So I was trying to create a classifier using the newly released dataset Imagenette. I was able to achieve an accuracy of 89% on the 160px dataset by training it for approx 25 epochs. I think I can improve the accuracy by training for more epochs but I have not tried it. Code can be found here. I have not used a pre-trained model as using pre-trained model will be cheating.
A couple of questions.

  • The labels naming does not seem right (‘n01440764’, ‘n02102040’). Is there a way to improve them by using human-readable labels.

  • Validation loss shoots up to values like 39277.097656 which does not seem right to me. Not sure what is causing this. Am I doing something wrong?


Also if you can point to some other improvements that also will be useful.

Tried another architectures like Adding AdaptiveAvgPooling layer before flatten and adding a Fully Connected layer after the Flatten layer but I was not able to breach 89% accuracy mark. If any one has tried this dataset please let me know your approach

Updated code: