Hi all,
I just started using pytorch. I wonder how could I define a layer that combine convolution with deconvolution. More precisely, for a tensor whose shape is [I_T, I_H, I_W], what I want is do convolution at the second and third dimension, and do deconvolution at the first dimension, so the shape of the output will be [(I_T - 1) * S_T + K_T - 2 * p_T, (I_H + 2 * P_H - K_H) / S_H + 1, (I_W + 2 * P_W - K_W) / S_W + 1].(T、H、W represents three dimension respectively, I means input, s means stride, k means kernel size, p means padding).
Meanwhile, I know I can get the output of the same shape as I want by using convolution first, and then use deconvolution, but I want to do both them at the same time, so I have to define a new layer.
Any suggestions? If demo code is too complex, could you give me some idea? Any help would be appreciated.