shithub: opus

Info  •  Files  •  Log  •  Branches

WIP: Adding a constraint

--- a/dnn/training_tf2/lpcnet.py

+++ b/dnn/training_tf2/lpcnet.py

@@ -29,6 +29,7 @@

 from tensorflow.keras.models import Model

 from tensorflow.keras.layers import Input, GRU, Dense, Embedding, Reshape, Concatenate, Lambda, Conv1D, Multiply, Add, Bidirectional, MaxPooling1D, Activation

 from tensorflow.keras import backend as K

+from tensorflow.keras.constraints import Constraint

 from tensorflow.keras.initializers import Initializer

 from tensorflow.keras.callbacks import Callback

 from mdense import MDense

@@ -115,6 +116,21 @@

             'seed': self.seed

         }

+class WeightClip(Constraint):

+    '''Clips the weights incident to each hidden unit to be inside a range

+    '''

+    def __init__(self, c=2):

+        self.c = c

+

+    def __call__(self, p):

+        return K.clip(p, -self.c, self.c)

+

+    def get_config(self):

+        return {'name': self.__class__.__name__,

+            'c': self.c}

+

+constraint = WeightClip(0.999)

+

 def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 38, training=False, adaptation=False):

     pcm = Input(shape=(None, 3))

     feat = Input(shape=(None, nb_used_features))

@@ -142,8 +158,10 @@

     rep = Lambda(lambda x: K.repeat_elements(x, frame_size, 1))

-    rnn = GRU(rnn_units1, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_a')

-    rnn2 = GRU(rnn_units2, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_b')

+    rnn = GRU(rnn_units1, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_a',

+              recurrent_constraint = constraint)

+    rnn2 = GRU(rnn_units2, return_sequences=True, return_state=True, recurrent_activation="sigmoid", reset_after='true', name='gru_b',

+               kernel_constraint=constraint)

     rnn_in = Concatenate()([cpcm, rep(cfeat)])

     md = MDense(pcm_levels, activation='softmax', name='dual_fc')

--- a/dnn/training_tf2/train_lpcnet.py

+++ b/dnn/training_tf2/train_lpcnet.py

@@ -102,14 +102,15 @@

 del in_exc

 # dump models to disk as we go

-checkpoint = ModelCheckpoint('lpcnet32v_384_10_G16_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet32y_384_10_G16_{epoch:02d}.h5')

 #Set this to True to adapt an existing model (e.g. on new data)

 adaptation = False

+model.load_weights('lpcnet32v_384_10_G16_00.h5')

 if adaptation:

     #Adapting from an existing model

-    model.load_weights('lpcnet24c_384_10_G16_120.h5')

+    model.load_weights('lpcnet32v_384_10_G16_100.h5')

     sparsify = lpcnet.Sparsify(0, 0, 1, (0.05, 0.05, 0.2))

     lr = 0.0001

     decay = 0

@@ -120,5 +121,5 @@

     decay = 5e-5

 model.compile(optimizer=Adam(lr, decay=decay, beta_2=0.99), loss='sparse_categorical_crossentropy')

-model.save_weights('lpcnet32v_384_10_G16_00.h5');

+model.save_weights('lpcnet32y_384_10_G16_00.h5');

 model.fit([in_data, features, periods], out_exc, batch_size=batch_size, epochs=nb_epochs, validation_split=0.0, callbacks=[checkpoint, sparsify])

-- 

⑨