Design Experiments

[DaeBlicki]

This issue describes important point for design an experiment so that is scientific solid and useful. Therefore, this points should be addressed before any code is written.

• Loss function (MSE, RMSE, L1, L2, ...)
• Num epochs
• Optimizer (Adam?)
• Reproducibility (seed all RNGs : PyTorch, NumPy - multiple measurements)
• Test and validation set split (80/20)
• Mini-batch training (batchsize)
• Do we want to observe the loss curve, convergence discussion?
• Do we considering fine-tuning or study transfer-learning?