pnnx convert onnx depth2space, add pixelshuffle pixelunshuffle test #5967
Conversation
There was a problem hiding this comment.
Pull Request Overview
This pull request adds new tests for pixel shuffle and pixel unshuffle operations on ONNX conversion using pnnx. The changes introduce unit tests for both the nn module and functional APIs related to pixel shuffle/unshuffle operations, ensuring conversion correctness.
Reviewed Changes
Copilot reviewed 4 out of 6 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
| tools/pnnx/tests/onnx/test_nn_PixelUnshuffle.py | Added test for nn.PixelUnshuffle conversion |
| tools/pnnx/tests/onnx/test_nn_PixelShuffle.py | Added test for nn.PixelShuffle conversion |
| tools/pnnx/tests/onnx/test_F_pixel_unshuffle.py | Added test for F.pixel_unshuffle conversion |
| tools/pnnx/tests/onnx/test_F_pixel_shuffle.py | Added test for F.pixel_shuffle conversion |
Files not reviewed (2)
- tools/pnnx/src/pass_level2/F_pixel_shuffle.cpp: Language not supported
- tools/pnnx/tests/onnx/CMakeLists.txt: Language not supported
| return torch.equal(a0, b0) and torch.equal(a1, b1) | ||
|
|
There was a problem hiding this comment.
Using torch.equal for comparing floating-point tensors may lead to false negatives due to minor numerical differences. Consider switching to torch.allclose with appropriate tolerance.
| return torch.equal(a0, b0) and torch.equal(a1, b1) | |
| return torch.allclose(a0, b0) and torch.allclose(a1, b1) |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
| return torch.equal(a0, b0) and torch.equal(a1, b1) | ||
|
|
There was a problem hiding this comment.
Using torch.equal for comparing floating-point tensors can be too strict after ONNX conversion. It may be more robust to use torch.allclose with a tolerance to account for potential numerical variability.
| return torch.equal(a0, b0) and torch.equal(a1, b1) | |
| return torch.allclose(a0, b0, atol=1e-6) and torch.allclose(a1, b1, atol=1e-6) |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
| return torch.equal(a, b) | ||
|
|
There was a problem hiding this comment.
Comparing tensors with torch.equal may not accommodate minor numerical discrepancies post-conversion. Consider using torch.allclose with a defined tolerance for a more reliable test.
| return torch.equal(a, b) | |
| return torch.allclose(a, b) |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
| return torch.equal(a, b) | ||
|
|
There was a problem hiding this comment.
Using torch.equal for floating-point tensor comparison might be overly strict. It is advisable to use torch.allclose with an appropriate tolerance to ensure robust test verification.
| return torch.equal(a, b) | |
| return torch.allclose(a, b, atol=1e-08) |
Copilot is powered by AI, so mistakes are possible. Review output carefully before use.
No description provided.