[tmva][sofie] Account for dilation in Conv2D same padding calculation

[harz05]

Changes

The padding='same' calculation in the Keras Conv2D parser was using the raw kernel size instead of the effective kernel size, which for dilated convolutions is dilation * (kernel - 1) + 1. fAttrDilations was extracted from the layer attributes but never factored into the padding formula.

# before
padding_height = max((outputHeight - 1) * strides[0] + kernel_shape[0] - inputHeight, 0)

# after
effective_kH = dilations[0] * (kernel_shape[0] - 1) + 1
padding_height = max((outputHeight - 1) * strides[0] + effective_kH - inputHeight, 0)

For dilation=2, kernel=3, input=8, stride=1 the old formula produced padding=2 while the correct value is 4, causing the generated inference code to operate on tensors with wrong spatial dimensions. With dilation=1 the effective kernel size reduces to the raw kernel size, so existing models are unaffected.

Checklist

• tested changes locally

This PR fixes #21684

[sanjibansg]

Hi @harz05,
Thanks for this PR. Can you add a test so that we can verify this patch?

[github-actions]

Test Results

    22 files      22 suites   3d 7h 32m 39s ⏱️
 3 834 tests  3 833 ✅  1 💤 0 ❌
75 676 runs  75 658 ✅ 18 💤 0 ❌

Results for commit 345e072.

♻️ This comment has been updated with latest results.

[harz05]

Hello @sanjibansg, I tried adding a numerical test with Conv2D(kernel=3, padding='same', dilation_rate=2) on 8×8 input and on testing locally I found that the output shape was correct (8x8x4=256 vs 144 before the fix), confirming the padding calculation works.
However the numerical values don't match Keras. The reason I could figure out is that ROperator_Conv::ShapeInference in ROperator_Conv.hxx overwrites fAttrKernelShape with the dilated kernel size and then Im2col applies dilation again. This is a seperate double dilation bug and a full numerical test might fail due to this and not due to the padding fix.
Should I raise a follow up issue first?

[sanjibansg]

Hello @sanjibansg, I tried adding a numerical test with Conv2D(kernel=3, padding='same', dilation_rate=2) on 8×8 input and on testing locally I found that the output shape was correct (8x8x4=256 vs 144 before the fix), confirming the padding calculation works. However the numerical values don't match Keras. The reason I could figure out is that ROperator_Conv::ShapeInference in ROperator_Conv.hxx overwrites fAttrKernelShape with the dilated kernel size and then Im2col applies dilation again. This is a seperate double dilation bug and a full numerical test might fail due to this and not due to the padding fix. Should I raise a follow up issue first?

Hi @harz05,
Could that be fixed within this PR itself, that should be fine and we can expand its scope.

[harz05]

hello @sanjibansg
i've changed Im2col calls in ROperator_Conv.hxx to pass dilation as 1, also added a dilated Conv2d test and corresponding model generator. Pls review it once.
Thank you.

[harz05]

There was a clang and ruff formatting issue during the run, I fixed that. CI was also getting Windows x86 failure but i'm not really sure what's the issue in that.
pls review

[harz05]

hello @sanjibansg ,sorry for the noisy commit history, I was trying to fix the ruff formatting issue and accidentally did a rebase that pulled in a ton of upstream master commits into the branch. I've force pushed a clean branch with just the 4 PR commits now
ruff format is fixed as flagged by CI
all other tests were passing according to the CI run