Why is accuracy lost in a model based on the EfficientNetB0 architecture when saved in H5 format ?
I am doing some experiments training classification models from different architectures for mobile devices (MobileNet, NASNetMobile and EfficientNet) and by storing them in H5 format the only model that loses significantly decreases its accuracy are the models based on EfficientNet.
When entering a model based on EfficientNet and evaluating the accuracy, values in excess of 96% are obtained, but when saving the model in H5 format and testing again with the same dataset, the accuracy is reduced by up to 30% (obtaining an accuracy ranging between 66% and 69%).
I have not observed any difference in models accuracy when I tried testing these three pre-trained models (MobileNet, NASNetMobile and EfficientNet) after saving as .h5 file. I have tested these models using Google Colab and Jupyter notebook in Mac M1 with Python 3.10 and TensorFlow 2.13. (Attaching the replicated gist here for your reference - Colab gist, Jupyter gist).
For further analysis, Please let us know if you are using the same software environment(python version, TF version or system OS) where the model was trained or providing the same kind of dataset or have done same dataset preprocessing before feeding into the loaded model.
Please try again with the above points and let us know if the issue still persists. Thank you.