Feature Map:
Height: (input_height - kernel_size + 2*padding )/stride[0] + 1
Width: (input_width - kernel_size + 2*padding )/stride[1] + 1
反卷积：
ConvTranspose2d(in_channel, out_channel, kernel_size, stride,
padding, output_padding, groups=1, bias=True, dilation=1)
Height_out = (Height_in - 1) * stride[0] - 2 * padding[0] + kernel_size[0] + output_padding[0]
Width_out = (Width_in - 1) * stride[1] - 2 * padding[1] + kernel_size[1] + output_padding[1]