@scrypted/libav 1.0.167 → 1.0.169

package/opus/dnn/torch/fwgan/models/fwgan400.py

@@ -0,0 +1,308 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.utils import weight_norm
+import numpy as np
+
+which_norm = weight_norm
+
+#################### Definition of basic model components ####################
+
+#Convolutional layer with 1 frame look-ahead (used for feature PreCondNet)
+class ConvLookahead(nn.Module):
+    def __init__(self, in_ch, out_ch, kernel_size, dilation=1, groups=1, bias= False):
+        super(ConvLookahead, self).__init__()
+        torch.manual_seed(5)
+
+        self.padding_left = (kernel_size - 2) * dilation
+        self.padding_right = 1 * dilation
+
+        self.conv = which_norm(nn.Conv1d(in_ch,out_ch,kernel_size,dilation=dilation, groups=groups, bias= bias))
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x):
+
+        x = F.pad(x,(self.padding_left, self.padding_right))
+        conv_out = self.conv(x)
+        return conv_out
+
+#(modified) GLU Activation layer definition
+class GLU(nn.Module):
+    def __init__(self, feat_size):
+        super(GLU, self).__init__()
+
+        torch.manual_seed(5)
+
+        self.gate = which_norm(nn.Linear(feat_size, feat_size, bias=False))
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d)\
+            or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x):
+
+        out = torch.tanh(x) * torch.sigmoid(self.gate(x))
+
+        return out
+
+#GRU layer definition
+class ContForwardGRU(nn.Module):
+    def __init__(self, input_size, hidden_size, num_layers=1):
+        super(ContForwardGRU, self).__init__()
+
+        torch.manual_seed(5)
+
+        self.hidden_size = hidden_size
+
+        self.cont_fc = nn.Sequential(which_norm(nn.Linear(64, self.hidden_size, bias=False)),
+                        nn.Tanh())
+
+        self.gru = nn.GRU(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers, batch_first=True,\
+                          bias=False)
+
+        self.nl = GLU(self.hidden_size)
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x, x0):
+
+        self.gru.flatten_parameters()
+
+        h0 = self.cont_fc(x0).unsqueeze(0)
+
+        output, h0 = self.gru(x, h0)
+
+        return self.nl(output)
+
+# Framewise convolution layer definition
+class ContFramewiseConv(torch.nn.Module):
+
+    def __init__(self, frame_len, out_dim, frame_kernel_size=3, act='glu', causal=True):
+
+        super(ContFramewiseConv, self).__init__()
+        torch.manual_seed(5)
+
+        self.frame_kernel_size = frame_kernel_size
+        self.frame_len = frame_len
+
+        if (causal == True) or (self.frame_kernel_size == 2):
+
+            self.required_pad_left = (self.frame_kernel_size - 1) * self.frame_len
+            self.required_pad_right = 0
+
+            self.cont_fc = nn.Sequential(which_norm(nn.Linear(64, self.required_pad_left, bias=False)),
+                                    nn.Tanh()
+                                   )
+
+        else:
+
+            self.required_pad_left = (self.frame_kernel_size - 1)//2 * self.frame_len
+            self.required_pad_right = (self.frame_kernel_size - 1)//2 * self.frame_len
+
+        self.fc_input_dim = self.frame_kernel_size * self.frame_len
+        self.fc_out_dim = out_dim
+
+        if act=='glu':
+            self.fc = nn.Sequential(which_norm(nn.Linear(self.fc_input_dim, self.fc_out_dim, bias=False)),
+                                    GLU(self.fc_out_dim)
+                                   )
+        if act=='tanh':
+            self.fc = nn.Sequential(which_norm(nn.Linear(self.fc_input_dim, self.fc_out_dim, bias=False)),
+                                    nn.Tanh()
+                                   )
+
+        self.init_weights()
+
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or\
+            isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x, x0):
+
+        if self.frame_kernel_size == 1:
+            return self.fc(x)
+
+        x_flat = x.reshape(x.size(0),1,-1)
+        pad = self.cont_fc(x0).view(x0.size(0),1,-1)
+        x_flat_padded = torch.cat((pad, x_flat), dim=-1).unsqueeze(2)
+
+        x_flat_padded_unfolded = F.unfold(x_flat_padded,\
+                    kernel_size= (1,self.fc_input_dim), stride=self.frame_len).permute(0,2,1).contiguous()
+
+        out = self.fc(x_flat_padded_unfolded)
+        return out
+
+# A fully-connected based upsampling layer definition
+class UpsampleFC(nn.Module):
+    def __init__(self, in_ch, out_ch, upsample_factor):
+        super(UpsampleFC, self).__init__()
+        torch.manual_seed(5)
+
+        self.in_ch = in_ch
+        self.out_ch = out_ch
+        self.upsample_factor = upsample_factor
+        self.fc = nn.Linear(in_ch, out_ch * upsample_factor, bias=False)
+        self.nl = nn.Tanh()
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or\
+            isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x):
+
+        batch_size = x.size(0)
+        x = x.permute(0, 2, 1)
+        x = self.nl(self.fc(x))
+        x = x.reshape((batch_size, -1, self.out_ch))
+        x = x.permute(0, 2, 1)
+        return x
+
+########################### The complete model definition #################################
+
+class FWGAN400ContLarge(nn.Module):
+    def __init__(self):
+        super().__init__()
+        torch.manual_seed(5)
+
+        self.bfcc_with_corr_upsampler = UpsampleFC(19,80,4)
+
+        self.feat_in_conv1 = ConvLookahead(160,256,kernel_size=5)
+        self.feat_in_nl1 = GLU(256)
+
+        self.cont_net = nn.Sequential(which_norm(nn.Linear(321, 160, bias=False)),
+                                      nn.Tanh(),
+                                      which_norm(nn.Linear(160, 160, bias=False)),
+                                      nn.Tanh(),
+                                      which_norm(nn.Linear(160, 80, bias=False)),
+                                      nn.Tanh(),
+                                      which_norm(nn.Linear(80, 80, bias=False)),
+                                      nn.Tanh(),
+                                      which_norm(nn.Linear(80, 64, bias=False)),
+                                      nn.Tanh(),
+                                      which_norm(nn.Linear(64, 64, bias=False)),
+                                      nn.Tanh())
+
+        self.rnn = ContForwardGRU(256,256)
+
+        self.fwc1 = ContFramewiseConv(256, 256)
+        self.fwc2 = ContFramewiseConv(256, 128)
+        self.fwc3 = ContFramewiseConv(128, 128)
+        self.fwc4 = ContFramewiseConv(128, 64)
+        self.fwc5 = ContFramewiseConv(64, 64)
+        self.fwc6 = ContFramewiseConv(64, 40)
+        self.fwc7 = ContFramewiseConv(40, 40)
+
+        self.init_weights()
+        self.count_parameters()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or\
+            isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def count_parameters(self):
+        num_params =  sum(p.numel() for p in self.parameters() if p.requires_grad)
+        print(f"Total number of {self.__class__.__name__} network parameters = {num_params}\n")
+
+    def create_phase_signals(self, periods):
+
+        batch_size = periods.size(0)
+        progression = torch.arange(1, 160 + 1, dtype=periods.dtype, device=periods.device).view((1, -1))
+        progression = torch.repeat_interleave(progression, batch_size, 0)
+
+        phase0 = torch.zeros(batch_size, dtype=periods.dtype, device=periods.device).unsqueeze(-1)
+        chunks = []
+        for sframe in range(periods.size(1)):
+            f = (2.0 * torch.pi / periods[:, sframe]).unsqueeze(-1)
+
+            chunk_sin = torch.sin(f  * progression + phase0)
+            chunk_sin = chunk_sin.reshape(chunk_sin.size(0),-1,40)
+
+            chunk_cos = torch.cos(f  * progression + phase0)
+            chunk_cos = chunk_cos.reshape(chunk_cos.size(0),-1,40)
+
+            chunk = torch.cat((chunk_sin, chunk_cos), dim = -1)
+
+            phase0 = phase0 + 160 * f
+
+            chunks.append(chunk)
+
+        phase_signals = torch.cat(chunks, dim=1)
+
+        return phase_signals
+
+
+    def gain_multiply(self, x, c0):
+
+        gain = 10**(0.5*c0/np.sqrt(18.0))
+        gain = torch.repeat_interleave(gain, 160, dim=-1)
+        gain  = gain.reshape(gain.size(0),1,-1).squeeze(1)
+
+        return x * gain
+
+    def forward(self, pitch_period, bfcc_with_corr, x0):
+
+        norm_x0 = torch.norm(x0,2, dim=-1, keepdim=True)
+        x0 = x0 / torch.sqrt((1e-8) +  norm_x0**2)
+        x0 = torch.cat((torch.log(norm_x0 + 1e-7), x0), dim=-1)
+
+        p_embed = self.create_phase_signals(pitch_period).permute(0, 2, 1).contiguous()
+
+        envelope = self.bfcc_with_corr_upsampler(bfcc_with_corr.permute(0,2,1).contiguous())
+
+        feat_in = torch.cat((p_embed , envelope), dim=1)
+
+        wav_latent1 = self.feat_in_nl1(self.feat_in_conv1(feat_in).permute(0,2,1).contiguous())
+
+        cont_latent = self.cont_net(x0)
+
+        rnn_out = self.rnn(wav_latent1, cont_latent)
+
+        fwc1_out = self.fwc1(rnn_out, cont_latent)
+
+        fwc2_out = self.fwc2(fwc1_out, cont_latent)
+
+        fwc3_out = self.fwc3(fwc2_out, cont_latent)
+
+        fwc4_out = self.fwc4(fwc3_out, cont_latent)
+
+        fwc5_out = self.fwc5(fwc4_out, cont_latent)
+
+        fwc6_out = self.fwc6(fwc5_out, cont_latent)
+
+        fwc7_out = self.fwc7(fwc6_out, cont_latent)
+
+        waveform = fwc7_out.reshape(fwc7_out.size(0),1,-1).squeeze(1)
+
+        waveform = self.gain_multiply(waveform,bfcc_with_corr[:,:,:1])
+
+        return  waveform

package/opus/dnn/torch/fwgan/models/fwgan500.py

@@ -0,0 +1,260 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.utils import weight_norm
+import numpy as np
+
+
+which_norm = weight_norm
+
+#################### Definition of basic model components ####################
+
+#Convolutional layer with 1 frame look-ahead (used for feature PreCondNet)
+class ConvLookahead(nn.Module):
+    def __init__(self, in_ch, out_ch, kernel_size, dilation=1, groups=1, bias= False):
+        super(ConvLookahead, self).__init__()
+        torch.manual_seed(5)
+
+        self.padding_left = (kernel_size - 2) * dilation
+        self.padding_right = 1 * dilation
+
+        self.conv = which_norm(nn.Conv1d(in_ch,out_ch,kernel_size,dilation=dilation, groups=groups, bias= bias))
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x):
+
+        x = F.pad(x,(self.padding_left, self.padding_right))
+        conv_out = self.conv(x)
+        return conv_out
+
+#(modified) GLU Activation layer definition
+class GLU(nn.Module):
+    def __init__(self, feat_size):
+        super(GLU, self).__init__()
+
+        torch.manual_seed(5)
+
+        self.gate = which_norm(nn.Linear(feat_size, feat_size, bias=False))
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d)\
+            or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x):
+
+        out = torch.tanh(x) * torch.sigmoid(self.gate(x))
+
+        return out
+
+#GRU layer definition
+class ContForwardGRU(nn.Module):
+    def __init__(self, input_size, hidden_size, num_layers=1):
+        super(ContForwardGRU, self).__init__()
+
+        torch.manual_seed(5)
+
+        self.hidden_size = hidden_size
+
+        #This is to initialize the layer with history audio samples for continuation.
+        self.cont_fc = nn.Sequential(which_norm(nn.Linear(320, self.hidden_size, bias=False)),
+                        nn.Tanh())
+
+        self.gru = nn.GRU(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers, batch_first=True,\
+                          bias=False)
+
+        self.nl = GLU(self.hidden_size)
+
+        self.init_weights()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x, x0):
+
+        self.gru.flatten_parameters()
+
+        h0 = self.cont_fc(x0).unsqueeze(0)
+
+        output, h0 = self.gru(x, h0)
+
+        return self.nl(output)
+
+# Framewise convolution layer definition
+class ContFramewiseConv(torch.nn.Module):
+
+    def __init__(self, frame_len, out_dim, frame_kernel_size=3, act='glu', causal=True):
+
+        super(ContFramewiseConv, self).__init__()
+        torch.manual_seed(5)
+
+        self.frame_kernel_size = frame_kernel_size
+        self.frame_len = frame_len
+
+        if (causal == True) or (self.frame_kernel_size == 2):
+
+            self.required_pad_left = (self.frame_kernel_size - 1) * self.frame_len
+            self.required_pad_right = 0
+
+            #This is to initialize the layer with history audio samples for continuation.
+            self.cont_fc = nn.Sequential(which_norm(nn.Linear(320, self.required_pad_left, bias=False)),
+                                    nn.Tanh()
+                                   )
+
+        else:
+            #This means non-causal frame-wise convolution. We don't use it at the moment
+            self.required_pad_left = (self.frame_kernel_size - 1)//2 * self.frame_len
+            self.required_pad_right = (self.frame_kernel_size - 1)//2 * self.frame_len
+
+        self.fc_input_dim = self.frame_kernel_size * self.frame_len
+        self.fc_out_dim = out_dim
+
+        if act=='glu':
+            self.fc = nn.Sequential(which_norm(nn.Linear(self.fc_input_dim, self.fc_out_dim, bias=False)),
+                                    GLU(self.fc_out_dim)
+                                   )
+        if act=='tanh':
+            self.fc = nn.Sequential(which_norm(nn.Linear(self.fc_input_dim, self.fc_out_dim, bias=False)),
+                                    nn.Tanh()
+                                   )
+
+        self.init_weights()
+
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or\
+            isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def forward(self, x, x0):
+
+        if self.frame_kernel_size == 1:
+            return self.fc(x)
+
+        x_flat = x.reshape(x.size(0),1,-1)
+        pad = self.cont_fc(x0).view(x0.size(0),1,-1)
+        x_flat_padded = torch.cat((pad, x_flat), dim=-1).unsqueeze(2)
+
+        x_flat_padded_unfolded = F.unfold(x_flat_padded,\
+                    kernel_size= (1,self.fc_input_dim), stride=self.frame_len).permute(0,2,1).contiguous()
+
+        out = self.fc(x_flat_padded_unfolded)
+        return out
+
+########################### The complete model definition #################################
+
+class FWGAN500Cont(nn.Module):
+    def __init__(self):
+        super().__init__()
+        torch.manual_seed(5)
+
+        #PrecondNet:
+        self.bfcc_with_corr_upsampler = nn.Sequential(nn.ConvTranspose1d(19,64,kernel_size=5,stride=5,padding=0,\
+                                                      bias=False),
+                                                      nn.Tanh())
+
+        self.feat_in_conv = ConvLookahead(128,256,kernel_size=5)
+        self.feat_in_nl = GLU(256)
+
+        #GRU:
+        self.rnn = ContForwardGRU(256,256)
+
+        #Frame-wise convolution stack:
+        self.fwc1 = ContFramewiseConv(256, 256)
+        self.fwc2 = ContFramewiseConv(256, 128)
+        self.fwc3 = ContFramewiseConv(128, 128)
+        self.fwc4 = ContFramewiseConv(128, 64)
+        self.fwc5 = ContFramewiseConv(64, 64)
+        self.fwc6 = ContFramewiseConv(64, 32)
+        self.fwc7 = ContFramewiseConv(32, 32, act='tanh')
+
+        self.init_weights()
+        self.count_parameters()
+
+    def init_weights(self):
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d) or isinstance(m, nn.ConvTranspose1d) or isinstance(m, nn.Linear) or\
+            isinstance(m, nn.Embedding):
+                nn.init.orthogonal_(m.weight.data)
+
+    def count_parameters(self):
+        num_params =  sum(p.numel() for p in self.parameters() if p.requires_grad)
+        print(f"Total number of {self.__class__.__name__} network parameters = {num_params}\n")
+
+    def create_phase_signals(self, periods):
+
+        batch_size = periods.size(0)
+        progression = torch.arange(1, 160 + 1, dtype=periods.dtype, device=periods.device).view((1, -1))
+        progression = torch.repeat_interleave(progression, batch_size, 0)
+
+        phase0 = torch.zeros(batch_size, dtype=periods.dtype, device=periods.device).unsqueeze(-1)
+        chunks = []
+        for sframe in range(periods.size(1)):
+            f = (2.0 * torch.pi / periods[:, sframe]).unsqueeze(-1)
+
+            chunk_sin = torch.sin(f  * progression + phase0)
+            chunk_sin = chunk_sin.reshape(chunk_sin.size(0),-1,32)
+
+            chunk_cos = torch.cos(f  * progression + phase0)
+            chunk_cos = chunk_cos.reshape(chunk_cos.size(0),-1,32)
+
+            chunk = torch.cat((chunk_sin, chunk_cos), dim = -1)
+
+            phase0 = phase0 + 160 * f
+
+            chunks.append(chunk)
+
+        phase_signals = torch.cat(chunks, dim=1)
+
+        return phase_signals
+
+
+    def gain_multiply(self, x, c0):
+
+        gain = 10**(0.5*c0/np.sqrt(18.0))
+        gain = torch.repeat_interleave(gain, 160, dim=-1)
+        gain  = gain.reshape(gain.size(0),1,-1).squeeze(1)
+
+        return x * gain
+
+    def forward(self, pitch_period, bfcc_with_corr, x0):
+
+        #This should create a latent representation of shape [Batch_dim, 500 frames, 256 elemets per frame]
+        p_embed = self.create_phase_signals(pitch_period).permute(0, 2, 1).contiguous()
+        envelope = self.bfcc_with_corr_upsampler(bfcc_with_corr.permute(0,2,1).contiguous())
+        feat_in = torch.cat((p_embed , envelope), dim=1)
+        wav_latent = self.feat_in_nl(self.feat_in_conv(feat_in).permute(0,2,1).contiguous())
+
+        #Generation with continuation using history samples x0 starts from here:
+
+        rnn_out = self.rnn(wav_latent, x0)
+
+        fwc1_out = self.fwc1(rnn_out, x0)
+        fwc2_out = self.fwc2(fwc1_out, x0)
+        fwc3_out = self.fwc3(fwc2_out, x0)
+        fwc4_out = self.fwc4(fwc3_out, x0)
+        fwc5_out = self.fwc5(fwc4_out, x0)
+        fwc6_out = self.fwc6(fwc5_out, x0)
+        fwc7_out = self.fwc7(fwc6_out, x0)
+
+        waveform_unscaled = fwc7_out.reshape(fwc7_out.size(0),1,-1).squeeze(1)
+        waveform = self.gain_multiply(waveform_unscaled,bfcc_with_corr[:,:,:1])
+
+        return  waveform

package/opus/dnn/torch/lossgen/README.md

@@ -0,0 +1,27 @@
+#Packet loss simulator
+
+This code is an attempt at simulating better packet loss scenarios. The most common way of simulating
+packet loss is to use a random sequence where each packet loss event is uncorrelated with previous events.
+That is a simplistic model since we know that losses often occur in bursts. This model uses real data
+to build a generative model for packet loss.
+
+We use the training data provided for the Audio Deep Packet Loss Concealment Challenge, which is available at:
+
+http://plcchallenge2022pub.blob.core.windows.net/plcchallengearchive/test_train.tar.gz
+
+To create the training data, run:
+
+`./process_data.sh /<path>/test_train/train/lossy_signals/`
+
+That will create an ascii loss\_sorted.txt file with all loss data sorted in increasing packet loss
+percentage. Then just run:
+
+`python ./train_lossgen.py`
+
+to train a model
+
+To generate a sequence, run
+
+`python3 ./test_lossgen.py <checkpoint> <percentage> output.txt --length 10000`
+
+where <checkpoint> is the .pth model file and <percentage> is the amount of loss (e.g. 0.2 for 20% loss).

package/opus/dnn/torch/lossgen/export_lossgen.py

@@ -0,0 +1,101 @@
+"""
+/* Copyright (c) 2022 Amazon
+   Written by Jan Buethe */
+/*
+   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.
+*/
+"""
+
+import os
+import argparse
+import sys
+
+sys.path.append(os.path.join(os.path.dirname(__file__), '../weight-exchange'))
+
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument('checkpoint', type=str, help='model checkpoint')
+parser.add_argument('output_dir', type=str, help='output folder')
+
+args = parser.parse_args()
+
+import torch
+import numpy as np
+
+import lossgen
+from wexchange.torch import dump_torch_weights
+from wexchange.c_export import CWriter, print_vector
+
+def c_export(args, model):
+
+    message = f"Auto generated from checkpoint {os.path.basename(args.checkpoint)}"
+
+    writer = CWriter(os.path.join(args.output_dir, "lossgen_data"), message=message, model_struct_name='LossGen', enable_binary_blob=False, add_typedef=True)
+    writer.header.write(
+f"""
+#include "opus_types.h"
+"""
+        )
+
+    dense_layers = [
+        ('dense_in', "lossgen_dense_in"),
+        ('dense_out', "lossgen_dense_out")
+    ]
+
+
+    for name, export_name in dense_layers:
+        layer = model.get_submodule(name)
+        dump_torch_weights(writer, layer, name=export_name, verbose=True, quantize=False, scale=None)
+
+
+    gru_layers = [
+        ("gru1", "lossgen_gru1"),
+        ("gru2", "lossgen_gru2"),
+    ]
+
+    max_rnn_units = max([dump_torch_weights(writer, model.get_submodule(name), export_name, verbose=True, input_sparse=False, quantize=True, scale=None, recurrent_scale=None)
+                             for name, export_name in gru_layers])
+
+    writer.header.write(
+f"""
+
+#define LOSSGEN_MAX_RNN_UNITS {max_rnn_units}
+
+"""
+        )
+
+    writer.close()
+
+
+if __name__ == "__main__":
+
+    os.makedirs(args.output_dir, exist_ok=True)
+    checkpoint = torch.load(args.checkpoint, map_location='cpu')
+    model = lossgen.LossGen(*checkpoint['model_args'], **checkpoint['model_kwargs'])
+    model.load_state_dict(checkpoint['state_dict'], strict=False)
+    #model = LossGen()
+    #checkpoint = torch.load(args.checkpoint, map_location='cpu')
+    #model.load_state_dict(checkpoint['state_dict'])
+    c_export(args, model)