keras-io
keras-io copied to clipboard
Add adamatch HF space and model
A little import fix and I added Hugging Face model and space links.
There is a bit of a problem with reproducing the original result by Sayak. With the settings as they are, so far the best test accuracy on the target set I could get was about 14%, which is still lower than Sayak's 19%. This shows in the generated md file. I did about 10 runs by now. What is your suggestion in this case @fchollet? Should I keep on trying to get close to 19%, leave it at 14% or is there another, better solution?
@sayakpaul Hi Sayak, do you have any suggestion on how to reproduce the original results?
Btw it may be beneficial to use the RandAugment version from KerasCV now that it's available.
Btw it may be beneficial to use the RandAugment version from KerasCV now that it's available.
Sure thing. I will try submitting a PR with that. Cc: @LukeWood
@sayakpaul Hi Sayak, do you have any suggestion on how to reproduce the original results?
I suspect this to be a version mismatch problem. But @johko mentions that even after 10 runs the performance stays there. Is it the same across different seeds?
Sorry for the late reply, I was out of office for a bit. I'll run some more experiments across different seeds, as @sayakpaul suggested and will report back on how it worked out.
@fchollet I finally managed to get a good run and added the generated files
Awesome. Thanks @johko! Appreciate your willingness.
cla/google Failing after 1s — ❌ Missing CLA from one or more contributors
Please sign the CLA.
@fchollet this one should be good now as well :)