Incorrect ConvTranspose2D padding behavior (forward padding vs. output crop) #420
Description
What
ConvTranspose2d padding was treated as input padding, which changes numerical results compared to PyTorch.
For transposed convolutions, PyTorch's padding parameter does not represent input padding — instead it represents a symmetric output crop.
Prior to this issue, the conversion pipeline inserted a Pad node before the transposed convolution when padding != 0. However, in a transposed convolution, adding input padding modifies the upsampling phase and kernel overlap pattern, which results in different values, even if the output shape matches.
This issue corrects the behavior by:
- Never inserting input padding for transposed convolutions
- Mapping the builtin op with VALID padding
- Applying a Slice (NHWC) crop after the transposed convolution when needed
As a result, the converted model now produces values consistent with the original PyTorch model.