shithub: opus

Info  •  Files  •  Log  •  Branches

gru2 -> gruB and dotp included in dump_rdovae

--- a/dnn/dred_rdovae_dec.c

+++ b/dnn/dred_rdovae_dec.c

@@ -55,6 +55,7 @@

     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 input_index = 0;

+    float zero_vector[1024] = {0};

     /* run encoder stack and concatenate output in buffer*/

     _lpcnet_compute_dense(&dec_dense1, &buffer[output_index], input);

@@ -61,7 +62,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += DEC_DENSE2_OUT_SIZE;

@@ -70,7 +71,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += DEC_DENSE4_OUT_SIZE;

@@ -79,7 +80,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += DEC_DENSE6_OUT_SIZE;

--- a/dnn/dred_rdovae_enc.c

+++ b/dnn/dred_rdovae_enc.c

@@ -45,6 +45,7 @@

     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 input_index = 0;

+    float zero_vector[1024] = {0};

     /* run encoder stack and concatenate output in buffer*/

     _lpcnet_compute_dense(&enc_dense1, &buffer[output_index], input);

@@ -51,7 +52,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += ENC_DENSE2_OUT_SIZE;

@@ -60,7 +61,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += ENC_DENSE4_OUT_SIZE;

@@ -69,7 +70,7 @@

     input_index = output_index;

     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));

     input_index = output_index;

     output_index += ENC_DENSE6_OUT_SIZE;

--- a/dnn/training_tf2/dump_rdovae.py

+++ b/dnn/training_tf2/dump_rdovae.py

@@ -50,7 +50,7 @@

 from rdovae import new_rdovae_model

 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(

 f"""

 #ifndef {header_guard}

@@ -72,6 +72,10 @@

 f"""

 /* this source file was automatically generated from weight file {weight_file} */

+#ifdef HAVE_CONFIG_H

+#include "config.h"

+#endif

+

 #include "{header_name}"

 """

@@ -159,7 +163,7 @@

     # dump GRUs

     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

         ]

     )

@@ -254,7 +258,7 @@

     # dump GRUs

     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

         ]

     )

--- a/dnn/training_tf2/keraslayerdump.py

+++ b/dnn/training_tf2/keraslayerdump.py

@@ -149,8 +149,8 @@

         reset_after = 1

     neurons = weights[0].shape[1]//3

     max_rnn_neurons = neurons

-    f.write('const GRULayer {} = {{\n   {}_bias,\n   {}_subias,\n   {}_weights,\n   NULL,\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))

+    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_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 {}_STATE_SIZE {}\n'.format(name.upper(), weights[0].shape[1]//3))

     hf.write('extern const GRULayer {};\n\n'.format(name));

-- 

⑨