resuming segment training

segment_data.py

@@ -113,7 +113,9 @@ def plot_segments(collection_name = 'story_test_segments'):
 def generate_spec(aiff_file):
     phrase_sample = pm_snd(aiff_file)
     phrase_spec = phrase_sample.to_spectrogram(window_length=SPEC_WINDOW_SIZE, maximum_frequency=SPEC_MAX_FREQUENCY)
-    sg_db = 10 * np.log10(phrase_spec.values)
+    sshow_abs = np.abs(phrase_spec.values + np.finfo(phrase_spec.values.dtype).eps)
+    sg_db = 10 * np.log10(sshow_abs)
+    sg_db[sg_db < 0] = 0
     return sg_db,phrase_spec
 
 
@@ -142,6 +144,7 @@ def create_segments_tfrecords(collection_name='story_test_segments',sample_count
             fname = g.iloc[0]['filename']
             sg_db,phrase_spec = generate_spec(g.iloc[0]['file_path'])
             phon_stops = []
+            phrase_groups.set_postfix(phrase=ph)
             spec_n,spec_w = sg_db.shape
             spec = sg_db.reshape(-1)
             for (i,phon) in g.iterrows():
@@ -177,7 +180,7 @@ def create_segments_tfrecords(collection_name='story_test_segments',sample_count
     wg_sampled = reservoir_sample(word_groups,sample_count) if sample_count > 0 else word_groups
     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'))
 
@@ -235,10 +238,8 @@ def read_segments_tfrecords_generator(collection_name='audio',batch_size=32,test
     random_samples = enumerate(reservoir_sample(te_re_iterator,test_size))
     for (i,string_record) in tqdm(random_samples,total=test_size):
         # (i,string_record) = next(random_samples)
-        # string_record
         example = tf.train.Example()
         example.ParseFromString(string_record)
-        # example.features.feature['spec'].float_list.value
         spec_n = example.features.feature['spec_n'].int64_list.value[0]
         spec_w = example.features.feature['spec_w'].int64_list.value[0]
         spec = np.array(example.features.feature['spec'].float_list.value).reshape(spec_n,spec_w)
@@ -254,9 +255,9 @@ if __name__ == '__main__':
     # plot_random_phrases()
     # fix_csv('story_test_segments')
     # plot_segments('story_test_segments')
-    # fix_csv('story_test')
+    # fix_csv('story_phrases')
-    pass
+    # pass
-    # create_segments_tfrecords('story_test')
+    create_segments_tfrecords('story_phrases', sample_count=100)
     # 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

@@ -36,7 +36,7 @@ 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)
+    # 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)
@@ -54,8 +54,6 @@ def segment_model(input_dim):
     return Model(inp, oup)
 
 def simple_segment_model(input_dim):
-    # input_dim = (100,100)
-    input_dim = (506,743)
     inp = Input(shape=input_dim)
     b_gr1 = Bidirectional(GRU(256, return_sequences=True),merge_mode='sum')(inp)
     # b_gr1
@@ -75,9 +73,9 @@ def load_model_arch(mod_file):
     model_f.close()
     return mod
 
-def train_segment(collection_name = 'test'):
+def train_segment(collection_name = 'test',resume_weights='',initial_epoch=0):
     # collection_name = 'story_test'
-    batch_size = 128
+    batch_size = 64
     # batch_size  = 4
     model_dir = './models/segment/'+collection_name
     create_dir(model_dir)
@@ -101,7 +99,7 @@ def train_segment(collection_name = 'test'):
         embeddings_freq=0,
         embeddings_layer_names=None,
         embeddings_metadata=None)
-    cp_file_fmt = model_dir+'/siamese_speech_model-{epoch:02d}-epoch-{val_loss:0.2f}\
+    cp_file_fmt = model_dir+'/speech_segment_model-{epoch:02d}-epoch-{val_loss:0.2f}\
 -acc.h5'
 
     cp_cb = ModelCheckpoint(
@@ -115,14 +113,16 @@ def train_segment(collection_name = 'test'):
     # train
     rms = RMSprop()
     model.compile(loss=categorical_crossentropy, optimizer=rms, metrics=[accuracy])
-    write_model_arch(model,model_dir+'/siamese_speech_model_arch.yaml')
+    write_model_arch(model,model_dir+'/speech_segment_model_arch.yaml')
     epoch_n_steps = step_count(n_records,batch_size)
+    if resume_weights != '':
+        model.load_weights(resume_weights)
     model.fit_generator(tr_gen
-                        , epochs=1000
+                        , epochs=10000
                         , steps_per_epoch=epoch_n_steps
                         , validation_data=(te_x, te_y)
                         , max_queue_size=32
-                        , callbacks=[tb_cb, cp_cb])
+                        , callbacks=[tb_cb, cp_cb],initial_epoch=initial_epoch)
     model.save(model_dir+'/speech_segment_model-final.h5')
 
     # y_pred = model.predict([te_pairs[:, 0], te_pairs[:, 1]])
@@ -133,4 +133,4 @@ def train_segment(collection_name = 'test'):
 
 if __name__ == '__main__':
     # pass
-    train_segment('test')
+    train_segment('story_phrases','./models/segment/story_phrases.1000/speech_segment_model-final.h5',1001)