C implementation of FARGAN

autogen.sh

@@ -9,7 +9,7 @@ set -e
 srcdir=`dirname $0`
 test -n "$srcdir" && cd "$srcdir"
 
-dnn/download_model.sh f68e31d
+dnn/download_model.sh 9e76a7b
 
 echo "Updating build configuration files, please wait...."
 

dnn/fargan.c

@@ -0,0 +1,220 @@
+/* Copyright (c) 2023 Amazon */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "fargan.h"
+#include "os_support.h"
+#include "freq.h"
+#include "fargan_data.h"
+#include "lpcnet.h"
+#include "pitch.h"
+#include "nnet.h"
+#include "lpcnet_private.h"
+
+#define FARGAN_FEATURES (NB_FEATURES)
+
+static void compute_fargan_cond(FARGANState *st, float *cond, const float *features, int period)
+{
+  FARGAN *model;
+  float dense_in[NB_FEATURES+COND_NET_PEMBED_OUT_SIZE];
+  float conv1_in[COND_NET_FCONV1_IN_SIZE];
+  float conv2_in[COND_NET_FCONV2_IN_SIZE];
+  model = &st->model;
+  celt_assert(FARGAN_FEATURES+COND_NET_PEMBED_OUT_SIZE == model->cond_net_fdense1.nb_inputs);
+  celt_assert(COND_NET_FCONV1_IN_SIZE == model->cond_net_fdense1.nb_outputs);
+  celt_assert(COND_NET_FCONV2_IN_SIZE == model->cond_net_fconv1.nb_outputs);
+  OPUS_COPY(&dense_in[NB_FEATURES], &model->cond_net_pembed.float_weights[IMAX(0,IMIN(period-32, 224))*COND_NET_PEMBED_OUT_SIZE], COND_NET_PEMBED_OUT_SIZE);
+  OPUS_COPY(dense_in, features, NB_FEATURES);
+
+  compute_generic_dense(&model->cond_net_fdense1, conv1_in, dense_in, ACTIVATION_TANH);
+  compute_generic_conv1d(&model->cond_net_fconv1, conv2_in, st->cond_conv1_state, conv1_in, COND_NET_FCONV1_IN_SIZE, ACTIVATION_TANH);
+  compute_generic_conv1d(&model->cond_net_fconv2, cond, st->cond_conv2_state, conv2_in, COND_NET_FCONV2_IN_SIZE, ACTIVATION_TANH);
+}
+
+static void fargan_deemphasis(float *pcm, float *deemph_mem) {
+  int i;
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) {
+    pcm[i] += FARGAN_DEEMPHASIS * *deemph_mem;
+    *deemph_mem = pcm[i];
+  }
+}
+
+static void run_fargan_subframe(FARGANState *st, float *pcm, const float *cond, int period)
+{
+  int i, pos;
+  float fwc0_in[SIG_NET_INPUT_SIZE];
+  float gru1_in[SIG_NET_FWC0_CONV_OUT_SIZE+2*FARGAN_SUBFRAME_SIZE];
+  float gru2_in[SIG_NET_GRU1_OUT_SIZE+2*FARGAN_SUBFRAME_SIZE];
+  float gru3_in[SIG_NET_GRU2_OUT_SIZE+2*FARGAN_SUBFRAME_SIZE];
+  float pred[FARGAN_SUBFRAME_SIZE+4];
+  float prev[FARGAN_SUBFRAME_SIZE];
+  float pitch_gate[4];
+  float gain;
+  float gain_1;
+  float skip_cat[10000];
+  float skip_out[SIG_NET_SKIP_DENSE_OUT_SIZE];
+  FARGAN *model;
+
+  celt_assert(st->cont_initialized);
+  model = &st->model;
+
+  compute_generic_dense(&model->sig_net_cond_gain_dense, &gain, cond, ACTIVATION_LINEAR);
+  gain = exp(gain);
+  gain_1 = 1.f/(1e-5 + gain);
+
+  pos = PITCH_MAX_PERIOD-period-2;
+  for (i=0;i<FARGAN_SUBFRAME_SIZE+4;i++) {
+    pred[i] = MIN32(1.f, MAX32(-1.f, gain_1*st->pitch_buf[IMAX(0, pos)]));
+    pos++;
+    if (pos == PITCH_MAX_PERIOD) pos -= period;
+  }
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) prev[i] = MAX32(-1.f, MIN16(1.f, gain_1*st->pitch_buf[PITCH_MAX_PERIOD-FARGAN_SUBFRAME_SIZE+i]));
+
+  OPUS_COPY(&fwc0_in[0], &cond[0], FARGAN_COND_SIZE);
+  OPUS_COPY(&fwc0_in[FARGAN_COND_SIZE], pred, FARGAN_SUBFRAME_SIZE+4);
+  OPUS_COPY(&fwc0_in[FARGAN_COND_SIZE+FARGAN_SUBFRAME_SIZE+4], prev, FARGAN_SUBFRAME_SIZE);
+
+  compute_generic_conv1d(&model->sig_net_fwc0_conv, gru1_in, st->fwc0_mem, fwc0_in, SIG_NET_INPUT_SIZE, ACTIVATION_TANH);
+  celt_assert(SIG_NET_FWC0_GLU_GATE_OUT_SIZE == model->sig_net_fwc0_glu_gate.nb_outputs);
+  compute_glu(&model->sig_net_fwc0_glu_gate, gru1_in, gru1_in);
+
+  compute_generic_dense(&model->sig_net_gain_dense_out, pitch_gate, gru1_in, ACTIVATION_SIGMOID);
+
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) gru1_in[SIG_NET_FWC0_GLU_GATE_OUT_SIZE+i] = pitch_gate[0]*pred[i+2];
+  OPUS_COPY(&gru1_in[SIG_NET_FWC0_GLU_GATE_OUT_SIZE+FARGAN_SUBFRAME_SIZE], prev, FARGAN_SUBFRAME_SIZE);
+  compute_generic_gru(&model->sig_net_gru1_input, &model->sig_net_gru1_recurrent, st->gru1_state, gru1_in);
+  compute_glu(&model->sig_net_gru1_glu_gate, gru2_in, st->gru1_state);
+
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) gru2_in[SIG_NET_GRU1_OUT_SIZE+i] = pitch_gate[1]*pred[i+2];
+  OPUS_COPY(&gru2_in[SIG_NET_GRU1_OUT_SIZE+FARGAN_SUBFRAME_SIZE], prev, FARGAN_SUBFRAME_SIZE);
+  compute_generic_gru(&model->sig_net_gru2_input, &model->sig_net_gru2_recurrent, st->gru2_state, gru2_in);
+  compute_glu(&model->sig_net_gru2_glu_gate, gru3_in, st->gru2_state);
+
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) gru3_in[SIG_NET_GRU2_OUT_SIZE+i] = pitch_gate[2]*pred[i+2];
+  OPUS_COPY(&gru3_in[SIG_NET_GRU2_OUT_SIZE+FARGAN_SUBFRAME_SIZE], prev, FARGAN_SUBFRAME_SIZE);
+  compute_generic_gru(&model->sig_net_gru3_input, &model->sig_net_gru3_recurrent, st->gru3_state, gru3_in);
+  compute_glu(&model->sig_net_gru3_glu_gate, &skip_cat[SIG_NET_GRU1_OUT_SIZE+SIG_NET_GRU2_OUT_SIZE], st->gru3_state);
+
+  OPUS_COPY(skip_cat, gru2_in, SIG_NET_GRU1_OUT_SIZE);
+  OPUS_COPY(&skip_cat[SIG_NET_GRU1_OUT_SIZE], gru3_in, SIG_NET_GRU2_OUT_SIZE);
+  OPUS_COPY(&skip_cat[SIG_NET_GRU1_OUT_SIZE+SIG_NET_GRU2_OUT_SIZE+SIG_NET_GRU3_OUT_SIZE], gru1_in, SIG_NET_FWC0_CONV_OUT_SIZE);
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) skip_cat[SIG_NET_GRU1_OUT_SIZE+SIG_NET_GRU2_OUT_SIZE+SIG_NET_GRU3_OUT_SIZE+SIG_NET_FWC0_CONV_OUT_SIZE+i] = pitch_gate[3]*pred[i+2];
+  OPUS_COPY(&skip_cat[SIG_NET_GRU1_OUT_SIZE+SIG_NET_GRU2_OUT_SIZE+SIG_NET_GRU3_OUT_SIZE+SIG_NET_FWC0_CONV_OUT_SIZE+FARGAN_SUBFRAME_SIZE], prev, FARGAN_SUBFRAME_SIZE);
+
+  compute_generic_dense(&model->sig_net_skip_dense, skip_out, skip_cat, ACTIVATION_TANH);
+  compute_glu(&model->sig_net_skip_glu_gate, skip_out, skip_out);
+
+  compute_generic_dense(&model->sig_net_sig_dense_out, pcm, skip_out, ACTIVATION_TANH);
+  for (i=0;i<FARGAN_SUBFRAME_SIZE;i++) pcm[i] *= gain;
+
+  OPUS_MOVE(st->pitch_buf, &st->pitch_buf[FARGAN_SUBFRAME_SIZE], PITCH_MAX_PERIOD-FARGAN_SUBFRAME_SIZE);
+  OPUS_COPY(&st->pitch_buf[PITCH_MAX_PERIOD-FARGAN_SUBFRAME_SIZE], pcm, FARGAN_SUBFRAME_SIZE);
+  fargan_deemphasis(pcm, &st->deemph_mem);
+}
+
+void fargan_cont(FARGANState *st, const float *pcm0, const float *features0)
+{
+  int i;
+  float cond[COND_NET_FCONV2_OUT_SIZE];
+  float x0[FARGAN_CONT_SAMPLES];
+  float dummy[FARGAN_SUBFRAME_SIZE];
+  int period=0;
+
+  /* Pre-load features. */
+  for (i=0;i<5;i++) {
+    const float *features = &features0[i*NB_FEATURES];
+    st->last_period = period;
+    period = (int)floor(.5+256./pow(2.f,((1./60.)*((features[NB_BANDS]+1.5)*60))));
+    compute_fargan_cond(st, cond, features, period);
+  }
+
+  x0[0] = 0;
+  for (i=1;i<FARGAN_CONT_SAMPLES;i++) {
+    x0[i] = pcm0[i] - FARGAN_DEEMPHASIS*pcm0[i-1];
+  }
+
+  OPUS_COPY(&st->pitch_buf[PITCH_MAX_PERIOD-FARGAN_FRAME_SIZE], x0, FARGAN_FRAME_SIZE);
+  st->cont_initialized = 1;
+
+  for (i=0;i<FARGAN_NB_SUBFRAMES;i++) {
+    run_fargan_subframe(st, dummy, &cond[i*FARGAN_COND_SIZE], st->last_period);
+    OPUS_COPY(&st->pitch_buf[PITCH_MAX_PERIOD-FARGAN_SUBFRAME_SIZE], &x0[FARGAN_FRAME_SIZE+i*FARGAN_SUBFRAME_SIZE], FARGAN_SUBFRAME_SIZE);
+  }
+  st->deemph_mem = pcm0[FARGAN_CONT_SAMPLES-1];
+}
+
+
+void fargan_init(FARGANState *st)
+{
+  int ret;
+  OPUS_CLEAR(st, 1);
+  ret = init_fargan(&st->model, fargan_arrays);
+  celt_assert(ret == 0);
+  /* FIXME: perform arch detection. */
+}
+
+int fargan_load_model(FARGANState *st, const unsigned char *data, int len) {
+  WeightArray *list;
+  int ret;
+  parse_weights(&list, data, len);
+  ret = init_fargan(&st->model, list);
+  free(list);
+  if (ret == 0) return 0;
+  else return -1;
+}
+
+static void fargan_synthesize_impl(FARGANState *st, float *pcm, const float *features)
+{
+  int subframe;
+  float cond[COND_NET_FCONV2_OUT_SIZE];
+  int period;
+  celt_assert(st->cont_initialized);
+
+  period = (int)floor(.5+256./pow(2.f,((1./60.)*((features[NB_BANDS]+1.5)*60))));
+  compute_fargan_cond(st, cond, features, period);
+  for (subframe=0;subframe<FARGAN_NB_SUBFRAMES;subframe++) {
+    float *sub_cond;
+    sub_cond = &cond[subframe*FARGAN_COND_SIZE];
+    run_fargan_subframe(st, &pcm[subframe*FARGAN_SUBFRAME_SIZE], sub_cond, st->last_period);
+  }
+  st->last_period = period;
+}
+
+void fargan_synthesize(FARGANState *st, float *pcm, const float *features)
+{
+  fargan_synthesize_impl(st, pcm, features);
+}
+
+void fargan_synthesize_int(FARGANState *st, opus_int16 *pcm, const float *features)
+{
+  int i;
+  float fpcm[FARGAN_FRAME_SIZE];
+  fargan_synthesize(st, fpcm, features);
+  for (i=0;i<LPCNET_FRAME_SIZE;i++) pcm[i] = (int)floor(.5 + MIN32(32767, MAX32(-32767, 32768.f*fpcm[i])));
+}

dnn/fargan.h

@@ -0,0 +1,68 @@
+/* Copyright (c) 2023 Amazon */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef FARGAN_H
+#define FARGAN_H
+
+#include "freq.h"
+#include "fargan_data.h"
+#include "pitchdnn.h"
+
+#define FARGAN_CONT_SAMPLES 320
+#define FARGAN_NB_SUBFRAMES 4
+#define FARGAN_SUBFRAME_SIZE 40
+#define FARGAN_FRAME_SIZE (FARGAN_NB_SUBFRAMES*FARGAN_SUBFRAME_SIZE)
+#define FARGAN_COND_SIZE (COND_NET_FCONV2_OUT_SIZE/FARGAN_NB_SUBFRAMES)
+#define FARGAN_DEEMPHASIS 0.85f
+
+#define SIG_NET_INPUT_SIZE (FARGAN_COND_SIZE+2*FARGAN_SUBFRAME_SIZE+4)
+#define SIG_NET_FWC0_STATE_SIZE (2*SIG_NET_INPUT_SIZE)
+
+typedef struct {
+  FARGAN model;
+  int arch;
+  int cont_initialized;
+  float deemph_mem;
+  float pitch_buf[PITCH_MAX_PERIOD];
+  float cond_conv1_state[COND_NET_FCONV1_STATE_SIZE];
+  float cond_conv2_state[COND_NET_FCONV2_STATE_SIZE];
+  float fwc0_mem[SIG_NET_FWC0_STATE_SIZE];
+  float gru1_state[SIG_NET_GRU1_STATE_SIZE];
+  float gru2_state[SIG_NET_GRU2_STATE_SIZE];
+  float gru3_state[SIG_NET_GRU3_STATE_SIZE];
+  int last_period;
+} FARGANState;
+
+void fargan_init(FARGANState *st);
+int fargan_load_model(FARGANState *st, const unsigned char *data, int len);
+
+void fargan_cont(FARGANState *st, const float *pcm0, const float *features0);
+
+void fargan_synthesize(FARGANState *st, float *pcm, const float *features);
+void fargan_synthesize_int(FARGANState *st, opus_int16 *pcm, const float *features);
+
+
+#endif /* FARGAN_H */

dnn/lpcnet_demo.c

@@ -37,6 +37,7 @@
 #include "freq.h"
 #include "os_support.h"
 #include "fwgan.h"
+#include "fargan.h"
 
 #ifdef USE_WEIGHTS_FILE
 # if __unix__
@@ -86,6 +87,7 @@ void free_blob(unsigned char *blob, int len) {
 #define MODE_PLC 4
 #define MODE_ADDLPC 5
 #define MODE_FWGAN_SYNTHESIS 6
+#define MODE_FARGAN_SYNTHESIS 7
 
 void usage(void) {
     fprintf(stderr, "usage: lpcnet_demo -features <input.pcm> <features.f32>\n");
@@ -115,6 +117,7 @@ int main(int argc, char **argv) {
     if (strcmp(argv[1], "-features") == 0) mode=MODE_FEATURES;
     else if (strcmp(argv[1], "-synthesis") == 0) mode=MODE_SYNTHESIS;
     else if (strcmp(argv[1], "-fwgan-synthesis") == 0) mode=MODE_FWGAN_SYNTHESIS;
+    else if (strcmp(argv[1], "-fargan-synthesis") == 0) mode=MODE_FARGAN_SYNTHESIS;
     else if (strcmp(argv[1], "-plc") == 0) {
         mode=MODE_PLC;
         plc_options = argv[2];
@@ -210,6 +213,32 @@ int main(int argc, char **argv) {
             for (i=0;i<LPCNET_FRAME_SIZE;i++) pcm[i] = (int)floor(.5 + MIN32(32767, MAX32(-32767, 32768.f*fpcm[i])));
             fwrite(pcm, sizeof(pcm[0]), LPCNET_FRAME_SIZE, fout);
         }
+    } else if (mode == MODE_FARGAN_SYNTHESIS) {
+        FARGANState fargan;
+        size_t ret, i;
+        float in_features[5*NB_TOTAL_FEATURES];
+        float zeros[320] = {0};
+        fargan_init(&fargan);
+#ifdef USE_WEIGHTS_FILE
+        fargan_load_model(fwgan, data, len);
+#endif
+        /* uncomment the following to align with Python code */
+        /*ret = fread(&in_features[0], sizeof(in_features[0]), NB_TOTAL_FEATURES, fin);*/
+        for (i=0;i<5;i++) {
+          ret = fread(&in_features[i*NB_FEATURES], sizeof(in_features[0]), NB_TOTAL_FEATURES, fin);
+        }
+        fargan_cont(&fargan, zeros, in_features);
+        while (1) {
+            float features[NB_FEATURES];
+            float fpcm[LPCNET_FRAME_SIZE];
+            opus_int16 pcm[LPCNET_FRAME_SIZE];
+            ret = fread(in_features, sizeof(features[0]), NB_TOTAL_FEATURES, fin);
+            if (feof(fin) || ret != NB_TOTAL_FEATURES) break;
+            OPUS_COPY(features, in_features, NB_FEATURES);
+            fargan_synthesize(&fargan, fpcm, features);
+            for (i=0;i<LPCNET_FRAME_SIZE;i++) pcm[i] = (int)floor(.5 + MIN32(32767, MAX32(-32767, 32768.f*fpcm[i])));
+            fwrite(pcm, sizeof(pcm[0]), LPCNET_FRAME_SIZE, fout);
+        }
     } else if (mode == MODE_PLC) {
         opus_int16 pcm[FRAME_SIZE];
         int count=0;

dnn/nnet.c

@@ -143,6 +143,16 @@ void compute_generic_gru(const LinearLayer *input_weights, const LinearLayer *re
      state[i] = h[i];
 }
 
+void compute_glu(const LinearLayer *layer, float *output, const float *input)
+{
+   int i;
+   float act2[MAX_INPUTS];
+   celt_assert(layer->nb_inputs == layer->nb_outputs);
+   compute_linear(layer, act2, input);
+   compute_activation(act2, act2, layer->nb_outputs, ACTIVATION_SIGMOID);
+   for (i=0;i<layer->nb_outputs;i++) output[i] = input[i]*act2[i];
+}
+
 void compute_gated_activation(const LinearLayer *layer, float *output, const float *input, int activation)
 {
    int i;

dnn/nnet.h

@@ -146,6 +146,7 @@ void compute_generic_dense(const LinearLayer *layer, float *output, const float
 void compute_generic_gru(const LinearLayer *input_weights, const LinearLayer *recurrent_weights, float *state, const float *in);
 void compute_generic_conv1d(const LinearLayer *layer, float *output, float *mem, const float *input, int input_size, int activation);
 void compute_generic_conv1d_dilation(const LinearLayer *layer, float *output, float *mem, const float *input, int input_size, int dilation, int activation);
+void compute_glu(const LinearLayer *layer, float *output, const float *input);
 void compute_gated_activation(const LinearLayer *layer, float *output, const float *input, int activation);
 
 void compute_activation(float *output, const float *input, int N, int activation);
@@ -176,6 +177,7 @@ extern const WeightArray lpcnet_plc_arrays[];
 extern const WeightArray rdovaeenc_arrays[];
 extern const WeightArray rdovaedec_arrays[];
 extern const WeightArray fwgan_arrays[];
+extern const WeightArray fargan_arrays[];
 extern const WeightArray pitchdnn_arrays[];
 
 int linear_init(LinearLayer *layer, const WeightArray *arrays,

dnn/torch/fargan/dump_fargan_weights.py

@@ -0,0 +1,112 @@
+import os
+import sys
+import argparse
+
+import torch
+from torch import nn
+
+
+sys.path.append(os.path.join(os.path.split(__file__)[0], '../weight-exchange'))
+import wexchange.torch
+
+import fargan
+#from models import model_dict
+
+unquantized = [ 'cond_net.pembed', 'cond_net.fdense1', 'sig_net.cond_gain_dense', 'sig_net.gain_dense_out' ]
+
+unquantized2 = [
+    'cond_net.pembed',
+    'cond_net.fdense1',
+    'cond_net.fconv1',
+    'cond_net.fconv2',
+    'cont_net.0',
+    'sig_net.cond_gain_dense',
+    'sig_net.fwc0.conv',
+    'sig_net.fwc0.glu.gate',
+    'sig_net.dense1_glu.gate',
+    'sig_net.gru1_glu.gate',
+    'sig_net.gru2_glu.gate',
+    'sig_net.gru3_glu.gate',
+    'sig_net.skip_glu.gate',
+    'sig_net.skip_dense',
+    'sig_net.sig_dense_out',
+    'sig_net.gain_dense_out'
+]
+
+description=f"""
+This is an unsafe dumping script for FARGAN models. It assumes that all weights are included in Linear, Conv1d or GRU layer
+and will fail to export any other weights.
+
+Furthermore, the quanitze option relies on the following explicit list of layers to be excluded:
+{unquantized}.
+
+Modify this script manually if adjustments are needed.
+"""
+
+parser = argparse.ArgumentParser(description=description)
+parser.add_argument('weightfile', type=str, help='weight file path')
+parser.add_argument('export_folder', type=str)
+parser.add_argument('--export-filename', type=str, default='fargan_data', help='filename for source and header file (.c and .h will be added), defaults to fargan_data')
+parser.add_argument('--struct-name', type=str, default='FARGAN', help='name for C struct, defaults to FARGAN')
+parser.add_argument('--quantize', action='store_true', help='apply quantization')
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+
+    print(f"loading weights from {args.weightfile}...")
+    saved_gen= torch.load(args.weightfile, map_location='cpu')
+    saved_gen['model_args']    = ()
+    saved_gen['model_kwargs']  = {'cond_size': 256, 'gamma': 0.9}
+
+    model = fargan.FARGAN(*saved_gen['model_args'], **saved_gen['model_kwargs'])
+    model.load_state_dict(saved_gen['state_dict'], strict=False)
+    def _remove_weight_norm(m):
+        try:
+            torch.nn.utils.remove_weight_norm(m)
+        except ValueError:  # this module didn't have weight norm
+            return
+    model.apply(_remove_weight_norm)
+
+
+    print("dumping model...")
+    quantize_model=args.quantize
+
+    output_folder = args.export_folder
+    os.makedirs(output_folder, exist_ok=True)
+
+    writer = wexchange.c_export.c_writer.CWriter(os.path.join(output_folder, args.export_filename), model_struct_name=args.struct_name)
+
+    for name, module in model.named_modules():
+
+        if quantize_model:
+            quantize=name not in unquantized
+            scale = None if quantize else 1/128
+        else:
+            quantize=False
+            scale=1/128
+
+        if isinstance(module, nn.Linear):
+            print(f"dumping linear layer {name}...")
+            wexchange.torch.dump_torch_dense_weights(writer, module, name.replace('.', '_'), quantize=quantize, scale=scale)
+
+        elif isinstance(module, nn.Conv1d):
+            print(f"dumping conv1d layer {name}...")
+            wexchange.torch.dump_torch_conv1d_weights(writer, module, name.replace('.', '_'), quantize=quantize, scale=scale)
+
+        elif isinstance(module, nn.GRU):
+            print(f"dumping GRU layer {name}...")
+            wexchange.torch.dump_torch_gru_weights(writer, module, name.replace('.', '_'), quantize=quantize, scale=scale, recurrent_scale=scale)
+
+        elif isinstance(module, nn.GRUCell):
+            print(f"dumping GRUCell layer {name}...")
+            wexchange.torch.dump_torch_grucell_weights(writer, module, name.replace('.', '_'), quantize=quantize, scale=scale, recurrent_scale=scale)
+
+        elif isinstance(module, nn.Embedding):
+            print(f"dumping Embedding layer {name}...")
+            wexchange.torch.dump_torch_embedding_weights(writer, module, name.replace('.', '_'), quantize=quantize, scale=scale)
+            #wexchange.torch.dump_torch_embedding_weights(writer, module)
+
+        else:
+            print(f"Ignoring layer {name}...")
+
+    writer.close()

dnn/torch/weight-exchange/wexchange/torch/__init__.py

@@ -31,5 +31,6 @@ from .torch import dump_torch_conv1d_weights, load_torch_conv1d_weights
 from .torch import dump_torch_conv2d_weights, load_torch_conv2d_weights
 from .torch import dump_torch_dense_weights, load_torch_dense_weights
 from .torch import dump_torch_gru_weights, load_torch_gru_weights
+from .torch import dump_torch_grucell_weights
 from .torch import dump_torch_embedding_weights, load_torch_embedding_weights
 from .torch import dump_torch_weights, load_torch_weights

dnn/torch/weight-exchange/wexchange/torch/torch.py

@@ -61,6 +61,30 @@ def dump_torch_gru_weights(where, gru, name='gru', input_sparse=False, recurrent
         np.save(os.path.join(where, 'bias_hh_rzn.npy'), b_hh)
 
 
+def dump_torch_grucell_weights(where, gru, name='gru', input_sparse=False, recurrent_sparse=False, quantize=False, scale=1/128, recurrent_scale=1/128):
+
+    w_ih = gru.weight_ih.detach().cpu().numpy().copy()
+    w_hh = gru.weight_hh.detach().cpu().numpy().copy()
+    if hasattr(gru, 'bias_ih') and gru.bias_ih is not None:
+        b_ih = gru.bias_ih.detach().cpu().numpy().copy()
+    else:
+        b_ih = None
+    if hasattr(gru, 'bias_hh') and gru.bias_hh is not None:
+        b_hh = gru.bias_hh.detach().cpu().numpy().copy()
+    else:
+        b_hh = None
+
+    if isinstance(where, CWriter):
+        return print_gru_layer(where, name, w_ih, w_hh, b_ih, b_hh, format='torch', input_sparse=input_sparse, recurrent_sparse=recurrent_sparse, quantize=quantize, scale=scale, recurrent_scale=recurrent_scale)
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight_ih_rzn.npy'), w_ih)
+        np.save(os.path.join(where, 'weight_hh_rzn.npy'), w_hh)
+        np.save(os.path.join(where, 'bias_ih_rzn.npy'), b_ih)
+        np.save(os.path.join(where, 'bias_hh_rzn.npy'), b_hh)
+
+
 
 def load_torch_gru_weights(where, gru):
 
@@ -165,11 +189,20 @@ def load_torch_conv2d_weights(where, conv):
                 conv.bias.set_(torch.from_numpy(b))
 
 
-def dump_torch_embedding_weights(where, emb):
+def dump_torch_embedding_weights(where, embed, name='embed', scale=1/128, sparse=False, diagonal=False, quantize=False):
-    os.makedirs(where, exist_ok=True)
 
-    w = emb.weight.detach().cpu().numpy().copy()
+    print("quantize = ", quantize)
-    np.save(os.path.join(where, 'weight.npy'), w)
+    w = embed.weight.detach().cpu().numpy().copy().transpose()
+    b = np.zeros(w.shape[0], dtype=w.dtype)
+
+    if isinstance(where, CWriter):
+        return print_dense_layer(where, name, w, b, scale=scale, format='torch', sparse=sparse, diagonal=diagonal, quantize=quantize)
+
+    else:
+        os.makedirs(where, exist_ok=True)
+
+        np.save(os.path.join(where, 'weight.npy'), w)
+        np.save(os.path.join(where, 'bias.npy'), b)
 
 
 def load_torch_embedding_weights(where, emb):
@@ -187,6 +220,8 @@ def dump_torch_weights(where, module, name=None, verbose=False, **kwargs):
         return dump_torch_dense_weights(where, module, name, **kwargs)
     elif isinstance(module, torch.nn.GRU):
         return dump_torch_gru_weights(where, module, name, **kwargs)
+    elif isinstance(module, torch.nn.GRUCell):
+        return dump_torch_grucell_weights(where, module, name, **kwargs)
     elif isinstance(module, torch.nn.Conv1d):
         return dump_torch_conv1d_weights(where, module, name, **kwargs)
     elif isinstance(module, torch.nn.Conv2d):
@@ -209,4 +244,4 @@ def load_torch_weights(where, module):
     elif isinstance(module, torch.nn.Embedding):
         load_torch_embedding_weights(where, module)
     else:
-        raise ValueError(f'dump_torch_weights: layer of type {type(module)} not supported')
+        raise ValueError(f'dump_torch_weights: layer of type {type(module)} not supported')

lpcnet_headers.mk

@@ -8,6 +8,8 @@ dnn/lpcnet.h \
 dnn/burg.h \
 dnn/common.h \
 dnn/freq.h \
+dnn/fargan.h \
+dnn/fargan_data.h \
 dnn/fwgan.h \
 dnn/fwgan_data.h \
 dnn/kiss99.h \

lpcnet_sources.mk

@@ -1,6 +1,8 @@
 LPCNET_SOURCES = \
 dnn/burg.c \
 dnn/freq.c \
+dnn/fargan.c \
+dnn/fargan_data.c \
 dnn/fwgan.c \
 dnn/fwgan_data.c \
 dnn/kiss99.c \