How to define a new convolution layer?

[lidehui]

We can use torch.nn.Conv2d to create an normal convolution layer, but how to create a convolution layer with a kernel of novel shape, such as ‘T’ shape(means with kernel weight of [w1 w2 w3; 0 w4 0; 0 w5 0] )?

One way to achieve this layer is using torch.nn.Conv2d to define a 3x3 normal convolution layer firstly (named NormalLayer), and then set the corresponding position as zero in NormalLayer.weight.data before every time I use NormalLayer. But the calculated amount will equal to 3x3 normal convolution (9 points) in this way, while the true calculated amount is 5 points (w1 to w5) in ‘T’ shape kernel. Apparently, this solution is not what I want.

Other solution maybe defines a new convolution layer. Just as a 3x3 normal convolution with dilation of 2. Its kernel size is 5x5, but have calculated amount of 9 (3x3=9) points only rather than 25 (5x5=25). So, how to define a new convolution layer to define ‘T’ shape kernel with calculated amount of 5 points? Maybe use the Extending in PyTorch?

Could somebody help me? Thanks very much!

[InnovArul (Arul)]

@lidehui

One way to achieve this layer is using torch.nn.Conv2d to define a 3x3 normal convolution layer firstly (named NormalLayer), and then set the corresponding position as zero in NormalLayer.weight.data before every time I use NormalLayer. But the calculated amount will equal to 3x3 normal convolution (9 points) in this way, while the true calculated amount is 5 points (w1 to w5) in ‘T’ shape kernel. Apparently, this solution is not what I want.

Why do you think the calculated result is of 3x3 normal convolution? Since you are setting non ‘T’ elements to 0, don’t you think the convolution only calculates 5 multiplications effectively ?

[lidehui]

What I mention about is calculated amount. Since I using 3x3 normal convolution, the calculation amount of single kernel is 9 (3x3) absolutely, nothing but some values of the kernel are set as zero artificially.