shithub: opus

Info  •  Files  •  Log  •  Branches

reduce memory use of training code

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -86,7 +86,9 @@

 in_data = np.concatenate([data[0:1], data[:-1]]);

 noise = np.concatenate([np.zeros((len(data)*1//5)), np.random.randint(-3, 3, len(data)*1//5), np.random.randint(-2, 2, len(data)*1//5), np.random.randint(-1, 1, len(data)*2//5)])

 #noise = np.round(np.concatenate([np.zeros((len(data)*1//5)), np.random.laplace(0, 1.2, len(data)*1//5), np.random.laplace(0, .77, len(data)*1//5), np.random.laplace(0, .33, len(data)*1//5), np.random.randint(-1, 1, len(data)*1//5)]))

+del data

 in_data = in_data + noise

+del noise

 in_data = np.clip(in_data, 0, 255)

 features = np.reshape(features, (nb_frames*feature_chunk_size, nb_features))

@@ -105,6 +107,7 @@

     for k in range(16):

         upred[i*frame_size:(i+1)*frame_size] = upred[i*frame_size:(i+1)*frame_size] - \

             pred_in[i*frame_size-k:(i+1)*frame_size-k]*features[i, nb_features-16+k]

+del pred_in

 pred = lin2ulaw(upred)

@@ -116,6 +119,8 @@

 # ideal excitation in_exc

 out_data = lin2ulaw(udata-upred)

+del upred

+del udata

 in_exc = np.concatenate([out_data[0:1], out_data[:-1]]);

 out_data = np.reshape(out_data, (nb_frames, pcm_chunk_size, 1))

@@ -134,6 +139,8 @@

 periods = (50*features[:,:,36:37]+100).astype('int16')

 in_data = np.concatenate([in_data, pred], axis=-1)

+

+del pred

 # dump models to disk as we go

 checkpoint = ModelCheckpoint('lpcnet14_384_10_G16_{epoch:02d}.h5')

-- 

⑨