Experiments to run

[mavaylon1]

1. Train teacher with KL-Div on current dataset regime |

• [ ] Train Student on same dataset but with weighted cross entropy
• [ ] Train Student on smaller dataset but with weighted cross entropy
• [ ] Train Student on smaller dataset (not seen by teacher) but with weighted cross entropy

2. Train teacher with KL-Div and MSE on current dataset regime |

• [ ] Train Student on same dataset @carbonscott (Could you add what else is to try with your current loss)

3. Explore new datasets to see what to include beyond confident overlaps. Repeat 1 and 2 with any dataset we come up with. It is important to remember that the masks already include information about pixels that are not overlapping but peaks found by peakfinder. These are just marked as not peaks. However, down the line we could look into including index only peaks.

4. Data -> Pattern vs our current approach of Data -> Peaks -> Pattern

[carbonscott]

Here's the link to the experiment tracking table:

https://docs.google.com/spreadsheets/d/1LVxf8516awO4g8NgGefzzc_c72R2HvGCWv5K0J2Kl24/edit?usp=sharing