Considering resource availability, I have taken a subset of the provided data. I have considered 7 categories from the provided data. Here are the following classes I have chosen: capsule_crack, cadpluse_faulty_imprint, capsule_good, hazelnut_crack, hazelnut_cut, hazelnut_good, hazelnut_hole.
I have used VGG16 to classify the data. Unfortunately, I spend very little time on the problem. I only got a chance to train for 2 epochs and my accuracy was 37% on test data. The size of my saved model is greater than 180MB and I couldn't add it to github.
The inference.py scripts takes the model and test image and generates the prediction.