ref: 824dbecaecaf0796696ee79081ebd70f1cffb455
parent: 06511ba5a4b154b7600ed1bdd6b5345494467e44
author: Jean-Marc Valin <jmvalin@jmvalin.ca>
date: Mon Jul 9 14:20:52 EDT 2018
decoder wip
--- a/dnn/lpcnet.py
+++ b/dnn/lpcnet.py
@@ -19,6 +19,7 @@
pcm = Input(shape=(None, 1))
pitch = Input(shape=(None, 1))
feat = Input(shape=(None, nb_used_features))
+ dec_feat = Input(shape=(None, 32))
conv1 = Conv1D(16, 7, padding='causal')
pconv1 = Conv1D(16, 5, padding='same')
@@ -26,7 +27,7 @@
fconv1 = Conv1D(128, 3, padding='same')
fconv2 = Conv1D(32, 3, padding='same')
- if True:
+ if False:
cpcm = conv1(pcm)
cpitch = pconv2(pconv1(pitch))
else:
@@ -37,10 +38,18 @@
rep = Lambda(lambda x: K.repeat_elements(x, 160, 1))
- rnn = CuDNNGRU(rnn_units, return_sequences=True)
+ rnn = CuDNNGRU(rnn_units, return_sequences=True, return_state=True)
rnn_in = Concatenate()([cpcm, cpitch, rep(cfeat)])
md = MDense(pcm_levels, activation='softmax')
- ulaw_prob = md(rnn(rnn_in))
+ gru_out, state = rnn(rnn_in)
+ ulaw_prob = md(gru_out)
model = Model([pcm, pitch, feat], ulaw_prob)
- return model
+ encoder = Model(feat, cfeat)
+
+ dec_rnn_in = Concatenate()([cpcm, cpitch, dec_feat])
+ dec_gru_out, state = rnn(dec_rnn_in)
+ dec_ulaw_prob = md(dec_gru_out)
+
+ decoder = Model([pcm, pitch, dec_feat], [dec_ulaw_prob, state])
+ return model, encoder, decoder
--- /dev/null
+++ b/dnn/test_lpcnet.py
@@ -1,0 +1,64 @@
+#!/usr/bin/python3
+
+import lpcnet
+import sys
+import numpy as np
+from keras.optimizers import Adam
+from keras.callbacks import ModelCheckpoint
+from ulaw import ulaw2lin, lin2ulaw
+import keras.backend as K
+import h5py
+from adadiff import Adadiff
+
+#import tensorflow as tf
+#from keras.backend.tensorflow_backend import set_session
+#config = tf.ConfigProto()
+#config.gpu_options.per_process_gpu_memory_fraction = 0.28
+#set_session(tf.Session(config=config))
+
+nb_epochs = 40
+batch_size = 64
+
+model, enc, dec = lpcnet.new_wavernn_model()
+model.compile(optimizer=Adadiff(), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
+model.summary()
+
+pcmfile = sys.argv[1]
+feature_file = sys.argv[2]
+frame_size = 160
+nb_features = 54
+nb_used_features = lpcnet.nb_used_features
+feature_chunk_size = 15
+pcm_chunk_size = frame_size*feature_chunk_size
+
+data = np.fromfile(pcmfile, dtype='int8')
+nb_frames = len(data)//pcm_chunk_size
+
+features = np.fromfile(feature_file, dtype='float32')
+
+data = data[:nb_frames*pcm_chunk_size]
+features = features[:nb_frames*feature_chunk_size*nb_features]
+
+in_data = np.concatenate([data[0:1], data[:-1]])/16.;
+
+features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features))
+
+in_data = np.reshape(in_data, (nb_frames*pcm_chunk_size, 1))
+out_data = np.reshape(data, (nb_frames*pcm_chunk_size, 1))
+
+
+model.load_weights('lpcnet1h_30.h5')+
+order = 16
+
+pcm = 0.*out_data
+for c in range(1, nb_frames):
+ for fr in range(1, feature_chunk_size):
+ f = c*feature_chunk_size + fr
+ a = features[c, fr, nb_used_features+1:]
+ #print(a)
+ gain = 1;
+ for i in range(frame_size):
+ pcm[f*frame_size + i, 0] = gain*out_data[f*frame_size + i, 0] - sum(a*pcm[f*frame_size + i - 1:f*frame_size + i - order-1:-1, 0])
+ print(pcm[f*frame_size + i, 0])
+
--- a/dnn/train_lpcnet.py
+++ b/dnn/train_lpcnet.py
@@ -8,11 +8,12 @@
from ulaw import ulaw2lin, lin2ulaw
import keras.backend as K
import h5py
+from adadiff import Adadiff
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
config = tf.ConfigProto()
-config.gpu_options.per_process_gpu_memory_fraction = 0.44
+config.gpu_options.per_process_gpu_memory_fraction = 0.28
set_session(tf.Session(config=config))
nb_epochs = 40
@@ -19,7 +20,7 @@
batch_size = 64
model = lpcnet.new_wavernn_model()
-model.compile(optimizer=Adam(0.0008), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
+model.compile(optimizer=Adadiff(), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
model.summary()
pcmfile = sys.argv[1]
@@ -62,8 +63,8 @@
# f.create_dataset('data', data=in_data[:50000, :, :]) # f.create_dataset('feat', data=features[:50000, :, :])-checkpoint = ModelCheckpoint('lpcnet1g_{epoch:02d}.h5')+checkpoint = ModelCheckpoint('lpcnet1k_{epoch:02d}.h5') #model.load_weights('wavernn1c_01.h5')-model.compile(optimizer=Adam(0.002, amsgrad=True, decay=2e-4), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
+model.compile(optimizer=Adadiff(), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
model.fit([in_data, in_pitch, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])
--
⑨