results are not reproducible
#1
by
mattiaGCom
- opened
Hi, I'm trying to reproduce the results you're mentioning in the model description (such as 77% accuracy from step 100).
I started from the Google ViT backbone (as you suggested) and the loaded the datasets as you did in your notebook available on Kaggle.
However, results are not the same, as my initial accuracy is 28% (at step 100 in the first epoch).
Maybe we can get your numbers if we start the training with your already fine-tuned model, instead of using the Google ViT.
Any thoughts/suggestions?
Hey,
Yes, exactly. If you check the notebook, you'll see that I used my fine-tuned model, not the ViT.
I hope this helps with your task.