diff --git a/src/experiments/NOTES.md b/src/experiments/NOTES.md index b389656..6e1b845 100644 --- a/src/experiments/NOTES.md +++ b/src/experiments/NOTES.md @@ -467,4 +467,9 @@ Maybe >30 PSNR is ok? Would be interesting to compare the order of the logits -- ### May 10, 2023 Tried a few other measures of similarity (KL Div, Jaccard of the top k) and they all agree with PSNR. Jaccard is a little easier for me to intuit -- 90PSNR is 1 jaccard, 60PSNR is ~.95 jaccard, 35PSNR is ~0.84, 20PSNR is ~0.40. Think that's good enough to be usable (anecdotally it is, but this gives me a bit more confidence). Still feels odd to me that f16 would contribute so much error -- once the torch mps f16 issues are fixed would be interesting to compare and rule out CoreML oddity. -Turns out that is fixed in the nightly torch release. PSNR is 42 which is a smidge higher than CoreML. Possible I missed some float32 somewhere. The torch amp docs have a section for CUDA ops that are numerically unstable in float16, might be an interesting cross-reference. \ No newline at end of file +Turns out that is fixed in the nightly torch release. PSNR is 42 which is a smidge higher than CoreML. Possible I missed some float32 somewhere. The torch amp docs have a section for CUDA ops that are numerically unstable in float16, might be an interesting cross-reference. + +### May 29, 2023 +TIL zip has a max size of 4GB and macOS doesn't support zip64. New commands to compress + decompress large files: +compress: `tar -czvf - model.mlpackage | split -b 1800m -d -a 1 - model.mlpackage.tar.gz.` +decompress: `cat directory.tar.gz.* | tar -xzvf -` \ No newline at end of file