replication

Todos

Add the things you already did here too so we can keep track of them.

So far done:

• Dataset class in Pytorch: some preprocessing steps and normalization but we need interpolation and padding
• NSN network: correct architecture and training loop but we need to train it and save the weights
• NDN network: correct architecture and training loop but it shrinks the depth too much at deeper layers (will be fixed with padding)
• Metrics: implemented but not tested

To do:

• BiCubic Interpolation
• Mirror padding
• Data augmentation
• (Aral; done but not tested) All the metrics described in metrics.py
• Metrics fixed and unit tested
• Water-based marker shed for instance segmentation
• (Dani) Visualizing the training and test error per epoch
• Visualizing inference on cells
• Quantitative observations about the inferred cells
• Training early stopping criterion: "The trained model with the highest IoU in cross-validation was used to analyse the test dataset" (p. 2).

TA Meeting

Deliverables

• Blog post indicating approach, difficulties in reproducing, comparison with paper results; and explain differences.

• Poster presentation

• Contact supervisor about data format; need a few labelled ground truths as well probably.

Planning

Four Criteria:

• Replicate: A full implementation from scratch without using any pre-existing code? (does looking at an existing implementation count?)
• New Code Variant: Rewrote or ported existing code to be more efficient/readable. (does reimplementing in pytorch count?)
• New Data: Evaluate different datasets to obtain similar results. (part of the assignment: replicate for new cell polarity data)

Pick one

• Reproduce?: Existing code was evaluated. (is it fine to just check the code quality, or do we have to find the original model weights and run the entire thing?)
• Hyperparams Check: Evaluating sensitivity to hyperparameters.

Probably Not

• New algorithm variant: Evaluating a slightly different variant (only Dani seems to know alternatives)
• Ablation study: Additional ablation studies. (both NDN and NSN are dependent on each other, so not sure how to proceed.)