shithub: opus

Info  •  Files  •  Log  •  Branches

Fix non-128 batch sizes

Avoid hardcoding the batch size in the model

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

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

@@ -230,10 +230,10 @@

 constraint = WeightClip(0.992)

-def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 20, training=False, adaptation=False, quantize=False, flag_e2e = False):

-    pcm = Input(shape=(None, 3), batch_size=128)

-    feat = Input(shape=(None, nb_used_features), batch_size=128)

-    pitch = Input(shape=(None, 1), batch_size=128)

+def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features=20, batch_size=128, training=False, adaptation=False, quantize=False, flag_e2e = False):

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

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

+    pitch = Input(shape=(None, 1), batch_size=batch_size)

     dec_feat = Input(shape=(None, 128))

     dec_state1 = Input(shape=(rnn_units1,))

     dec_state2 = Input(shape=(rnn_units2,))

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

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

@@ -121,7 +121,7 @@

 strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()

 with strategy.scope():

-    model, _, _ = lpcnet.new_lpcnet_model(rnn_units1=args.grua_size, rnn_units2=args.grub_size, training=True, quantize=quantize, flag_e2e = flag_e2e)

+    model, _, _ = lpcnet.new_lpcnet_model(rnn_units1=args.grua_size, rnn_units2=args.grub_size, batch_size=batch_size, training=True, quantize=quantize, flag_e2e = flag_e2e)

     if not flag_e2e:

         model.compile(optimizer=opt, loss='sparse_categorical_crossentropy', metrics='sparse_categorical_crossentropy')

     else:

-- 

⑨