eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-05-18 09:28:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Frame network seems to be working

Browse source
This commit is contained in:
Jean-Marc Valin 2018-11-26 18:41:54 -05:00
parent 538f25565a
commit 4ccfbdff04
3 changed files with 23 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

1
dnn/dump_lpcnet.py
View file

@ -135,6 +135,7 @@ def dump_conv1d_layer(self, f, hf):
.format(name, name, name, weights[0].shape[1], weights[0].shape[0], weights[0].shape[2], activation))
hf.write('#define {}_OUT_SIZE {}\n'.format(name.upper(), weights[0].shape[2]))
hf.write('#define {}_STATE_SIZE ({}*{})\n'.format(name.upper(), weights[0].shape[1], (weights[0].shape[0]-1)))
hf.write('#define {}_DELAY {}\n'.format(name.upper(), (weights[0].shape[0]-1)//2))
hf.write('extern const Conv1DLayer {};\n\n'.format(name));
return True
Conv1D.dump_layer = dump_conv1d_layer

24
dnn/lpcnet.c
View file

@ -33,6 +33,8 @@
#include "lpcnet.h"
#define NB_FEATURES 38
#define NB_TOTAL_FEATURES 55
#define LPC_ORDER 16
@ -43,10 +45,12 @@
#define SAMPLE_INPUT_SIZE (2*EMBED_SIG_OUT_SIZE + EMBED_EXC_OUT_SIZE + FEATURE_DENSE2_OUT_SIZE)
#define FEATURES_DELAY (FEATURE_CONV1_DELAY + FEATURE_CONV2_DELAY)
struct LPCNetState {
NNetState nnet;
int last_exc;
short last_sig[LPC_ORDER];
float old_input[FEATURES_DELAY][FEATURE_CONV2_OUT_SIZE];
};
@ -73,19 +77,23 @@ static int lin2ulaw(int x)
return (int)floor(.5 + u);
}
void run_frame_network(NNetState *net, float *condition, float *lpc, const float *features, int pitch)
void run_frame_network(LPCNetState *lpcnet, float *condition, float *lpc, const float *features, int pitch)
{
int i;
NNetState *net;
float in[FRAME_INPUT_SIZE];
float conv1_out[FEATURE_CONV1_OUT_SIZE];
float conv2_out[FEATURE_CONV2_OUT_SIZE];
float dense1_out[FEATURE_DENSE1_OUT_SIZE];
net = &lpcnet->nnet;
RNN_COPY(in, features, NB_FEATURES);
compute_embedding(&embed_pitch, &in[NB_FEATURES], pitch);
compute_conv1d(&feature_conv1, conv1_out, net->feature_conv1_state, in);
compute_conv1d(&feature_conv2, conv2_out, net->feature_conv2_state, conv1_out);
celt_assert(FRAME_INPUT_SIZE == FEATURE_CONV2_OUT_SIZE);
for (i=0;i<FEATURE_CONV2_OUT_SIZE;i++) conv2_out[i] += in[i];
for (i=0;i<FEATURE_CONV2_OUT_SIZE;i++) conv2_out[i] += lpcnet->old_input[FEATURES_DELAY-1][i];
memmove(lpcnet->old_input[1], lpcnet->old_input[0], (FEATURES_DELAY-1)*FRAME_INPUT_SIZE*sizeof(in[0]));
memcpy(lpcnet->old_input[0], in, FRAME_INPUT_SIZE*sizeof(in[0]));
compute_dense(&feature_dense1, dense1_out, conv2_out);
compute_dense(&feature_dense2, condition, dense1_out);
/* FIXME: Actually compute the LPC on the middle frame. */
@ -127,10 +135,11 @@ void lpcnet_synthesize(LPCNetState *lpcnet, short *output, const float *features
float pdf[DUAL_FC_OUT_SIZE];
int pitch;
float pitch_gain;
pitch = (int)floor(.5 + 50*features[36]+100);
/* FIXME: Do proper rounding once the Python code rounds properly. */
pitch = (int)floor(50*features[36]+100);
/* FIXME: get the pitch gain from 2 frames in the past. */
pitch_gain = features[PITCH_GAIN_FEATURE];
run_frame_network(&lpcnet->nnet, condition, lpc, features, pitch);
run_frame_network(lpcnet, condition, lpc, features, pitch);
for (i=0;i<N;i++)
{
int j;
@ -154,13 +163,16 @@ void lpcnet_synthesize(LPCNetState *lpcnet, short *output, const float *features
#if 1
#define FRAME_SIZE 160
int main(int argc, char **argv) {
int main() {
LPCNetState *net;
net = lpcnet_create();
while (1) {
float in_features[NB_TOTAL_FEATURES];
float features[NB_FEATURES];
short pcm[FRAME_SIZE];
fread(features, sizeof(features[0]), NB_FEATURES, stdin);
fread(in_features, sizeof(features[0]), NB_TOTAL_FEATURES, stdin);
RNN_COPY(features, in_features, NB_FEATURES);
RNN_CLEAR(&features[18], 18);
if (feof(stdin)) break;
lpcnet_synthesize(net, pcm, features, FRAME_SIZE);
fwrite(pcm, sizeof(pcm[0]), FRAME_SIZE, stdout);

5
dnn/nnet.c
View file

@ -109,7 +109,7 @@ void compute_activation(float *output, float *input, int N, int activation)
for (i=0;i<N;i++)
output[i] = sum*output[i];
} else {
celt_assert(layer->activation == ACTIVATION_LINEAR);
celt_assert(activation == ACTIVATION_LINEAR);
for (i=0;i<N;i++)
output[i] = input[i];
}
@ -231,6 +231,9 @@ void compute_conv1d(const Conv1DLayer *layer, float *output, float *mem, const f
void compute_embedding(const EmbeddingLayer *layer, float *output, int input)
{
int i;
celt_assert(input >= 0);
celt_assert(input < layer->nb_inputs);
/*if (layer->dim == 64) printf("%d\n", input);*/
for (i=0;i<layer->dim;i++)
{
output[i] = layer->embedding_weights[input*layer->dim + i];