eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-05-22 11:28:31 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

gru2 -> gruB and dotp included in dump_rdovae

Browse source
This commit is contained in:
jbuethe 2023-01-12 14:15:39 +00:00
parent 5b9b4381eb
commit a8673d0e25
4 changed files with 17 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

7
dnn/dred_rdovae_dec.c
View file

@ -55,13 +55,14 @@ void dred_rdovae_decode_qframe(
float buffer[DEC_DENSE1_OUT_SIZE + DEC_DENSE2_OUT_SIZE + DEC_DENSE3_OUT_SIZE + DEC_DENSE4_OUT_SIZE + DEC_DENSE5_OUT_SIZE + DEC_DENSE6_OUT_SIZE + DEC_DENSE7_OUT_SIZE + DEC_DENSE8_OUT_SIZE]; float buffer[DEC_DENSE1_OUT_SIZE + DEC_DENSE2_OUT_SIZE + DEC_DENSE3_OUT_SIZE + DEC_DENSE4_OUT_SIZE + DEC_DENSE5_OUT_SIZE + DEC_DENSE6_OUT_SIZE + DEC_DENSE7_OUT_SIZE + DEC_DENSE8_OUT_SIZE];
int output_index = 0; int output_index = 0;
int input_index = 0; int input_index = 0;
float zero_vector[1024] = {0};
/* run encoder stack and concatenate output in buffer*/ /* run encoder stack and concatenate output in buffer*/
_lpcnet_compute_dense(&dec_dense1, &buffer[output_index], input); _lpcnet_compute_dense(&dec_dense1, &buffer[output_index], input);
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE1_OUT_SIZE; output_index += DEC_DENSE1_OUT_SIZE;
compute_gru2(&dec_dense2, dec_state->dense2_state, &buffer[input_index]); compute_gruB(&dec_dense2, zero_vector, dec_state->dense2_state, &buffer[input_index]);
memcpy(&buffer[output_index], dec_state->dense2_state, DEC_DENSE2_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], dec_state->dense2_state, DEC_DENSE2_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE2_OUT_SIZE; output_index += DEC_DENSE2_OUT_SIZE;
@ -70,7 +71,7 @@ void dred_rdovae_decode_qframe(
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE3_OUT_SIZE; output_index += DEC_DENSE3_OUT_SIZE;
compute_gru2(&dec_dense4, dec_state->dense4_state, &buffer[input_index]); compute_gruB(&dec_dense4, zero_vector, dec_state->dense4_state, &buffer[input_index]);
memcpy(&buffer[output_index], dec_state->dense4_state, DEC_DENSE4_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], dec_state->dense4_state, DEC_DENSE4_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE4_OUT_SIZE; output_index += DEC_DENSE4_OUT_SIZE;
@ -79,7 +80,7 @@ void dred_rdovae_decode_qframe(
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE5_OUT_SIZE; output_index += DEC_DENSE5_OUT_SIZE;
compute_gru2(&dec_dense6, dec_state->dense6_state, &buffer[input_index]); compute_gruB(&dec_dense6, zero_vector, dec_state->dense6_state, &buffer[input_index]);
memcpy(&buffer[output_index], dec_state->dense6_state, DEC_DENSE6_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], dec_state->dense6_state, DEC_DENSE6_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += DEC_DENSE6_OUT_SIZE; output_index += DEC_DENSE6_OUT_SIZE;

7
dnn/dred_rdovae_enc.c
View file

@ -45,13 +45,14 @@ void dred_rdovae_encode_dframe(
float buffer[ENC_DENSE1_OUT_SIZE + ENC_DENSE2_OUT_SIZE + ENC_DENSE3_OUT_SIZE + ENC_DENSE4_OUT_SIZE + ENC_DENSE5_OUT_SIZE + ENC_DENSE6_OUT_SIZE + ENC_DENSE7_OUT_SIZE + ENC_DENSE8_OUT_SIZE + GDENSE1_OUT_SIZE]; float buffer[ENC_DENSE1_OUT_SIZE + ENC_DENSE2_OUT_SIZE + ENC_DENSE3_OUT_SIZE + ENC_DENSE4_OUT_SIZE + ENC_DENSE5_OUT_SIZE + ENC_DENSE6_OUT_SIZE + ENC_DENSE7_OUT_SIZE + ENC_DENSE8_OUT_SIZE + GDENSE1_OUT_SIZE];
int output_index = 0; int output_index = 0;
int input_index = 0; int input_index = 0;
float zero_vector[1024] = {0};
/* run encoder stack and concatenate output in buffer*/ /* run encoder stack and concatenate output in buffer*/
_lpcnet_compute_dense(&enc_dense1, &buffer[output_index], input); _lpcnet_compute_dense(&enc_dense1, &buffer[output_index], input);
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE1_OUT_SIZE; output_index += ENC_DENSE1_OUT_SIZE;
compute_gru2(&enc_dense2, enc_state->dense2_state, &buffer[input_index]); compute_gruB(&enc_dense2, zero_vector, enc_state->dense2_state, &buffer[input_index]);
memcpy(&buffer[output_index], enc_state->dense2_state, ENC_DENSE2_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], enc_state->dense2_state, ENC_DENSE2_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE2_OUT_SIZE; output_index += ENC_DENSE2_OUT_SIZE;
@ -60,7 +61,7 @@ void dred_rdovae_encode_dframe(
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE3_OUT_SIZE; output_index += ENC_DENSE3_OUT_SIZE;
compute_gru2(&enc_dense4, enc_state->dense4_state, &buffer[input_index]); compute_gruB(&enc_dense4, zero_vector, enc_state->dense4_state, &buffer[input_index]);
memcpy(&buffer[output_index], enc_state->dense4_state, ENC_DENSE4_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], enc_state->dense4_state, ENC_DENSE4_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE4_OUT_SIZE; output_index += ENC_DENSE4_OUT_SIZE;
@ -69,7 +70,7 @@ void dred_rdovae_encode_dframe(
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE5_OUT_SIZE; output_index += ENC_DENSE5_OUT_SIZE;
compute_gru2(&enc_dense6, enc_state->dense6_state, &buffer[input_index]); compute_gruB(&enc_dense6, zero_vector, enc_state->dense6_state, &buffer[input_index]);
memcpy(&buffer[output_index], enc_state->dense6_state, ENC_DENSE6_OUT_SIZE * sizeof(float)); memcpy(&buffer[output_index], enc_state->dense6_state, ENC_DENSE6_OUT_SIZE * sizeof(float));
input_index = output_index; input_index = output_index;
output_index += ENC_DENSE6_OUT_SIZE; output_index += ENC_DENSE6_OUT_SIZE;

10
dnn/training_tf2/dump_rdovae.py
View file

@ -50,7 +50,7 @@ from keraslayerdump import dump_conv1d_layer, dump_dense_layer, dump_gru_layer,
from rdovae import new_rdovae_model from rdovae import new_rdovae_model
def start_header(header_fid, header_name): def start_header(header_fid, header_name):
header_guard = "_" + os.path.basename(header_name)[:-2].upper() + "_H" header_guard = os.path.basename(header_name)[:-2].upper() + "_H"
header_fid.write( header_fid.write(
f""" f"""
#ifndef {header_guard} #ifndef {header_guard}
@ -72,6 +72,10 @@ def start_source(source_fid, header_name, weight_file):
f""" f"""
/* this source file was automatically generated from weight file {weight_file} */ /* this source file was automatically generated from weight file {weight_file} */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "{header_name}" #include "{header_name}"
""" """
@ -159,7 +163,7 @@ f"""
# dump GRUs # dump GRUs
max_rnn_neurons_enc = max( max_rnn_neurons_enc = max(
[ [
dump_gru_layer(encoder.get_layer(name), source_fid, header_fid) dump_gru_layer(encoder.get_layer(name), source_fid, header_fid, dotp=True, sparse=True)
for name in encoder_gru_names for name in encoder_gru_names
] ]
) )
@ -254,7 +258,7 @@ f"""
# dump GRUs # dump GRUs
max_rnn_neurons_dec = max( max_rnn_neurons_dec = max(
[ [
dump_gru_layer(decoder.get_layer(name), source_fid, header_fid) dump_gru_layer(decoder.get_layer(name), source_fid, header_fid, dotp=True, sparse=True)
for name in decoder_gru_names for name in decoder_gru_names
] ]
) )

4
dnn/training_tf2/keraslayerdump.py
View file

@ -149,8 +149,8 @@ def dump_gru_layer(self, f, hf, dotp=False, sparse=False):
reset_after = 1 reset_after = 1
neurons = weights[0].shape[1]//3 neurons = weights[0].shape[1]//3
max_rnn_neurons = neurons max_rnn_neurons = neurons
f.write('const GRULayer {} = {{\n {}_bias,\n {}_subias,\n {}_weights,\n NULL,\n {}_recurrent_weights,\n {}, {}, ACTIVATION_{}, {}\n}};\n\n' f.write('const GRULayer {} = {{\n {}_bias,\n {}_subias,\n {}_weights,\n {},\n {}_recurrent_weights,\n {}, {}, ACTIVATION_{}, {}\n}};\n\n'
.format(name, name, name, name, name, weights[0].shape[0], weights[0].shape[1]//3, activation, reset_after)) .format(name, name, name, name, name + "_weights_idx" if sparse else "NULL", name, weights[0].shape[0], weights[0].shape[1]//3, activation, reset_after))
hf.write('#define {}_OUT_SIZE {}\n'.format(name.upper(), weights[0].shape[1]//3)) hf.write('#define {}_OUT_SIZE {}\n'.format(name.upper(), weights[0].shape[1]//3))
hf.write('#define {}_STATE_SIZE {}\n'.format(name.upper(), weights[0].shape[1]//3)) hf.write('#define {}_STATE_SIZE {}\n'.format(name.upper(), weights[0].shape[1]//3))
hf.write('extern const GRULayer {};\n\n'.format(name)); hf.write('extern const GRULayer {};\n\n'.format(name));