implemented padding and pipeline is complete

speech_data.py

@@ -95,17 +95,34 @@ def create_spectrogram_tfrecords(audio_group='audio'):
                 writer.write(example.SerializeToString())
     writer.close()
 
+def padd_zeros(spgr, max_samples):
+    return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
+                      'constant')
+
+def find_max_n(trf):
+    max_n = 0
+    max_n_it = tf.python_io.tf_record_iterator(path=trf)
+    for string_record in max_n_it:
+        example = tf.train.Example()
+        example.ParseFromString(string_record)
+        spec_n1 = example.features.feature['spec_n1'].int64_list.value[0]
+        spec_n2 = example.features.feature['spec_n2'].int64_list.value[0]
+        max_n = max([max_n,spec_n1,spec_n2])
+    return max_n
+
 def read_siamese_tfrecords(audio_group='audio'):
     records_file  = os.path.join('./outputs',audio_group+'.tfrecords')
     record_iterator = tf.python_io.tf_record_iterator(path=records_file)
+    # input1,input2 = [],[]
     input_pairs = []
     output_class = []
-    input_words = []
+    max_n = find_max_n(records_file)
+    spec_w1 = 0
     for string_record in record_iterator:
         example = tf.train.Example()
         example.ParseFromString(string_record)
-        word = example.features.feature['word'].bytes_list.value[0]
-        input_words.append(word)
+        # word = example.features.feature['word'].bytes_list.value[0]
+        # input_words.append(word)
         example.features.feature['spec2'].float_list.value[0]
         spec_n1 = example.features.feature['spec_n1'].int64_list.value[0]
         spec_n2 = example.features.feature['spec_n2'].int64_list.value[0]
@@ -113,10 +130,23 @@ def read_siamese_tfrecords(audio_group='audio'):
         spec_w2 = example.features.feature['spec_w2'].int64_list.value[0]
         spec1 = np.array(example.features.feature['spec1'].float_list.value).reshape(spec_n1,spec_w1)
         spec2 = np.array(example.features.feature['spec2'].float_list.value).reshape(spec_n2,spec_w2)
-        input_pairs.append([spec1,spec2])
+        p_spec1,p_spec2 = padd_zeros(spec1,max_n),padd_zeros(spec2,max_n)
+        # input1.append(spec1)
+        # input2.append(spec2)
+        input_pairs.append(np.asarray([p_spec1,p_spec2]))
+        # input_pairs.append([spec1,spec2])
         output = example.features.feature['output'].int64_list.value
-        output_class.append(output)
-    return input_pairs,output_class
+        output_class.append(np.asarray(output))
+        n_features = spec_w1
+        # if len(input_pairs) > 50:
+        #     break
+    input_data,output_data = np.asarray(input_pairs),np.asarray(output_class)
+    # import pdb; pdb.set_trace()
+    # tr_x1,te_x1,tr_x2,te_x2,tr_y,te_y = train_test_split(input1,input2,output_class)
+    tr_pairs,te_pairs,tr_y,te_y = train_test_split(input_data,output_data)
+    # return (tr_x1,te_x1,tr_x2,te_x2,tr_y,te_y)
+    n_step,n_features = int(max_n),int(spec_w1)
+    return (tr_pairs,te_pairs,tr_y,te_y,n_step,n_features)
 
 def audio_samples_word_count(audio_group='audio'):
     audio_group = 'story_all'
@@ -157,7 +187,8 @@ if __name__ == '__main__':
     # create_spectrogram_data()
     # create_spectrogram_data('story_words')
     # create_spectrogram_tfrecords('story_words')
-    create_spectrogram_tfrecords('story_all')
+    # create_spectrogram_tfrecords('story_all')
+    read_siamese_tfrecords('story_all')
     # create_padded_spectrogram()
     # create_speech_pairs_data()
     # print(speech_model_data())

speech_siamese.py

@@ -82,10 +82,10 @@ def siamese_model(input_dim):
 def train_siamese():
     # the data, shuffled and split between train and test sets
     # tr_pairs, te_pairs, tr_y_e, te_y_e = speech_model_data()
-    pairs,y = read_siamese_tfrecords('story_words')
+    (tr_pairs,te_pairs,tr_y,te_y,n_step,n_features) = read_siamese_tfrecords('story_words')
     # tr_y = to_categorical(tr_y_e, num_classes=2)
     # te_y = to_categorical(te_y_e, num_classes=2)
-    input_dim = (None, 1654)
+    input_dim = (n_step, n_features)
 
     model = siamese_model(input_dim)
 
@@ -114,11 +114,11 @@ def train_siamese():
     rms = RMSprop(lr=0.001)
     model.compile(loss=categorical_crossentropy, optimizer=rms, metrics=[accuracy])
     model.fit(
-        [tr_pairs[:, 0], tr_pairs[:, 1]],
+        [tr_x1, tr_x2],
         tr_y,
         batch_size=128,
         epochs=50,
-        validation_data=([te_pairs[:, 0], te_pairs[:, 1]], te_y),
+        validation_data=([tr_pairs[:, 0], tr_pairs[:, 1]], te_y),
         callbacks=[tb_cb, cp_cb])
 
     model.save('./models/siamese_speech_model-final.h5')