Method forward

stream/modules/convolution.py:232–262

x: [bs,C,1,F] cache: [bs,C,T-1,F]

(self, x, cache)

                                        bias = bias)
        
    def forward(self, x, cache):
        """
        x: [bs,C,1,F]
        cache: [bs,C,T-1,F]
        """
        # [bs,C,T,F]
        inp = torch.cat([cache, x], dim = 2)
        out_cache = inp[:, :, 1:]
        bs, C, T, F = inp.shape
        
        # Upsampling operation
        if self.F_stride > 1: 
            # [bs,C,T,F] -> [bs,C,T,F,1] -> [bs,C,T,F,F_stride] -> [bs,C,T,F_out]
            inp = torch.cat([inp[:,:,:,:,None], torch.zeros([bs,C,T,F,self.F_stride-1])], dim = -1).reshape([bs,C,T,-1])
            left_pad = self.F_stride - 1
            if self.F_size > 1:
                if left_pad <= self.F_size - 1:
                    inp = torch.nn.functional.pad(inp, pad = [(self.F_size - 1)*self.F_dilation-self.F_pad, (self.F_size - 1)*self.F_dilation-self.F_pad - left_pad, 0, 0])
                else:
                    # inp = torch.nn.functional.pad(inp, pad = [self.F_size - 1, 0, 0, 0])[:,:,:,: - (left_pad - self.F_stride + 1)]
                    raise(NotImplementedError)
            else:
                # inp = inp[:,:,:,:-left_pad]
                raise(NotImplementedError)

        else: # F_stride = 1
            inp = torch.nn.functional.pad(inp, pad=[(self.F_size-1)*self.F_dilation-self.F_pad, (self.F_size-1)*self.F_dilation-self.F_pad])
                
        outp = self.ConvTranspose2d(inp)
    
        return outp, out_cache
    

if __name__ == '__main__':