Model Prototype

[RyanCaudill04]

Improve Model Accuracy from baseline of 67%

• Take in image
• Process
• Guess
• Maybe confidence

[theodv]

1.1 Model has metrics at 15% epochs for base and finetune training:

Accuracy: 66.5% AUC: 0.923 Loss: 0.974.

While the accuracy number leaves more to be desired, it is actually the very high AUC that needs to be looked at. 92.3% of the time, the model will rank a randomly chosen positive case higher (assign it a better score) than a randomly chosen negative case.

This imbalance of the numbers mostly comes from the classes within the data itself. The HAM10000 is a fairly imbalanced class dataset, leaving the model to struggle with underrepresented classes.

Next step: Run the same code but with EfficientNetB0 to see how it compares. Then play around with number of epochs and quantizing.