eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-06-06 07:21:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

auto-detect end-to-end models

Browse source
This commit is contained in:
Jean-Marc Valin 2021-10-13 22:12:39 -04:00
parent d5b6087f48
commit a3ef596822
2 changed files with 26 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

28
dnn/dump_data.c
View file

@ -75,7 +75,7 @@ void compute_noise(int *noise, float noise_std) {
} }
void write_audio(LPCNetEncState *st, const short *pcm, const int *noise, FILE *file, int nframes) { void write_audio(LPCNetEncState *st, const short *pcm, const int *noise, FILE *file, int nframes, int e2e) {
int i, k; int i, k;
for (k=0;k<nframes;k++) { for (k=0;k<nframes;k++) {
unsigned char data[4*FRAME_SIZE]; unsigned char data[4*FRAME_SIZE];
@ -92,11 +92,11 @@ void write_audio(LPCNetEncState *st, const short *pcm, const int *noise, FILE *f
/* Excitation in. */ /* Excitation in. */
data[4*i+2] = st->exc_mem; data[4*i+2] = st->exc_mem;
/* Excitation out. */ /* Excitation out. */
#ifdef END2END if (e2e) {
data[4*i+3] = lin2ulaw(pcm[k*FRAME_SIZE+i]); data[4*i+3] = lin2ulaw(pcm[k*FRAME_SIZE+i]);
#else } else {
data[4*i+3] = e; data[4*i+3] = e;
#endif }
/* Simulate error on excitation. */ /* Simulate error on excitation. */
e += noise[k*FRAME_SIZE+i]; e += noise[k*FRAME_SIZE+i];
e = IMIN(255, IMAX(0, e)); e = IMIN(255, IMAX(0, e));
@ -118,6 +118,8 @@ static short float2short(float x)
int main(int argc, char **argv) { int main(int argc, char **argv) {
int i; int i;
char *argv0;
int e2e=0;
int count=0; int count=0;
static const float a_hp[2] = {-1.99599, 0.99600}; static const float a_hp[2] = {-1.99599, 0.99600};
static const float b_hp[2] = {-2, 1}; static const float b_hp[2] = {-2, 1};
@ -148,6 +150,12 @@ int main(int argc, char **argv) {
int quantize = 0; int quantize = 0;
srand(getpid()); srand(getpid());
st = lpcnet_encoder_create(); st = lpcnet_encoder_create();
argv0=argv[0];
if (argc > 2 && strcmp(argv[1], "-end2end")==0) {
e2e = 1;
argv++;
argc--;
}
if (argc == 5 && strcmp(argv[1], "-train")==0) training = 1; if (argc == 5 && strcmp(argv[1], "-train")==0) training = 1;
if (argc == 5 && strcmp(argv[1], "-qtrain")==0) { if (argc == 5 && strcmp(argv[1], "-qtrain")==0) {
training = 1; training = 1;
@ -168,8 +176,8 @@ int main(int argc, char **argv) {
decode = 1; decode = 1;
} }
if (training == -1) { if (training == -1) {
fprintf(stderr, "usage: %s -train <speech> <features out> <pcm out>\n", argv[0]); fprintf(stderr, "usage: %s -train <speech> <features out> <pcm out>\n", argv0);
fprintf(stderr, " or %s -test <speech> <features out>\n", argv[0]); fprintf(stderr, " or %s -test <speech> <features out>\n", argv0);
return 1; return 1;
} }
f1 = fopen(argv[2], "r"); f1 = fopen(argv[2], "r");
@ -273,7 +281,7 @@ int main(int argc, char **argv) {
if (!quantize) { if (!quantize) {
process_single_frame(st, ffeat); process_single_frame(st, ffeat);
if (fpcm) write_audio(st, pcm, &noisebuf[st->pcount*FRAME_SIZE], fpcm, 1); if (fpcm) write_audio(st, pcm, &noisebuf[st->pcount*FRAME_SIZE], fpcm, 1, e2e);
} }
st->pcount++; st->pcount++;
/* Running on groups of 4 frames. */ /* Running on groups of 4 frames. */
@ -281,7 +289,7 @@ int main(int argc, char **argv) {
if (quantize) { if (quantize) {
unsigned char buf[8]; unsigned char buf[8];
process_superframe(st, buf, ffeat, encode, quantize); process_superframe(st, buf, ffeat, encode, quantize);
if (fpcm) write_audio(st, pcmbuf, noisebuf, fpcm, 4); if (fpcm) write_audio(st, pcmbuf, noisebuf, fpcm, 4, e2e);
} }
st->pcount = 0; st->pcount = 0;
} }

8
dnn/training_tf2/dump_lpcnet.py
View file

@ -250,6 +250,7 @@ with h5py.File(filename, "r") as f:
units = min(f['model_weights']['gru_a']['gru_a']['recurrent_kernel:0'].shape) units = min(f['model_weights']['gru_a']['gru_a']['recurrent_kernel:0'].shape)
units2 = min(f['model_weights']['gru_b']['gru_b']['recurrent_kernel:0'].shape) units2 = min(f['model_weights']['gru_b']['gru_b']['recurrent_kernel:0'].shape)
cond_size = min(f['model_weights']['feature_dense1']['feature_dense1']['kernel:0'].shape) cond_size = min(f['model_weights']['feature_dense1']['feature_dense1']['kernel:0'].shape)
e2e = 'rc2lpc' in f['model_weights']
model, _, _ = lpcnet.new_lpcnet_model(rnn_units1=units, rnn_units2=units2, flag_e2e = flag_e2e, cond_size=cond_size) model, _, _ = lpcnet.new_lpcnet_model(rnn_units1=units, rnn_units2=units2, flag_e2e = flag_e2e, cond_size=cond_size)
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy']) model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])
@ -276,6 +277,13 @@ f.write('#ifdef HAVE_CONFIG_H\n#include "config.h"\n#endif\n\n#include "nnet.h"\
hf.write('/*This file is automatically generated from a Keras model*/\n\n') hf.write('/*This file is automatically generated from a Keras model*/\n\n')
hf.write('#ifndef RNN_DATA_H\n#define RNN_DATA_H\n\n#include "nnet.h"\n\n') hf.write('#ifndef RNN_DATA_H\n#define RNN_DATA_H\n\n#include "nnet.h"\n\n')
if e2e:
hf.write('/* This is an end-to-end model */\n')
hf.write('#define END2END\n\n')
else:
hf.write('/* This is *not* an end-to-end model */\n')
hf.write('/* #define END2END */\n\n')
embed_size = lpcnet.embed_size embed_size = lpcnet.embed_size
E = model.get_layer('embed_sig').get_weights()[0] E = model.get_layer('embed_sig').get_weights()[0]