-
Notifications
You must be signed in to change notification settings - Fork 1
/
NTMCell.py
204 lines (131 loc) · 11.2 KB
/
NTMCell.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
import tensorflow as tf
import numpy as np
import Batch_Focusing
import Batch_RWV_Generation
class NTMCell(tf.keras.layers.AbstractRNNCell):
def __init__(self, rnn_size, memory_rows, memory_columns, num_read_heads, num_write_heads, num_bits_per_output_vector, addressing_type = 'LOC', shift_range = tf.range(-1,2), **kwargs
):
super().__init__(**kwargs)
self.rnn_size = rnn_size
self.memory_rows = memory_rows #The "N" or "size of memory" in Literature
self.memory_columns = memory_columns #The "M" or "memory vector's dimension" in Literature
self.num_read_heads = num_read_heads
self.num_write_heads = num_write_heads
self.num_bits_per_output_vector = num_bits_per_output_vector
self.addressing_type = addressing_type
if ((self.addressing_type != 'LOC') and (self.addressing_type != 'CONT')):
raise ValueError('Incorrect Addressing Type: Allowed values are "LOC" for Location based Focusing and "CONT" for Content based Focusing.')
self.shift_range = shift_range
self.total_num_heads = self.num_read_heads + self.num_write_heads
self.controller = tf.keras.layers.LSTMCell(self.rnn_size)
self.output_dim = self.num_bits_per_output_vector #vector_dim
self.total_parameters = ( 3 * self.memory_columns + 3 + self.shift_range.shape[0] )*(self.num_write_heads + self.num_read_heads)
self.PMG_Layer = tf.keras.layers.Dense(units= self.total_parameters, use_bias=True,) #PMG_Layer = Parameter Matrix GeneratingLayer CHECK INTITIALISATION OF PARAMETERS FOR IMPROVEMENT
self.NTM_ouput_gen_layer = tf.keras.layers.Dense(units= self.output_dim,use_bias = True)
@tf.function
def call(self, inputs, previous_states):
'''
inputs: shape = (Batch_size, input_features) where input_features is equal to num_bits_per_output_vector.
previous_states: dictionary, contains: 1. controller_state (list of two matrices, one for Memory and one for Carry, both of size [Batch_size, RNN_size]),
2. All_Read_vectors of size [Num_Read_Heads, Batch_size, Memory_dim(M)]
3. All_Weights of size [Num_ALL_Heads, Batch_size, Memory_size(N)]
4. Memory_Matrix of size (Batch_size, Memory_size(N), Memory_dim(M))
'''
#Since controller itself is a LSTMCell, thus it would demand a input of shape [Batch_size, features].
#We construct a controller whose input will be of size [Batch_size, features_for_controller]
#where features_for_controller is Num_Read_Heads * Memory_dim(M) + input_features
All_prev_read_vectors = previous_states['All_Read_vectors']
prev_controller_state = previous_states['controller_state']
M_prev = previous_states['Memory_Matrix']
w_prev = previous_states['All_Weight_vectors']
#^Of shape [num_total_heads, batch_size, N]
assert inputs.shape[1] == self.num_bits_per_output_vector
PRV = [All_prev_read_vectors[i] for i in range(All_prev_read_vectors.shape[0])] #PRV: Previous Read Vectors
PRV.insert(0, inputs)
controller_input = tf.concat(PRV, axis = 1)
assert controller_input.shape[1] == self.num_read_heads * self.memory_columns + inputs.shape[1]
controller_output, controller_state = self.controller(controller_input, prev_controller_state)
#controller_output is of the same shape as the controller_input
Parameter_Matrix = self.PMG_Layer(controller_output)
#Parameter_Matrix is of shape [Batch_size, self.total_parameters]
Each_Heads_PM_list = tf.split(Parameter_Matrix, self.num_read_heads + self.num_write_heads,axis = 1)
#Contains Each Head's Parameter matrix; is of total length self.num_read_heads + self.num_write_heads.
All_Heads_W_list = []
All_Heads_R_list = []
#To get the weights for each Head in the whole Batch
#To get the Read Vectors and Updated Memory Matrix, we assume first self.num_read_heads to be READ Heads and rest to be WRITE Heads
for i,Head_PM in enumerate(Each_Heads_PM_list):
k_t, beta_t, g_t, s_t, gamma_t, a_t, e_t = tf.split(Head_PM, [self.memory_columns, 1, 1, self.shift_range.shape[0], 1, self.memory_columns, self.memory_columns], axis = 1)
#EXPERIMENT WITH OTHER VALID COMBINATIONS OF THE BELOW USED ACTIVATIONS
#For k_t:-
k_t = tf.tanh(k_t )
#For beta_t:-
beta_t = tf.sigmoid(beta_t) * 10
#For g_t:-
g_t = tf.sigmoid(g_t)
#For s_t:-
s_t = tf.nn.softmax(s_t,axis = 1)
#The above s_t is one of the points where we can improve
#For gamma_t:-
gamma_t = tf.math.log(tf.exp(gamma_t) + 1 ) + 1
#For a_t:-
a_t = tf.tanh(a_t)
#For e_t:-
e_t = tf.sigmoid(e_t)
if self.addressing_type == 'LOC':
Heads_w_t = (Batch_Focusing.LocationFocusing( k_t, M_prev, beta_t, g_t, w_prev[i], s_t, gamma_t, K = None))
elif self.addressing_type == 'CONT':
Heads_w_t = (Batch_Focusing.ContentFocusing( k_t, M_prev, beta_t, K = None))
#^Should be of shape [batch_size,N]
if i<self.num_read_heads:
r_t = Batch_RWV_Generation.ReadVector(M_prev,Heads_w_t)
All_Heads_R_list.append(r_t)
elif i>=self.num_read_heads:
M_prev = Batch_RWV_Generation.WriteOnMemory(M_prev,Heads_w_t,e_t,a_t)
All_Heads_W_list.append(Heads_w_t)
#Please Note that at this point M_prev has been updated to the new weight Matrix
All_W_Matrix = tf.convert_to_tensor(All_Heads_W_list) #W for Weights
#^Of shape [num_total_heads, batch_size, N]
All_R_Matrix = tf.convert_to_tensor(All_Heads_R_list) #R for Read
#^Of Shape [num_Read_Heads, batch_size, M]
#TODO:: COMPLETE THE CONVOLUTION OPERATION IN FOCUSING AND THEN COMPLETE THIS CLASS
NTM_output = self.NTM_ouput_gen_layer(controller_output)
current_states = {
'All_Read_vectors' : All_R_Matrix,
'controller_state' : controller_state,
'Memory_Matrix' : M_prev,
'All_Weight_vectors' : All_W_Matrix
}
return NTM_output, current_states
@property
def state_size(self):
return {
'controller_state' : self.controller.state_size,
'All_Read_vectors' : tf.TensorShape((self.num_read_heads,None,self.memory_columns)),
'All_Weight_vectors' : tf.TensorShape(((self.total_num_heads, None, self.memory_rows))),
'Memory_Matrix' : tf.TensorShape([None,self.memory_rows,self.memory_columns])
}
@property
def output_size(self):
return self.output_dim
#CHANGE INITIAL STATES TO SOME OTHER VALUES AND OBSERVE WHETHER THE MODEL IMPROVES OR NOT
def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
# initial_state = {
# 'controller_state': [0.0000001 * tf.ones((batch_size,self.rnn_size)), 0.0000001 * tf.ones((batch_size,self.rnn_size))],
# 'All_Read_vectors': 0.0000001 * tf.ones((self.num_read_heads,batch_size,self.memory_columns)),
# 'All_Weight_vectors': 0.0000001 * tf.ones((self.total_num_heads, batch_size, self.memory_rows)),
# 'Memory_Matrix': 0.0000001 * tf.ones((batch_size,self.memory_rows,self.memory_columns))
# }
# initial_state = {
# 'controller_state': [tf.compat.v1.get_variable(name = 'controller_state_memory',shape=[batch_size, self.rnn_size], dtype = tf.float32, initializer=tf.random_uniform_initializer(minval=0.0000001, maxval=0.999999)), tf.compat.v1.get_variable(name = 'controller_state_carry',shape=[batch_size, self.rnn_size], dtype = tf.float32, initializer=tf.random_uniform_initializer(minval=0.0000001, maxval=0.999999))],
# 'All_Read_vectors': tf.compat.v1.get_variable(name = 'All_Read_vectors',shape=[self.num_read_heads, batch_size, self.memory_columns], dtype = tf.float32, initializer=tf.random_uniform_initializer(minval=0.0000001, maxval=0.999999)),
# 'All_Weight_vectors': tf.compat.v1.get_variable(name = 'All_Weight_vectors',shape=[self.total_num_heads, batch_size, self.memory_rows], dtype = tf.float32, initializer=tf.random_uniform_initializer(minval=0.0000001, maxval=0.999999)),
# 'Memory_Matrix': tf.compat.v1.get_variable(name = 'Memory_Matrix',shape=[batch_size, self.memory_rows, self.memory_columns], dtype = tf.float32, initializer=tf.random_uniform_initializer(minval=0.0000001, maxval=0.999999))
# }
initial_state = {
'controller_state': [tf.nn.tanh(tf.compat.v1.get_variable(name = 'controller_state_memory',shape=[batch_size, self.rnn_size], dtype = tf.float32, initializer=tf.random_normal_initializer(stddev = 0.5))), tf.nn.tanh(tf.compat.v1.get_variable(name = 'controller_state_carry',shape=[batch_size, self.rnn_size], dtype = tf.float32, initializer=tf.random_normal_initializer(stddev = 0.5)))],
'All_Read_vectors': tf.nn.tanh(tf.compat.v1.get_variable(name = 'All_Read_vectors',shape=[self.num_read_heads, batch_size, self.memory_columns], dtype = tf.float32, initializer=tf.random_normal_initializer(stddev = 0.5))),
'All_Weight_vectors': tf.nn.softmax(tf.compat.v1.get_variable(name = 'All_Weight_vectors',shape=[2, batch_size, self.memory_rows], dtype = tf.float32,initializer=tf.random_normal_initializer(stddev = 0.5)), axis = 2),
'Memory_Matrix': tf.tanh(tf.compat.v1.get_variable(name = 'Memory_Matrix',shape=[batch_size, self.memory_rows, self.memory_columns], dtype = tf.float32, initializer=tf.random_normal_initializer(stddev = 0.5)))
}
return initial_state