-
Notifications
You must be signed in to change notification settings - Fork 20
/
GeneralSoftmaxModel.py
268 lines (219 loc) · 11.9 KB
/
GeneralSoftmaxModel.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
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
import tensorflow as tf
class GeneralSoftmaxModel:
""" General fully connected softmax neural network for classification of input data.
It contains multiple hidden layers (provided by the user) with
ReLU activation functions and the output layer is Softmax. Dropout is added for training as well. """
def __init__(self, initial_weights=None, model_size=None, training_algorithm='GD', learning_rate=0.1,
training_parameter=0.9, regulize=False, regularization_gain=5e-4, computation_precision=tf.float32):
self._computation_precision = computation_precision
self._graph = None
self._sess = None
self._trainOp = None
self._initializer = None
self._accuracy = None
# parameters of the neural network
self._input = None
self._output = None
self._target = None
self._gradients = None
self._keep_prob = None
self._nn_signals = []
self._nn_weights = []
self._nn_biases = []
self._num_parameters = 0
if initial_weights is not None:
self._create_initialized_graph(initial_weights)
elif model_size is not None:
self._create_random_graph(model_size)
else:
raise ValueError('Network size is not given.')
self._define_optimizer(training_algorithm, learning_rate, training_parameter, regulize, regularization_gain)
self._num_parameters = len(self._nn_weights)
self._sess = tf.Session(graph=self._graph)
# =========================================================================
# Create the computational graph, including:
# 1- Neural Network (placeholders, weights and output)
# 2- Initializer
# 3- Values and Gradients of the parameters of the NN
# 4- Optimizer
def _create_initialized_graph(self, initial_weights):
if self._graph is not None:
return
num_layers = len(initial_weights)
input_len = initial_weights[0][0].shape[0]
output_len = initial_weights[-1][0].shape[1]
self._graph = tf.Graph()
with self._graph.as_default():
self._x = tf.placeholder(self._computation_precision, [None, input_len])
self._target = tf.placeholder(self._computation_precision, [None, output_len])
self._keep_prob = tf.placeholder(tf.float32, 1)
# create layers
self._nn_signals = [self._x]
self._nn_weights = []
self._nn_biases = []
out_dim = input_len # number of output nodes from previous layer
output_signal = self._x
output_signal_drop = self._x # no dropout for the input signal
for h in range(num_layers):
init_w = initial_weights[h][0]
init_b = initial_weights[h][1]
if (init_w.shape[0] != out_dim) or (init_w.shape[1] != init_b.shape):
raise ValueError('Inconsistent dimensions for initial weights.')
W = tf.Variable(init_w)
b = tf.Variable(init_b)
output_signal = tf.matmul(output_signal_drop, W) + b
if h == num_layers - 1:
# final layer is Softmax
output_signal = tf.nn.softmax(output_signal)
else:
# hidden layers are ReLU
output_signal = tf.nn.relu(output_signal)
output_signal_drop = tf.nn.dropout(output_signal, self._keep_prob)
out_dim = init_w.shape[1]
self._nn_weights += [W]
self._nn_biases += [b]
self._nn_signals += [output_signal]
self._output = output_signal
# =========================================================================
def _create_random_graph(self, model_size):
if self._graph is not None:
return
num_layers = len(model_size) - 1
input_len = model_size[0]
output_len = model_size[-1]
self._graph = tf.Graph()
with self._graph.as_default():
self._x = tf.placeholder(self._computation_precision, [None, input_len])
self._target = tf.placeholder(self._computation_precision, [None, output_len])
self._keep_prob = tf.placeholder(self._computation_precision)
# create layers
self._nn_signals = [self._x]
self._nn_weights = []
self._nn_biases = []
output_signal = self._x
output_signal_drop = self._x # no dropout for the input signal
for h in range(num_layers):
in_dim = model_size[h] # number of input nodes to the hidden layer
out_dim = model_size[h + 1] # number of output nodes of the hidden layer
W = tf.Variable(tf.truncated_normal([in_dim, out_dim], 0, 0.1, dtype=self._computation_precision))
b = tf.Variable(0.1 * tf.ones([out_dim], dtype=self._computation_precision))
output_signal = tf.matmul(output_signal_drop, W) + b
if h == num_layers - 1:
# final layer is Softmax
output_signal = tf.nn.softmax(output_signal)
else:
# hidden layers are ReLU
output_signal = tf.nn.relu(output_signal)
output_signal_drop = tf.nn.dropout(output_signal, self._keep_prob)
self._nn_weights += [W]
self._nn_biases += [b]
self._nn_signals += [output_signal]
self._output = output_signal
# =========================================================================
def _define_optimizer(self, training_algorithm, learning_rate, training_parameter=0.9, regulize=False, gain=1e-4):
with self._graph.as_default():
# the cost function
cross_entropy = tf.reduce_mean(
tf.losses.softmax_cross_entropy(logits=self._output, onehot_labels=self._target))
if regulize:
# regulization of the W coefficients. it does not include the biases, b
regulizer_cost = 0
for w in self._nn_weights:
regulizer_cost += tf.nn.l2_loss(w)
cost = cross_entropy + gain * regulizer_cost
else:
cost = cross_entropy
# =================================================================
# define the appropriate optimizer to use
if (training_algorithm == 0) or (training_algorithm == 'GD'):
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
elif (training_algorithm == 1) or (training_algorithm == 'RMSProp'):
optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate)
elif (training_algorithm == 2) or (training_algorithm == 'Adam'):
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
elif (training_algorithm == 3) or (training_algorithm == 'AdaGrad'):
optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate)
elif (training_algorithm == 4) or (training_algorithm == 'AdaDelta'):
optimizer = tf.train.AdadeltaOptimizer(learning_rate=learning_rate)
elif (training_algorithm == 5) or (training_algorithm == 'Momentum'):
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=training_parameter)
else:
raise ValueError("Unknown training algorithm.")
# =================================================================
# training and initialization operators
parameter_list = self._nn_weights + self._nn_biases
gv = optimizer.compute_gradients(cost, var_list=parameter_list)
self._gradients = [g for (g, _) in gv]
self._trainOp = optimizer.minimize(cost, var_list=parameter_list)
self._initializer = tf.global_variables_initializer()
# =================================================================
# update (assign) operator for the parameters of the NN model
self._weight_assign_op = []
self._weight_placeholders = ()
self._bias_assign_op = []
self._bias_placeholders = ()
for w in self._nn_weights:
holder = tf.placeholder(dtype=self._computation_precision, shape=w.get_shape())
assign_op = w.assign(holder)
self._weight_assign_op.append(assign_op)
self._weight_placeholders = self._weight_placeholders + (holder,)
for b in self._nn_biases:
holder = tf.placeholder(dtype=self._computation_precision, shape=b.get_shape())
assign_op = b.assign(holder)
self._bias_assign_op.append(assign_op)
self._bias_placeholders = self._bias_placeholders + (holder,)
# =========================================================================
# Number of parameter pairs (W, b)
@property
def num_parameters(self):
return self._num_parameters
# =========================================================================
# Accuracy rate of the NN (defined as property)
@property
def accuracy(self):
if self._accuracy is None:
with self._graph.as_default():
matches = tf.equal(
tf.argmax(self._target, 1), tf.argmax(self._output, 1))
self._accuracy = tf.reduce_mean(tf.cast(matches, self._computation_precision))
return self._accuracy
# =========================================================================
# Initialize the computation graph
def initialize(self):
if self._initializer is not None:
self._sess.run(self._initializer)
else:
raise ValueError('Initializer has not been set.')
# =========================================================================
# One iteration of the training algorithm with input data
def train(self, batch_x, batch_y, keep_prob=1.0):
if self._trainOp is not None:
self._sess.run(self._trainOp,
feed_dict={self._x: batch_x, self._target: batch_y, self._keep_prob: keep_prob})
else:
raise ValueError('Training algorithm has not been set.')
# =========================================================================
# Get values of the parameters of the NN
def get_weights(self):
return self._sess.run(self._nn_weights), self._sess.run(self._nn_biases)
# =========================================================================
# Set values of the parameters of the NN
def set_weights(self, new_weights=None, new_biases=None):
if new_weights is not None:
self._sess.run(self._weight_assign_op, {self._weight_placeholders: tuple(new_weights)})
if new_biases is not None:
self._sess.run(self._bias_assign_op, {self._bias_placeholders: tuple(new_biases)})
# =========================================================================
# Compute the gradients of the parameters of the NN for the given input
def compute_gradients(self, x, target):
return self._sess.run(self._gradients,
feed_dict={self._x: x, self._target: target})
# =========================================================================
# Compute the accuracy of the NN using the given inputs
def compute_accuracy(self, x, target):
return self._sess.run(self.accuracy,
feed_dict={self._x: x, self._target: target, self._keep_prob: 1.0})
# =========================================================================
# Compute the signals in the NN
def compute_signals(self, x):
return self._sess.run(self._nn_signals, feed_dict={self._x: x, self._keep_prob: 1.0})