1+ # -*- coding:utf-8 -*-
2+
3+ """
4+ @author: Songgx
5+ @file: 1200_mtt_2dCNN.py
6+ @time: 2017/2/7 14:55
7+ """
8+
9+ from __future__ import print_function
10+
11+ import numpy as np
12+ import tensorflow as tf
13+ from sklearn .metrics import roc_auc_score
14+
15+ # https://indico.io/blog/tensorflow-data-inputs-part1-placeholders-protobufs-queues/
16+ top_50_tags_index = np .loadtxt ('data/top_50_tags.txt' , delimiter = ',' , skiprows = 0 , dtype = int )
17+
18+ # Parameters
19+ x_height = 96
20+ x_width = 1366
21+
22+ # 总共的tag数
23+ n_total_tags = 50
24+ learning_rate = 0.00001
25+ training_epochs = 1000 * 200 # 1000 * 200 iterations, 200 epochs
26+ display_step = 100
27+ num_threads = 8
28+ dropout = 0.5
29+ #L2_norm = 1e-9
30+ batch_size = 12
31+
32+
33+ def read_and_decode (filename ):
34+ filename_queue = tf .train .string_input_producer ([filename ])
35+
36+ reader = tf .TFRecordReader ()
37+ _ , serialized_example = reader .read (filename_queue )
38+ features = tf .parse_single_example (serialized_example ,
39+ features = {
40+ 'features_mel' : tf .FixedLenFeature ([], tf .string ),
41+ 'label' : tf .FixedLenFeature ([n_total_tags ], tf .float32 ),
42+ })
43+
44+ x = tf .decode_raw (features ['features_mel' ], tf .float32 )
45+ x = tf .reshape (x , [x_height , x_width , 1 ])
46+ y = tf .cast (features ['label' ], tf .float32 )
47+ return x , y
48+
49+
50+ def load_and_shuffle_to_batch_data (path , batch_size = batch_size ):
51+ features , label = read_and_decode (path )
52+ # 使用shuffle_batch可以随机打乱输入
53+ audio_batch , label_batch = tf .train .shuffle_batch ([features , label ],
54+ batch_size = batch_size , capacity = 2000 ,
55+ min_after_dequeue = 1000 )
56+ return audio_batch , label_batch
57+
58+
59+ # tf Graph input
60+ x = tf .placeholder (tf .float32 , (batch_size , x_height , x_width , 1 ), name = 'input_layer' )
61+ y = tf .placeholder (tf .float32 , (batch_size , n_total_tags ), name = 'output_layer' )
62+ keep_prob = tf .placeholder (tf .float32 ) #dropout (keep probability)
63+ # phase_train = tf.placeholder(tf.bool, name='phase_train')
64+
65+
66+ def batch_norm (x , n_out , phase_train , scope = 'bn' ):
67+ with tf .variable_scope (scope ):
68+ beta = tf .Variable (tf .constant (0.0 , shape = [n_out ]),name = 'beta' , trainable = True )
69+ gamma = tf .Variable (tf .constant (1.0 , shape = [n_out ]),name = 'gamma' , trainable = True )
70+ batch_mean , batch_var = tf .nn .moments (x , [0 ,1 ,2 ], name = 'moments' )
71+ ema = tf .train .ExponentialMovingAverage (decay = 0.5 )
72+
73+ def mean_var_with_update ():
74+ ema_apply_op = ema .apply ([batch_mean , batch_var ])
75+ with tf .control_dependencies ([ema_apply_op ]):
76+ return tf .identity (batch_mean ), tf .identity (batch_var )
77+
78+ mean , var = tf .cond (phase_train ,
79+ mean_var_with_update ,
80+ lambda : (ema .average (batch_mean ), ema .average (batch_var )))
81+ normed = tf .nn .batch_normalization (x , mean , var , beta , gamma , 1e-3 )
82+ return normed
83+
84+
85+ # Create model
86+ def conv_net (x , weights , phase_train = np .array (True )):
87+
88+ conv2_1 = tf .add (tf .nn .conv2d (x , weights ['wconv1' ], strides = [1 , 1 , 1 , 1 ], padding = 'SAME' ), weights ['bconv1' ])
89+ conv2_1 = tf .nn .relu (batch_norm (conv2_1 , 128 , phase_train ))
90+ mpool_1 = tf .nn .max_pool (conv2_1 , ksize = [1 , 2 , 4 , 1 ], strides = [1 , 2 , 4 , 1 ], padding = 'VALID' )
91+ dropout_1 = tf .nn .dropout (mpool_1 , 0.5 )
92+
93+ conv2_2 = tf .add (tf .nn .conv2d (dropout_1 , weights ['wconv2' ], strides = [1 , 1 , 1 , 1 ], padding = 'SAME' ), weights ['bconv2' ])
94+ conv2_2 = tf .nn .relu (batch_norm (conv2_2 , 384 , phase_train ))
95+ mpool_2 = tf .nn .max_pool (conv2_2 , ksize = [1 , 4 , 5 , 1 ], strides = [1 , 4 , 5 , 1 ], padding = 'VALID' )
96+ dropout_2 = tf .nn .dropout (mpool_2 , 0.5 )
97+
98+ conv2_3 = tf .add (tf .nn .conv2d (dropout_2 , weights ['wconv3' ], strides = [1 , 1 , 1 , 1 ], padding = 'SAME' ), weights ['bconv3' ])
99+ conv2_3 = tf .nn .relu (batch_norm (conv2_3 , 768 , phase_train ))
100+ mpool_3 = tf .nn .max_pool (conv2_3 , ksize = [1 , 3 , 8 , 1 ], strides = [1 , 3 , 8 , 1 ], padding = 'VALID' )
101+ dropout_3 = tf .nn .dropout (mpool_3 , 0.5 )
102+
103+ conv2_4 = tf .add (tf .nn .conv2d (dropout_3 , weights ['wconv4' ], strides = [1 , 1 , 1 , 1 ], padding = 'SAME' ), weights ['bconv4' ])
104+ conv2_4 = tf .nn .relu (batch_norm (conv2_4 , 2048 , phase_train ))
105+ mpool_4 = tf .nn .max_pool (conv2_4 , ksize = [1 , 4 , 8 , 1 ], strides = [1 , 4 , 8 , 1 ], padding = 'VALID' )
106+ dropout_4 = tf .nn .dropout (mpool_4 , 0.5 )
107+
108+ flat = tf .reshape (dropout_4 , [- 1 , weights ['woutput' ].get_shape ().as_list ()[0 ]])
109+ fc_out = tf .nn .sigmoid (tf .add (tf .matmul (flat , weights ['woutput' ]), weights ['boutput' ]))
110+
111+ return fc_out
112+
113+ # Store layers weight & bias
114+ def init_weights (shape ):
115+ return tf .Variable (tf .random_normal (shape , stddev = 0.01 ))
116+
117+
118+ def init_biases (shape ):
119+ return tf .Variable (tf .zeros (shape ))
120+
121+
122+ def get_roc_auc_scores (tags , logits ):
123+ final_acc = 0.
124+ num = batch_size
125+ for i in range (batch_size ):
126+ cur_tag_array = tags [i ]
127+ cur_logits_array = logits [i ]
128+ if is_zeros (cur_tag_array ):
129+ if num == 1 :
130+ continue
131+ else :
132+ num = num - 1
133+ continue
134+ roc_auc = roc_auc_score (cur_tag_array , cur_logits_array )
135+ final_acc += roc_auc
136+ return final_acc / num
137+
138+
139+ def is_zeros (arr ):
140+ for element in arr :
141+ if element != 0 :
142+ return False
143+ return True
144+
145+ weights = {
146+ 'wconv1' : init_weights ([3 , 3 , 1 , 128 ]),
147+ 'wconv2' : init_weights ([3 , 3 , 128 , 384 ]),
148+ 'wconv3' : init_weights ([3 , 3 , 384 , 768 ]),
149+ 'wconv4' : init_weights ([3 , 3 , 768 , 2048 ]),
150+ 'bconv1' : init_biases ([128 ]),
151+ 'bconv2' : init_biases ([384 ]),
152+ 'bconv3' : init_biases ([768 ]),
153+ 'bconv4' : init_biases ([2048 ]),
154+ 'woutput' : init_weights ([2048 , 50 ]),
155+ 'boutput' : init_biases ([50 ])
156+ }
157+
158+
159+ # Construct model
160+ logits = conv_net (x , weights )
161+
162+
163+ # Define loss and optimizer & correct_prediction
164+
165+ # NaN bug
166+ #cross_entropy = -tf.reduce_sum(y * tf.log(tf.clip_by_value(logits, 1e-10, 1.0)))
167+
168+ # cross_entropy_loss with L2 norm
169+ # cross_entropy_loss = -tf.reduce_sum(y * tf.log(logits) + L2_norm * tf.nn.l2_loss(weights['wd1']))
170+ cross_entropy_loss = tf .reduce_sum (tf .nn .softmax_cross_entropy_with_logits (logits , y ))
171+
172+ optimizer = tf .train .AdamOptimizer (learning_rate = learning_rate ).minimize (cross_entropy_loss )
173+ # load data
174+ audio_batch_training , label_batch_training = load_and_shuffle_to_batch_data ("data/merge/mtt_mel_training_filtered.tfrecords" , batch_size )
175+ audio_batch_validation , label_batch_validation = load_and_shuffle_to_batch_data ("data/merge/mtt_mel_validation_filtered.tfrecords" , batch_size )
176+ audio_batch_test , label_batch_test = load_and_shuffle_to_batch_data ("data/merge/mtt_mel_test_filtered.tfrecords" , batch_size )
177+
178+ # Launch the graph
179+ # Initializing the variables
180+ init = tf .global_variables_initializer ()
181+ saver = tf .train .Saver ()
182+
183+ # Launch the graph
184+ with tf .Session () as sess :
185+ sess .run (init )
186+
187+ # Start input enqueue threads.
188+ coord = tf .train .Coordinator ()
189+ threads = tf .train .start_queue_runners (sess = sess , coord = coord )
190+ # for epoch in range(int(8000/batch_size)):
191+ valdation_accuracy_final = 0.
192+ for epoch in range (training_epochs ):
193+ # pass it in through the feed_dict
194+ audio_batch_vals_training , label_batch_vals_training = sess .run ([audio_batch_training , label_batch_training ])
195+ _ , loss_val , pred_ = sess .run ([optimizer , cross_entropy_loss , logits ], feed_dict = {x :audio_batch_vals_training , y :label_batch_vals_training , keep_prob : dropout })
196+
197+ #print("Epoch:", '%06d' % (epoch + 1), "cost=", "{:.9f}".format(loss_val))
198+ #print(pred_, label_batch_vals_training)
199+
200+ # calculate accuracy at each display step
201+ if (epoch + 1 ) % display_step == 0 :
202+ validation_iterations = 100
203+ cur_validation_acc = 0.
204+ for _ in range (validation_iterations ):
205+ audio_batch_validation_vals , label_batch_validation_vals = sess .run ([audio_batch_validation , label_batch_validation ])
206+
207+ logits_validation , loss_val_validation = sess .run ([logits , cross_entropy_loss ], feed_dict = {
208+ x : audio_batch_validation_vals ,
209+ y : label_batch_validation_vals ,
210+ keep_prob : 1.0
211+ })
212+ validation_accuracy = get_roc_auc_scores (label_batch_validation_vals , logits_validation )
213+ cur_validation_acc += validation_accuracy
214+
215+ cur_validation_acc /= validation_iterations
216+ print ("iter %d, training loss: %f, validation accuracy: %f" % ((epoch + 1 ), loss_val , cur_validation_acc ))
217+ print ("######### Training finished. #########" )
218+
219+ # Test model
220+ # batch_test --> reduce_mean --> final_test_accuracy
221+
222+ test_epochs = 400
223+ test_accuracy_final = 0.
224+ for _ in range (test_epochs ):
225+ audio_test_vals , label_test_vals = sess .run ([audio_batch_test , label_batch_test ])
226+ logits_test , test_loss_val = sess .run ([logits , cross_entropy_loss ], feed_dict = {x : audio_test_vals , y :label_test_vals , keep_prob : 1.0 })
227+ test_accuracy = get_roc_auc_scores (label_test_vals , logits_test )
228+ test_accuracy_final += test_accuracy
229+ print ("test epoch: %d, test loss: %f, test accuracy: %f" % (_ , test_loss_val , test_accuracy ))
230+ test_accuracy_final /= test_epochs
231+ print ("final test accuracy: %f" % test_accuracy_final )
232+
233+ coord .request_stop ()
234+ coord .join (threads )
235+ sess .close ()
236+
237+ # console results is in FCN-5.pdf
0 commit comments