eden-emu/opus
17
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Pitch and fargan model updates

Browse source 
Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.
This commit is contained in:
Jean-Marc Valin 2023-10-28 23:33:47 -04:00
parent c99054dad9
commit ddd5669e79
No known key found for this signature in database
GPG key ID: 531A52533318F00A
5 changed files with 7 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

2
autogen.sh
View file

@ -9,7 +9,7 @@ set -e
srcdir=`dirname $0` srcdir=`dirname $0`
test -n "$srcdir" && cd "$srcdir" test -n "$srcdir" && cd "$srcdir"
dnn/download_model.sh 290be25 dnn/download_model.sh c99054d
echo "Updating build configuration files, please wait...." echo "Updating build configuration files, please wait...."

4
dnn/lpcnet_enc.c
View file

@ -81,8 +81,8 @@ static void biquad(float *y, float mem[2], const float *x, const float *b, const
float xi, yi; float xi, yi;
xi = x[i]; xi = x[i];
yi = x[i] + mem[0]; yi = x[i] + mem[0];
mem[0] = mem[1] + (b[0]*(double)xi - a[0]*(double)yi); mem[0] = mem[1] + (b[0]*xi - a[0]*yi);
mem[1] = (b[1]*(double)xi - a[1]*(double)yi); mem[1] = (b[1]*xi - a[1]*yi);
y[i] = yi; y[i] = yi;
} }
} }

3
dnn/pitchdnn.c
View file

@ -33,8 +33,7 @@ float compute_pitchdnn(
/* xcorr*/ /* xcorr*/
OPUS_COPY(&conv1_tmp1[1], xcorr_features, NB_XCORR_FEATURES); OPUS_COPY(&conv1_tmp1[1], xcorr_features, NB_XCORR_FEATURES);
compute_conv2d(&model->conv2d_1, &conv1_tmp2[1], st->xcorr_mem1, conv1_tmp1, NB_XCORR_FEATURES, NB_XCORR_FEATURES+2, ACTIVATION_TANH); compute_conv2d(&model->conv2d_1, &conv1_tmp2[1], st->xcorr_mem1, conv1_tmp1, NB_XCORR_FEATURES, NB_XCORR_FEATURES+2, ACTIVATION_TANH);
compute_conv2d(&model->conv2d_2, &conv1_tmp1[1], st->xcorr_mem2, conv1_tmp2, NB_XCORR_FEATURES, NB_XCORR_FEATURES+2, ACTIVATION_TANH); compute_conv2d(&model->conv2d_2, downsampler_in, st->xcorr_mem2, conv1_tmp2, NB_XCORR_FEATURES, NB_XCORR_FEATURES, ACTIVATION_TANH);
compute_conv2d(&model->conv2d_3, downsampler_in, st->xcorr_mem3, conv1_tmp1, NB_XCORR_FEATURES, NB_XCORR_FEATURES, ACTIVATION_TANH);
compute_generic_dense(&model->dense_downsampler, downsampler_out, downsampler_in, ACTIVATION_TANH); compute_generic_dense(&model->dense_downsampler, downsampler_out, downsampler_in, ACTIVATION_TANH);
compute_generic_gru(&model->gru_1_input, &model->gru_1_recurrent, st->gru_state, downsampler_out); compute_generic_gru(&model->gru_1_input, &model->gru_1_recurrent, st->gru_state, downsampler_out);

3
dnn/torch/neural-pitch/export_neuralpitch_weights.py
View file

@ -73,8 +73,7 @@ f"""
conv_layers = [ conv_layers = [
('conv.1', "conv2d_1"), ('conv.1', "conv2d_1"),
('conv.4', "conv2d_2"), ('conv.4', "conv2d_2")
('conv.7', "conv2d_3")
] ]

7
dnn/torch/neural-pitch/models.py
View file

@ -86,13 +86,10 @@ class PitchDNN(torch.nn.Module):
self.conv = torch.nn.Sequential( self.conv = torch.nn.Sequential(
torch.nn.ZeroPad2d((2,0,1,1)), torch.nn.ZeroPad2d((2,0,1,1)),
torch.nn.Conv2d(1, 8, 3, bias=True), torch.nn.Conv2d(1, 4, 3, bias=True),
self.activation, self.activation,
torch.nn.ZeroPad2d((2,0,1,1)), torch.nn.ZeroPad2d((2,0,1,1)),
torch.nn.Conv2d(8, 8, 3, bias=True), torch.nn.Conv2d(4, 1, 3, bias=True),
self.activation,
torch.nn.ZeroPad2d((2,0,1,1)),
torch.nn.Conv2d(8, 1, 3, bias=True),
self.activation, self.activation,
) )