PySINN/sinn_model.py at master · Philliams/PySINN · GitHub

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
# SINN model with k-means initiation
from sklearn.cluster import KMeans
import numpy as np
from keras.layers import Input, Activation, Dropout
from keras.models import Model
from sinn_layers import Metric, Shepard

class SINN_model:

	def __init__(self, input_matrix, output_matrix, num_of_clusters, nodes_per_cluster, batch_size, epochs, mode='classification', verbose=2, validation_split=0.05, shuffle_data=True, model_optimizer='RMSprop', loss_metric='mse', initiation_mode='k_means'):

		# keep training parameters
		self.batch_size 		= batch_size
		self.epochs  			= epochs
		self.verbose 		  	= verbose
		self.validation_split 	= validation_split
		self.shuffle_data		= shuffle_data


		(separated_data,separated_labels) = self.separate_data_by_class(input_matrix,output_matrix)

		( initiation_data , initiation_labels ) = ([],[])

		if( initiation_mode == 'k_means'):
			(initiation_data , initiation_labels) = self.k_means_initiation(separated_data,separated_labels,num_of_clusters,nodes_per_cluster)
		elif (initiation_mode == 'distributed'):
			(initiation_data,initiation_labels) = self.distributed_initiation(separated_data,separated_labels,num_of_clusters*nodes_per_cluster)

		print('number of encoded nodes : ' + str(len(initiation_data)))

		# build model from diverse dataset picked using k-means
		input_size = len(input_matrix[0])
		input_layer = Input(shape=(input_size,))

		# metric layer
		m = Metric(initiation_data)(input_layer)
		# d = Dropout(0.2)(m)
		# shepard layer
		s = Shepard(initiation_labels)(m)

		if(mode == 'classification'):

			# softmax = Dense(len(output_matrix[0]), activation='softmax')(s)
			softmax = Activation('softmax')(s)
			self.model = Model(input=input_layer, output=softmax)
			self.model.compile(optimizer=model_optimizer,loss=loss_metric,metrics=['categorical_accuracy'])

		else:

			self.model = Model(input=input_layer, output=s)
			self.model.compile(optimizer=model_optimizer,loss=loss_metric)

		print('model complete...')

	def k_means_initiation(self,separated_data,separated_labels,num_of_clusters,nodes_per_cluster):

		initiation_data   = []
		initiation_labels = []

		for i in range(len(separated_data)):

			(metric_initiation,shepard_initiation) = self.get_clusters(num_of_clusters,nodes_per_cluster,separated_data[i],separated_labels[i])

			initiation_data   += metric_initiation
			initiation_labels += shepard_initiation

		return (initiation_data,initiation_labels)


	def distributed_initiation(self,separated_data,separated_labels,nodes_per_class):

		initiation_data   = []
		initiation_labels = []

		for i in range(len(separated_data)):

			for j in range( min( nodes_per_class , len(separated_data[i]) )):

				initiation_data.append(separated_data[i][j])
				initiation_labels.append(separated_labels[i][j])

		return (initiation_data,initiation_labels)


	def get_clusters(self,num_of_clusters,nodes_per_cluster,input_matrix,output_matrix):
		# build k-means clusters
		print('building k-means clusters...')
		kmeans  = KMeans(n_clusters=num_of_clusters, random_state=0, max_iter=1000,n_init=20).fit(input_matrix)
		print('done...')
		centers = kmeans.cluster_centers_
		labels  = kmeans.labels_

		# seperate labelled data by cluster, a maximum of nodes_per_cluster data points are kept for each cluster
		cluster_nodes = []
		cluster_topology = []

		print('separating clusters...')

		for i in range(num_of_clusters):
			cluster_nodes.append([])
			cluster_topology.append([])

		for i in range(len(input_matrix)):

			index = labels[i]

			if(len(cluster_nodes[index]) < nodes_per_cluster):
				cluster_nodes[index].append(input_matrix[i])
				cluster_topology[index].append(output_matrix[i])

			if(self.is_matrix_valid(nodes_per_cluster,cluster_nodes)):
				break

		print('done...')
		print('building initiation matrices...')
		metric_initiation  = []
		shepard_initiation = []

		for i in range(len(cluster_nodes)):
			for j in range( min(len(cluster_nodes[i]),nodes_per_cluster)):
				metric_initiation.append(cluster_nodes[i][j])
				shepard_initiation.append(cluster_topology[i][j])

		return (metric_initiation,shepard_initiation)

	def separate_data_by_class(self, input_matrix, output_matrix):

		separated_data   = []
		separated_labels = []

		for i in range(len(output_matrix[0])):
			separated_data.append([])
			separated_labels.append([])

		for i in range(len(input_matrix)):

			index = np.argmax(output_matrix[i])

			separated_data[index].append(input_matrix[i])
			separated_labels[index].append(output_matrix[i])

		return (separated_data,separated_labels)

	def predict(self, input_matrix):

		return self.model.predict_on_batch(input_matrix)

	def train_model(self, input_matrix, output_matrix, input_test=None, output_test=None):

		print('training starting...')

		if( input_test != None):
			self.model.fit( input_matrix, output_matrix , batch_size=self.batch_size, nb_epoch=self.epochs, verbose=self.verbose, validation_data=(input_test,output_test), shuffle=self.shuffle_data)
		else:
			self.model.fit( input_matrix, output_matrix , batch_size=self.batch_size, nb_epoch=self.epochs, verbose=self.verbose, validation_split=self.validation_split, shuffle=self.shuffle_data)

	def evaluate_model(self,input_matrix,output_matrix):

		return self.model.evaluate(input_matrix, output_matrix, batch_size=self.batch_size, verbose=self.verbose)

	def is_matrix_valid(self,k,m):

		# check if matrix is not jagged (i.e. rectangular or square)
		for i in range(len(m)):
			if(len(m[i]) != k):
				return False

		return True