Unable to Reproduce cfar10 experiment results
Lewington-pitsos opened this issue · comments
lewington commented
I cloned the repo and installed requirements on the tensorflow-gpu docker image.
I made no changes other than to add a new dataset.
I ran the command:
python3 -m sam.sam_jax.train --dataset cifar10 --model_name WideResnet28x10
--output_dir /tmp/my_experiment --image_level_augmentations autoaugment
(1stpass folder)
And then re-ran it again just to be sure, but the error and loss scores I am getting are orders of magnitude worse than what was reported in the paper.
Possibly I've made some error, but I can't see what it could possibly be.
You can check out my results here: https://tensorboard.dev/experiment/vhlr4NRkSqSevdfqIVvIjg/#scalars
Devansh Bisla commented
@Lewington-pitsos were you able to fix these issues. I'm struggling with similar problem #6
lucasliunju commented
Could I ask how to solve it?
Thank you!