Accuracy not Improving as in Paper Results

[meher-mankikar]

Hello,
I cloned the repo and am running the training file, but my accuracy seems to be improving much slower than the results presented in the paper. Are there some hyperparameter settings that should be changed to reproduce your results? I have attached some partial training results I am getting right now.

Thank you for your help!

[Horea94]

Did you download the repo for tensorflow 1.8 or the latest one?
Bear in mind that the results in the paper were recorded with certain hardware and a much older version of tensorflow. It is difficult to estimate how the newer versions affect the overall training process and you may have to perform some exploratory search in order to find the best hyper parameters.

[meher-mankikar]

I was using the latest version for the results above. If using tf 1.8, were you using the code at image_classification_tf_1.8.0/network/fruit_train_net.py for training?
I was originally running image_classification/Fruits-360 CNN.py but I believe this is only compatible with tf 2.0?

[Horea94]

Yes. However, that version is deprecated and I would still recommend using the one for tf 2.0 as it should achieve comparable performance even if the training progress differs.