Great work! Since you shared the conversion script, I'd suggest adding metrics and equivalence tests to compare output drift between the Core ML models and PyTorch eager mode. This would help quantify numerical differences introduced during conversion and catch regressions early. I have been looking at this but only at module level (https://opdiff.com/). A key takeaway is that just because the Core ML model runs doesn’t mean it works as expected.
Great work! Since you shared the conversion script, I'd suggest adding metrics and equivalence tests to compare output drift between the Core ML models and PyTorch eager mode. This would help quantify numerical differences introduced during conversion and catch regressions early. I have been looking at this but only at module level (https://opdiff.com/). A key takeaway is that just because the Core ML model runs doesn’t mean it works as expected.