implemented the model, todo implement ctc and training queueing logic

requirements-linux.txt

@@ -41,7 +41,6 @@ parso==0.1.0
 partd==0.3.8
 pexpect==4.2.1
 pickleshare==0.7.4
-pkg-resources==0.0.0
 praat-parselmouth==0.2.0
 progressbar2==3.34.3
 prompt-toolkit==1.0.15

segment_data.py

@@ -7,5 +7,6 @@ def fix_csv(collection_name = 'test'):
 
 
 def segment_data_gen(collection_name = 'test'):
-    # collection_name = 'test'
+    collection_name = 'test'
     seg_data = pd.read_csv('./outputs/'+collection_name+'.fixed.csv',index_col=0)
+    

segment_model.py

@@ -2,8 +2,9 @@ from __future__ import absolute_import
 from __future__ import print_function
 import numpy as np
 from keras.models import Model,load_model,model_from_yaml
-from keras.layers import Input,Concatenate,Lambda, BatchNormalization, Dropout
+from keras.layers import Input,Concatenate,Lambda, Reshape, Dropout
-from keras.layers import Dense, LSTM, Bidirectional, GRU
+from keras.layers import Dense,Conv2D, LSTM, Bidirectional, GRU
+from keras.layers import BatchNormalization
 from keras.losses import categorical_crossentropy
 from keras.utils import to_categorical
 from keras.optimizers import RMSprop
@@ -11,34 +12,28 @@ from keras.callbacks import TensorBoard, ModelCheckpoint
 from keras import backend as K
 from keras.utils import plot_model
 from speech_tools import create_dir,step_count
-from speech_data import segment_data_gen
+from segment_data import segment_data_gen
 
 
+# TODO implement ctc losses
+# https://github.com/fchollet/keras/blob/master/examples/image_ocr.py
 def accuracy(y_true, y_pred):
     '''Compute classification accuracy with a fixed threshold on distances.
     '''
     return K.mean(K.equal(y_true, K.cast(y_pred > 0.5, y_true.dtype)))
 
-def dense_classifier(processed):
-    conc_proc = Concatenate()(processed)
-    d1 = Dense(64, activation='relu')(conc_proc)
-    # dr1 = Dropout(0.1)(d1)
-    # d2 = Dense(128, activation='relu')(d1)
-    d3 = Dense(8, activation='relu')(d1)
-    # dr2 = Dropout(0.1)(d2)
-    return Dense(2, activation='softmax')(d3)
-
 def segment_model(input_dim):
     inp = Input(shape=input_dim)
     # ls0 = LSTM(512, return_sequences=True)(inp)
-    ls1 = LSTM(128, return_sequences=True)(inp)
+    cnv1 = Conv2D(filters=512, kernel_size=(5,9))(inp)
-    ls2 = LSTM(64, return_sequences=True)(ls1)
+    cnv2 = Conv2D(filters=1, kernel_size=(5,9))(cnv1)
-    # ls3 = LSTM(32, return_sequences=True)(ls2)
+    dr_cnv2 = Dropout(rate=0.95)(cnv2)
-    ls4 = LSTM(32)(ls2)
+    cn_rnn_dim = (dr_cnv2.shape[1].value,dr_cnv2.shape[2].value)
-    d1 = Dense(64, activation='relu')(ls4)
+    r_dr_cnv2 = Reshape(target_shape=cn_rnn_dim)(dr_cnv2)
-    d3 = Dense(8, activation='relu')(d1)
+    b_gr1 = Bidirectional(GRU(512, return_sequences=True))(r_dr_cnv2)
-    oup = Dense(2, activation='softmax')(d3)
+    b_gr2 = Bidirectional(GRU(512, return_sequences=True))(b_gr1)
-    return Model(inp, oup)
+    b_gr3 = Bidirectional(GRU(512))(b_gr2)
+    return Model(inp, b_gr3)
 
 def write_model_arch(mod,mod_file):
     model_f = open(mod_file,'w')
@@ -99,10 +94,11 @@ def train_segment(collection_name = 'test'):
     model.save(model_dir+'/speech_segment_model-final.h5')
 
     y_pred = model.predict([te_pairs[:, 0], te_pairs[:, 1]])
-    te_acc = compute_accuracy(te_y, y_pred)
+    # te_acc = compute_accuracy(te_y, y_pred)
-    print('* Accuracy on test set: %0.2f%%' % (100 * te_acc))
+    # print('* Accuracy on test set: %0.2f%%' % (100 * te_acc))
 
 
 
 if __name__ == '__main__':
+    import pdb; pdb.set_trace()
     train_segment('test')

speech_model.py

@@ -20,9 +20,11 @@ def create_base_rnn_network(input_dim):
     inp = Input(shape=input_dim)
     # ls0 = LSTM(512, return_sequences=True)(inp)
     ls1 = LSTM(128, return_sequences=True)(inp)
-    ls2 = LSTM(64, return_sequences=True)(ls1)
+    bn_ls1 = BatchNormalization(momentum=0.98)(ls1)
+    ls2 = LSTM(64, return_sequences=True)(bn_ls1)
+    bn_ls2 = BatchNormalization(momentum=0.98)(ls2)
     # ls3 = LSTM(32, return_sequences=True)(ls2)
-    ls4 = LSTM(32)(ls2)
+    ls4 = LSTM(32)(bn_ls2)
     # d1 = Dense(128, activation='relu')(ls4)
     #d2 = Dense(64, activation='relu')(ls2)
     return Model(inp, ls4)