DepthWiseConv2dImplicitGEMM has no 'padding' class attribute(actually zero)

[zzzc18]

When I ran directly replknet.py with using DepthWiseConv2dImplicitGEMM, I got this error:

RuntimeError: The size of tensor a (56) must match the size of tensor b (26) at non-singleton dimension 3

which seems likely a padding issuse with large kernel size. Then I checked replknet.py and depthwise_conv2d_implicit_gemm.py and found there is no padding parameter.

With debugging, I find out that DepthWiseConv2dImplicitGEMM has no attribute of padding(actually zero), which leads to when calling model.structural_reparam() and ReparamLargeKernelConv.merge_kernel(), incorrect conv2d class will be created (miss the using condtion of DepthWiseConv2dImplicitGEMM and fallback to create a nn.Conv2d with 0 padding).

[StarTherlyn]

I got the same problem, have you solved it yet?

[StarTherlyn]

I got the same problem, have you solved it yet?
I add this at line56：
result.conv.padding = padding

[zzzc18]

I got the same problem, have you solved it yet?
I add this at line56：
result.conv.padding = padding

I didn't investigate it further, my exp finds author's Conv2d fails to converge on my mission, while nn.Conv2d is fine, so I just ignore them.

[jsczzzk]

I also got the same problem. Have you solved it yet? @StarTherlyn

[jsczzzk]

I changed this at line122：

and got the results:
.

Is it right?

[zzzc18]

I changed this at line122： and got the results: .

Is it right?

I think the difference is too large, I recommend you to use nn.Conv2d Instead

[jsczzzk]

Thank you for your suggestion ：）

[liyiersan]

I changed this at line122： and got the results: .

Is it right?

I guess it is right. As can be seen in depthwise_conv2d_implicit_gemm.py in cutlass, nn.Conv2d needs padding while DepthWiseConv2dImplicitGEMM does not.

I modified line 122 as you did, and the difference is quite small, just 1e-5.

The difference is accepted to me.

[liyiersan]

I changed this at line122： and got the results: .

Is it right?

Actually, another modification to convert tuple to int is needed. see #59