-
Notifications
You must be signed in to change notification settings - Fork 0
/
lstm.py
220 lines (153 loc) · 7.15 KB
/
lstm.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
import tensorflow as tf
import numpy as np
import os
import pickle
import mlflow
class LSTM:
def __init__(
self,
num_classes,
state_size=512,
layers=2,
heavy_device=None,
light_device=None,
restore=True
):
s# Initializes the lstm and builds the graph when run
self.state_size = state_size
self.layers = layers
self.num_classes = num_classes
self.heavy_device = heavy_device
self.light_device = light_device
def __graph__():
# Build the graph that will process one batch
def step(prev,x):
with tf.device(self.heavy_device):
# x will be a tensor of shape [batch_size,state_size]
# prev will be a tensor of shape [2, num_layers, batch_size, state_size]
# We will unstack this and return a tensor of the same shape to be passed into the next timestep
# Using embeddings we can reshape our number of features into the size of our desired hidden shape
# Get weights or initialize if not already in graph
W = tf.get_variable(name="W", shape=[self.layers, 4, self.state_size, self.state_size],initializer=tf.random_uniform_initializer(minval=0.08, maxval=0.08))
U = tf.get_variable(name = "U", shape=[self.layers, 4, self.state_size, self.state_size],initializer=tf.random_uniform_initializer(minval=-0.08, maxval=0.08))
b = tf.get_variable(name="b", shape=[self.layers, 4, self.state_size], initializer=tf.zeros_initializer())
h_prev,c_prev = tf.unstack(prev)
h_full, c_full = [], []
inp = x
for i in range(self.layers):
with tf.name_scope("gates"):
with tf.name_scope("ft"):
ft = tf.sigmoid(tf.matmul(inp,W[i][0]) + tf.matmul(h_prev[i],U[i][0]) + b[i][0])
with tf.name_scope("it"):
it = tf.sigmoid(tf.matmul(inp,W[i][1]) + tf.matmul(h_prev[i],U[i][1]) + b[i][1])
with tf.name_scope("ot"):
ot = tf.sigmoid(tf.matmul(inp,W[i][2]) + tf.matmul(h_prev[i],U[i][2]) + b[i][2])
with tf.name_scope("ct"):
ct = tf.tanh(tf.matmul(inp,W[i][3]) + tf.matmul(h_prev[i],U[i][3]) + b[i][3])
with tf.name_scope("c"):
c = (ft * c_prev[i]) + (it * ct)
with tf.name_scope("h"):
h = ot * tf.tanh(c)
h_full.append(h)
c_full.append(c)
inp = h
return tf.stack([h_full,c_full])
with tf.device(self.heavy_device):
# This will end up being shape [batch_size,state_size] currently it is [batch_size, seq_length]
x_ = tf.placeholder(shape=[None,None], dtype=tf.int64, name="x_")
y_ = tf.placeholder(shape=[None], dtype=tf.int64, name="y_")
initial_state = tf.placeholder(shape=[2, self.layers,None, self.state_size], dtype=tf.float32, name='initial_state')
# Create embedding. This will be a trainable parameter.
embedding = tf.get_variable("embedding",shape=[self.num_classes,self.state_size])
# This will create a tensor size of [batch_size, seq_length, state_size] which we can feed into our graph
inputs = tf.nn.embedding_lookup(embedding,x_)
# We need to reshape inputs to have seq_length as its first dimension so tf.scan can run over it giving us [batch_size,state_size] at each time step
inputs = tf.transpose(inputs, [1,0,2])
# Now we can pass inputs into tf.scan
# This will give us an output size of [seq_length, 2, num_layers, batch_size, state_size]
outputs = tf.scan(step, inputs, initial_state)
# TODO: Expose this later
last_state = outputs[-1]
# We only want the hidden state on the last layer
# This should be of size batch_size, seq_length, state_size
# if we reshape the first two dimensions we can do a matrix multiply with our new weights to compute logits
states = tf.transpose(outputs, [1,2,3,0,4])[0][-1]
# Now we create our final hidden layer weights
# Our Y value will in theory be size of [batch_size, seq_length * num_classes] except we will reshape this into a column vector
# and it will not be onehot because sparse softmax will handle this
W_f = tf.get_variable(name="W_f", shape=[self.state_size, self.num_classes], initializer=tf.random_uniform_initializer(minval=-0.08, maxval=0.08))
b_f = tf.get_variable(name="b_f", shape=[self.num_classes], initializer=tf.zeros_initializer())
# reshape to size [batch_size * seq_length, state_size]
logits = tf.matmul(tf.reshape(states,[-1,self.state_size]),W_f) + b_f
# Create our predictions
predictions = tf.nn.softmax(logits)
# Because this is sparse softmax, y_ will become onehot and of shape [batch_size * seq_length, num_classes]
# Remember it is already size [batch_length * seq_length] because each character is represented by an int index
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,labels=y_)
loss = tf.reduce_mean(losses)
optimize = tf.train.AdagradOptimizer(0.1).minimize(loss)
self.x_ = x_
self.y_ = y_
self.initial_state = initial_state
self.predictions = predictions
self.loss = loss
self.optimize = optimize
self.last_state = last_state
with tf.device(self.light_device):
self.saver = tf.train.Saver()
# Initialize session
self.sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
if restore:
self.saver.restore(self.sess, tf.train.latest_checkpoint('./saves'))
else:
self.sess.run(tf.global_variables_initializer())
print("Building Graph")
__graph__()
print("Done.")
def train(self, train_step, iterations=2000, restore=False):
# Called to train model
try:
i = 0
while True:
# Get our batch of random samples
# Now let's encode this in an embedding
x_sample, y_sample = train_step.__next__()
batch_size = x_sample.shape[0]
feed = {self.x_: x_sample,
self.y_: y_sample.flatten(),
self.initial_state: np.zeros(shape=(2, self.layers,batch_size, self.state_size), dtype=np.float32)}
_,loss = self.sess.run([self.optimize, self.loss], feed_dict=feed)
if (i%100 == 0):
print("Loss: " + str(loss))
if (i%1000 == 0):
print("Saving model....")
save_path = self.saver.save(self.sess, "./model_path/saves/model.ckpt")
i += 1
print(i, end="\r", flush=True)
except KeyboardInterrupt:
print("Interrupted... Saving model.")
save_path = self.saver.save(self.sess, "./model_path/saves/model.ckpt")
def generate(self, char2ix, ix2char, seq_length, input_seed='a'):
# Called to generate samples from trained model
#seed = np.random.choice(list(char2ix.values()))
seed = char2ix[input_seed]
out = input_seed
initialize = False
for i in range(seq_length):
if initialize == False:
# if it is the first letter in sequence, intialize with default hidden states
feed = {
self.x_: np.array([seed]).reshape(1,1),
self.initial_state: np.zeros(shape=(2, self.layers, 1, self.state_size), dtype=np.float32)
}
else:
# otherwise, we need to intialize previous hidden state with the returned last state
feed = {
self.x_: np.array([seed]).reshape(1, 1),
self.initial_state: last_state
}
predictions,last_state = self.sess.run([self.predictions, self.last_state], feed_dict = feed)
initialize = True
seed = np.random.choice(range(len(ix2char)), p=np.ravel(predictions))
out += ix2char[seed]
return out