1. fixed softmax output and overfit the model for small sample

2. updated to run on complete data

segment_data.py

@@ -156,8 +156,8 @@ def create_segments_tfrecords(collection_name='story_test_segments',sample_count
             f_bounds = [spec_frame(phrase_spec,b) for b in ph_bounds]
             valid_bounds = [i for i in f_bounds if 0 < i < spec_n]
             b_frames = np.asarray(valid_bounds)
-            # print(spec_n,b_frames)
-            result[b_frames] = 1
+            if len(b_frames) > 0:
+                result[b_frames] = 1
             nonlocal n_records,n_spec,n_features
             n_spec = max([n_spec,spec_n])
             n_features = spec_w
@@ -178,9 +178,10 @@ def create_segments_tfrecords(collection_name='story_test_segments',sample_count
 
     word_groups = [i for i in audio_samples.groupby('phrase')]
     wg_sampled = reservoir_sample(word_groups,sample_count) if sample_count > 0 else word_groups
+    # write_samples(word_groups,'all')
     tr_audio_samples,te_audio_samples = train_test_split(wg_sampled,test_size=train_test_ratio)
     write_samples(tr_audio_samples,'train')
-    # write_samples(te_audio_samples,'test')
+    write_samples(te_audio_samples,'test')
     const_file = './outputs/segments/'+collection_name+'/constants.pkl'
     pickle.dump((n_spec,n_features,n_records),open(const_file,'wb'))
 
@@ -255,9 +256,9 @@ if __name__ == '__main__':
     # plot_random_phrases()
     # fix_csv('story_test_segments')
     # plot_segments('story_test_segments')
-    # fix_csv('story_phrases')
+    # fix_csv('story_words')
     # pass
-    create_segments_tfrecords('story_phrases', sample_count=100)
+    create_segments_tfrecords('story_words.30', sample_count=36,train_test_ratio=0.1)
     # record_generator,input_data,output_data,copy_read_consts = read_segments_tfrecords_generator('story_test')
     # tr_gen = record_generator()
     # for i in tr_gen:

segment_model.py

@@ -4,10 +4,10 @@ import numpy as np
 from keras.models import Model,load_model,model_from_yaml
 from keras.layers import Input,Concatenate,Lambda, Reshape, Dropout
 from keras.layers import Dense,Conv2D, LSTM, Bidirectional, GRU
-from keras.layers import BatchNormalization
+from keras.layers import BatchNormalization,Activation
 from keras.losses import categorical_crossentropy
 from keras.utils import to_categorical
-from keras.optimizers import RMSprop
+from keras.optimizers import RMSprop,Adadelta,Adagrad,Adam,Nadam
 from keras.callbacks import TensorBoard, ModelCheckpoint
 from keras import backend as K
 from keras.utils import plot_model
@@ -36,30 +36,38 @@ def ctc_lambda_func(args):
     return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
 
 def segment_model(input_dim):
-    # input_dim = (100,100,1)
     inp = Input(shape=input_dim)
     cnv1 = Conv2D(filters=32, kernel_size=(5,9))(inp)
     cnv2 = Conv2D(filters=1, kernel_size=(5,9))(cnv1)
     dr_cnv2 = Dropout(rate=0.95)(cnv2)
-    # dr_cnv2
     cn_rnn_dim = (dr_cnv2.shape[1].value,dr_cnv2.shape[2].value)
     r_dr_cnv2 = Reshape(target_shape=cn_rnn_dim)(dr_cnv2)
     b_gr1 = Bidirectional(GRU(512, return_sequences=True),merge_mode='sum')(r_dr_cnv2)
-    # b_gr1
     b_gr2 = Bidirectional(GRU(512, return_sequences=True),merge_mode='sum')(b_gr1)
     b_gr3 = Bidirectional(GRU(512, return_sequences=True),merge_mode='sum')(b_gr2)
-    # b_gr3
     oup = Dense(2, activation='softmax')(b_gr3)
-    # oup
     return Model(inp, oup)
 
 def simple_segment_model(input_dim):
     inp = Input(shape=input_dim)
-    b_gr1 = Bidirectional(GRU(256, return_sequences=True),merge_mode='sum')(inp)
-    # b_gr1
-    b_gr2 = Bidirectional(GRU(64, return_sequences=True),merge_mode='sum')(b_gr1)
-    b_gr3 = Bidirectional(GRU(1, return_sequences=True),merge_mode='sum')(b_gr2)
-    oup = Reshape(target_shape=(input_dim[0],))(b_gr3)
+    b_gr1 = Bidirectional(LSTM(32, return_sequences=True))(inp)
+    b_gr1 = Bidirectional(LSTM(16, return_sequences=True),merge_mode='sum')(b_gr1)
+    b_gr1 = LSTM(1, return_sequences=True,activation='softmax')(b_gr1)
+    # b_gr1 = LSTM(4, return_sequences=True)(b_gr1)
+    # b_gr1 = LSTM(2, return_sequences=True)(b_gr1)
+    # bn_b_gr1 = BatchNormalization(momentum=0.98)(b_gr1)
+    # b_gr2 = GRU(64, return_sequences=True)(b_gr1)
+    # bn_b_gr2 = BatchNormalization(momentum=0.98)(b_gr2)
+    # d1 = Dense(32)(b_gr2)
+    # bn_d1 = BatchNormalization(momentum=0.98)(d1)
+    # bn_da1 = Activation('relu')(bn_d1)
+    # d2 = Dense(8)(bn_da1)
+    # bn_d2 = BatchNormalization(momentum=0.98)(d2)
+    # bn_da2 = Activation('relu')(bn_d2)
+    # d3 = Dense(1)(b_gr1)
+    # # bn_d3 = BatchNormalization(momentum=0.98)(d3)
+    # bn_da3 = Activation('softmax')(d3)
+    oup = Reshape(target_shape=(input_dim[0],))(b_gr1)
     return Model(inp, oup)
 
 def write_model_arch(mod,mod_file):
@@ -75,7 +83,7 @@ def load_model_arch(mod_file):
 
 def train_segment(collection_name = 'test',resume_weights='',initial_epoch=0):
     # collection_name = 'story_test'
-    batch_size = 64
+    batch_size = 128
     # batch_size  = 4
     model_dir = './models/segment/'+collection_name
     create_dir(model_dir)
@@ -111,8 +119,8 @@ def train_segment(collection_name = 'test',resume_weights='',initial_epoch=0):
         mode='auto',
         period=1)
     # train
-    rms = RMSprop()
-    model.compile(loss=categorical_crossentropy, optimizer=rms, metrics=[accuracy])
+    opt = RMSprop()
+    model.compile(loss=categorical_crossentropy, optimizer=opt, metrics=[accuracy])
     write_model_arch(model,model_dir+'/speech_segment_model_arch.yaml')
     epoch_n_steps = step_count(n_records,batch_size)
     if resume_weights != '':
@@ -133,4 +141,4 @@ def train_segment(collection_name = 'test',resume_weights='',initial_epoch=0):
 
 if __name__ == '__main__':
     # pass
-    train_segment('story_phrases','./models/segment/story_phrases.1000/speech_segment_model-final.h5',1001)
+    train_segment('story_words')#,'./models/segment/story_phrases.1000/speech_segment_model-final.h5',1001)