Hi!

I have a model trained by using the current 0.1.0 version of this library.

Later on, I loaded the model checkpoint in the new environment that also uses the 0.1.0 version. Model performance downgraded substantially.

The reason is that your requirements.txt mentions efficientnet-pytorch>=0.6.3. My old training environment used the 0.6.3 version. However, the current one from Aug 25 is 0.7.0. So, when I specified 0.1.0 as a requirement, it wasn't the same thing.

0.1.0 with 0.6.3 is a breaking difference when compared to 0.1.0 and 0.7.0.

Maybe it should be efficientnet-pytorch==0.6.3 if still staying at 0.1.0?