="weight_transform")
self.b_t = tf.Variable(tf.constant(carry_bias, shape=[size]), name="bias_transform")
self.w = tf.Variable(tf.truncated_normal([size, size], stddev=0.1), name="weight")
self.b = tf.Variable(tf.constant(0.1, shape=[size]), name="bias")
def call(self, inputs):
t = tf.sigmoid(tf.matmul(inputs, self.w_t) + self.b_t, name="transform_gate")
h = tf.keras.activations.relu((tf.matmul(inputs, self.w) + self.b),alpha = self.leaky_relu_alpha)
c = tf.sub(1.0, t, name="carry_gate")
# LOL thc
return tf.add(tf.mul(self.h, self.t), tf.mul(x, self.c), "y")
class Convolutional_Bullshit(tf.keras.layers.Layer):
def __init__(self, filter_sizes, num_filters, kernel_regularizer, table_len,
sequence_length, carry_bias, leaky_relu_alpha):
self.filter_sizes = filter_sizes
self.num_filters = num_filters
self.kernel_regularizer = kernel_regularizer
self.table_len = table_len
self.leaky_relu = tf.keras.layers.LeakyReLU(alpha = leaky_relu_alpha)
self.seq_len = sequence_length
self.carry_bias = carry_bias
def call(self, inputs):
pooled_outputs = []
for filter_size, num_filters in zip(self.filter_sizes, self.num_filters):
conv = tf.keras.layers.Conv2D(
filters = num_filters,
kernel_size = filter_size,
strides = 1,
padding='valid',
kernel_regularizer = self.kernel_regularizer)(inputs)
h = self.leaky_relu(conv)
pooled = tf.keras.layers.MaxPool2D(
pool_sizes = (filter_size, self.table_len),
strides = 1,
padding = 'valid'
)
pooled_outputs.append(pooled)
num_filters_total = sum(self.num_filters)
h_pool = tf.concat(pooled_outputs, 3)
h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])
h_highway = Highway(h_pool_flat, size = h_pool_flat.get_shape()[1], leaky_relu_alpha = 0.1, carry_bias=self.carry_bias)
return h_highway
class Discriminator(object):
def __init__(self, table_len, filter_sizes, num_filters, sequence_len = 128,
l2_reg_lambda=0.0, dropout_keep_prob = 1.0, learning_rate = 1e-4,
leaky_relu_alpha = 0.1, carry_bias = 0, metrics = ['loss','val_loss']):
self.table_len = table_len
self.input_shape = [None, self.table_len]
self.filter_sizes = filter_sizes
self.num_filters = num_filters,
self.dropout_keep_prob = dropout_keep_prob
self.leaky_relu_alpha = leaky_relu_alpha
self.seq_len = sequence_len
self.opt = tf.keras.optimizers.Adam(learning_rate = learning_rate)
self.metrics = metrics
self.kernel_regularizer = tf.keras.regularizers.l2(l2 = l2_reg_lambda)
self.carry_bias = carry_bias,
self.loss = tf.keras.losses.categorical_crossentropy
self.model = self.buid_model()
def buid_model(self):
model = tf.keras.Sequential()
model.add(
Convolutional_Bullshit(
filter_sizes= self.filter_size,
num_filters = self.num_filters,
kernel_regularizer = self.kernel_regularizer,
table_len = self.table_len,
sequence_length = self.seq_len,
carry_bias = self.carry_bias,
leaky_relu_alpha = self.leaky_relu_alpha
)
)
model.add(
tf.keras.layers.Dropout(rate = self.dropout_keep_prob)
)
model.add(
tf.keras.layers.Dense(
units = 1,
activation='sigmoid'
)
)
model.compile(
optimizer = self.opt,
loss = 'binary_crossentropy',
metrics = self.metrics
)
return model
def train(self, logits, labels, num_epochs, verbose = 1, callbacks=[], save_weights = False, filepath = ''):
history = self.model.fit(
x = logits,
y = labels,
epochs = num_epochs,
verbose = verbose,
callbacks = callbacks,
)
if (save_weights):
save_weights(filepath)
return history
def load_weights(self, filepath):
print("Loading weights from " + str(filepath))
self.model.load_weights(filepath)
print("Weights Loaded")
def save_weights(self, filepath):
print("Saving weights to " + str(filepath))
self.model.save_weights(filepath)
print("Weights Saved")
def get_predictions(logits):
return np.round(logits)