Reproducing evaluation results

[Michiexb]

Hi! I'm having some trouble with reproducing your evaluation results. I've run your eval bash file for all models with the checkpoint files that I downloaded from the link in the Readme, but the results are very different from the ones mentioned in your paper.

These are the accuracies I get: beta_1: --- I get 25.30 % --- Should be 67.30 % beta_2: --- I get 0.56 % ---- Should be 71.73 % beta_4: --- I get 0.21 % ---- Should be 73.69 % beta_8: --- I get 0.12 % ---- Should be 74.59 % beta_16: -- I get 0.76 % ---- Should be 75.54 % beta_32: -- I get 36.81 % --- Should be 76.18 % beta_inf: -- I get 74.40 % --- Should be 76.27 %

Do you have any idea why this might be? Because I would love to use your INN, but of course would need a better accuracy than what I am getting now.

[ZY123-GOOD]

Hello, I wonder if you know how to solve this problem.

[Michiexb]

I didn't fully reproduce the results yet since my time is currently more important to me than the accuracy of the model, but I did train the model for 10 more epochs starting from the checkpoint file (by setting resume_checkpoint in the config to True), and that did give me 67.55% for the beta_8 model. So probably, the checkpoints that they shared are not the fully trained checkpoints from the paper.

[ZY123-GOOD]

Thank you for your reply. Best wishes.

---Original--- From: @.> Date: Mon, Oct 18, 2021 19:15 PM To: @.>; Cc: @.**@.>; Subject: Re: [RayDeeA/ibinn_imagenet] Reproducing evaluation results (#7)

I didn't fully reproduce the results yet since my time is currently more important to me than the accuracy of the model, but I did train the model for 10 more epochs starting from the checkpoint file (by setting resume_checkpoint in the config to True), and that did give me 67.55% for the beta_8 model. So probably, the checkpoints that they shared are not the fully trained checkpoints from the paper.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[RayDeeA]

Hi, how are evaluating the model?

[Michiexb]

The way that it's mentioned in the readme

[RayDeeA]

Ok, that is strange. I’ll clone and check.

[ZY123-GOOD]

Ok, that is strange. I’ll clone and check.

Thank you! Please check it. I am looking forward to using your INN in our new work.

[ZY123-GOOD]

Ok, that is strange. I’ll clone and check.

Today I also get that beta_1: --- I get 25.30 % --- Should be 67.30 %. Could you please correct it in your busy schedule?

[jenellefeather]

I know this is an old issue and its possible the repo is no longer being maintained, however I am getting the same low performance results for the downloaded checkpoints (36.948% for beta_32).

Is there any update on the discrepancy?

[craymichael]

I am also getting the same issue with the exact same accuracies reported in the issue.

Conda (23.1.0) Environment:

• torch==1.7.1
• FrEIA at tag v0.2
• Python 3.9.18

The command I run to evaluate the model:

python -m ibinn_imagenet.eval.ibinn_imagenet_classifier \
    --model_file_path=checkpoints/beta_2.avg.pt \
    --evaluation=accuracy \
    --data_root_folder_val='/mnt/data/imagenet/' \
    --data_root_folder_train='/mnt/data/imagenet/' \
    --model_n_loss_dims_1d 3072 \
    --data_batch_size 32