Open AkiSakurai opened 1 month ago
Hi,
Thanks for filing this bug! I can confirm that problem exists when converting ConvTranspose2d in regards to the bias tensor. If bias=False
as you have above, a garbage bias value (AFAICT) is propagated through the edge model conversion and is part of the inference calculation. If bias=True
when running your script, the models behave as expected (agreement to ~1e-8 in my tests). I will update this bug again when a fix has landed.
Description of the bug:
Actual vs expected behavior:
No response
Any other information you'd like to share?
No response