1+ # -*- coding:utf-8 -*-
2+
3+ """
4+ @author: Songgx
5+ @file: 0504_tf_full_connect_NN.py
6+ @time: 11/30/16 3:44 PM
7+ """
8+
9+ from __future__ import print_function
10+
11+ import numpy as np
12+ import tensorflow as tf
13+
14+
15+ def csv_file_line_number (fname ):
16+ with open (fname , "r" ) as f :
17+ num = 0
18+ for line in f :
19+ num += 1
20+ return num
21+
22+
23+ def count_column_num (fname , field_delim ):
24+ with open (fname ) as f :
25+ line = f .readline ().split (field_delim )
26+ # the last column is the class number --> -1
27+ return len (line )
28+
29+
30+ def dense_to_one_hot (labels_dense , num_classes = 10 ):
31+ """Convert class labels from scalars to one-hot vectors."""
32+ num_labels = labels_dense .shape [0 ]
33+ index_offset = np .arange (num_labels ) * num_classes
34+ labels_one_hot = np .zeros ((num_labels , num_classes ))
35+ labels_one_hot .flat [index_offset + labels_dense .ravel ()] = 1
36+ return labels_one_hot
37+
38+
39+ def read_and_decode (filename ):
40+ filename_queue = tf .train .string_input_producer ([filename ])
41+
42+ reader = tf .TFRecordReader ()
43+ _ , serialized_example = reader .read (filename_queue )
44+ features = tf .parse_single_example (serialized_example ,
45+ features = {
46+ 'label' : tf .FixedLenFeature ([], tf .int64 ),
47+ # We know the length of both fields. If not the
48+ # tf.VarLenFeature could be used
49+ 'features' : tf .FixedLenFeature ([n_input ], tf .float32 ),
50+ })
51+
52+ X = tf .cast (features ['features' ], tf .float32 )
53+ y = tf .cast (features ['label' ], tf .int32 )
54+
55+ return X , y
56+
57+
58+ # Parameters
59+ learning_rate = 0.001
60+ training_epochs = 10000
61+ display_step = 50
62+ num_threads = 8
63+
64+ training_csv_file_path = "data/tvtsets/training_scat_data.txt"
65+ training_file_path = "data/tvtsets/training_scat_data.tfrecords"
66+ test_csv_file_path = "data/tvtsets/test_scat_data.txt"
67+ test_file_path = "data/tvtsets/test_scat_data.tfrecords"
68+ batch_size = 20
69+ column_num = count_column_num (training_csv_file_path , " " )
70+ # file_length = file_len(csv_file_path)
71+
72+ # Network Parameters
73+ n_input = 28 # data input 8660 * 1 = 433 * 20
74+ n_steps = 28 # timesteps
75+ n_hidden = 128 # hidden layer num of features
76+ n_classes = 10 # MNIST total classes (0-9 digits)
77+
78+
79+ n_input = column_num - 1
80+ n_classes = 10 # total classes (0-9 digits)
81+
82+ # tf Graph input
83+
84+ x = tf .placeholder (tf .float32 , [batch_size , n_steps , n_input ])
85+ y = tf .placeholder (tf .int32 , [batch_size ,])
86+
87+ # Define weights
88+ weights = {
89+ 'out' : tf .Variable (tf .random_normal ([n_hidden , n_classes ]))
90+ }
91+ biases = {
92+ 'out' : tf .Variable (tf .random_normal ([n_classes ]))
93+ }
94+
95+
96+ # Create model
97+ def LSTM (x , weights , biases ):
98+
99+ # Prepare data shape to match `rnn` function requirements
100+ # Current data input shape: (batch_size, n_steps, n_input)
101+ # Required shape: 'n_steps' tensors list of shape (batch_size, n_input)
102+
103+ # Permuting batch_size and n_steps
104+ x = tf .transpose (x , [1 , 0 , 2 ])
105+ # Reshaping to (n_steps*batch_size, n_input)
106+ x = tf .reshape (x , [- 1 , n_input ])
107+ # Split to get a list of 'n_steps' tensors of shape (batch_size, n_input)
108+ x = tf .split (0 , n_steps , x )
109+
110+ # Define a lstm cell with tensorflow
111+ lstm_cell = rnn_cell .BasicLSTMCell (n_hidden , forget_bias = 1.0 )
112+
113+ # Get lstm cell output
114+ outputs , states = rnn .rnn (lstm_cell , x , dtype = tf .float32 )
115+
116+ # Linear activation, using rnn inner loop last output
117+ return tf .matmul (outputs [- 1 ], weights ['out' ]) + biases ['out' ]
118+
119+
120+
121+ # Construct model
122+ pred = LSTM (x , weights , biases )
123+
124+ # Define loss and optimizer & correct_prediction
125+ cost = tf .reduce_mean (tf .nn .sparse_softmax_cross_entropy_with_logits (pred , y ))
126+ optimizer = tf .train .AdamOptimizer (learning_rate = learning_rate ).minimize (cost )
127+
128+ # Launch the graph
129+
130+ audio , label = read_and_decode (training_file_path )
131+ audio_test , label_test = read_and_decode (test_file_path )
132+
133+
134+ #使用shuffle_batch可以随机打乱输入
135+ audio_batch , label_batch = tf .train .shuffle_batch ([audio , label ],
136+ batch_size = batch_size , capacity = 2000 ,
137+ min_after_dequeue = 1000 )
138+
139+ # Initializing the variables
140+ init = tf .global_variables_initializer ()
141+
142+ # Launch the graph
143+ with tf .Session () as sess :
144+ sess .run (init )
145+
146+ # Start input enqueue threads.
147+ coord = tf .train .Coordinator ()
148+ threads = tf .train .start_queue_runners (sess = sess , coord = coord )
149+ for epoch in range (10000 ):
150+ # pass it in through the feed_dict
151+ audio_batch_vals , label_batch_vals = sess .run ([audio_batch , label_batch ])
152+
153+ _ , loss_val = sess .run ([optimizer , cost ], feed_dict = {x :audio_batch_vals , y :label_batch_vals })
154+ if (epoch + 1 ) % display_step == 0 :
155+ print ("Epoch:" , '%06d' % (epoch + 1 ), "cost=" , "{:.9f}" .format (loss_val ))
156+
157+ print ("Training finished." )
158+ coord .request_stop ()
159+ coord .join (threads )
160+
161+ # Test model
162+ # Calculate accuracy
163+ test_example_number = csv_file_line_number (test_csv_file_path )
164+ correct_num = 0
165+ for _ in range (test_example_number ):
166+ audio_test_val , label_test_val = sess .run ([audio_test , label_test ])
167+ audio_test_val_vector = np .array ([audio_test_val ])
168+ test_pred = multilayer_perceptron (audio_test_val_vector , weights , biases )
169+ '''
170+ print (sess.run([test_pred]))
171+
172+ [array([[ 11.13519478, 12.56501865, 14.68154907, 2.02798128,
173+ -5.89219952, -0.76298785, 0.46614531, 5.27717066,
174+ 7.54774714, 7.12729597]], dtype=float32)]
175+ '''
176+ pred_class_index = sess .run (tf .argmax (test_pred , 1 ))
177+ '''
178+ print (sess.run(tf.argmax(test_pred, 1)))
179+
180+ [8]
181+ 0
182+ [5]
183+ 0
184+ [9]
185+ 0
186+ [8]
187+ 0
188+ ....
189+
190+ [9]
191+ 9
192+ [9]
193+ 9
194+ [4]
195+ 9
196+ [1]
197+ 9
198+ '''
199+
200+ if label_test_val == pred_class_index [0 ]:
201+ correct_num += 1
202+ print ("%i / %i is correct." % (correct_num , test_example_number ))
203+ print ("Accuracy is %f ." % (float (correct_num ) / test_example_number ))
204+ sess .close ()
0 commit comments