americast / algo_bias Goto Github PK

It would be nice to experiment with other loss functions and alternate network architectures. Also, results in the form of an ablation study in the README would be great.

In order to improve code reproducibility, it would be great to have a fixed random seed for all the experiments. Experimentations may also be done using different seeds, however, they need to be updated in the README accordingly.

Improve the README describing the flow in kmain.py, and how LRP is being used to obtain the contributions from the inputs at every layer.

The heatmap generated is static and is based on the final trained model. It would be great to have different heatmaps based on the current iteration of training. A gif as a collection of all such heatmaps on the README would help in better analysis.

The heatmaps are generated using keras variant kmain.py. It would be nice to have:

• More modularity in the code (by adding more functions)
• Docstrings for every module.

Dropping of unnecessary Scale_ID takes place while training. This can be done while preprocessing to speed up training.

• Distribute train and test data 🙄
• Add tanh activation 🎃
• Change layer sizes 🤕