# Weekly Progress (1/12 - 1/16)

## Hyperparameter Tuning for Filopodia Data

This week, the main focus was on hyperparameter tuning for the IsoScope model using filopodia microscopy data.

### GitHub Repositories

Training code:
- https://github.com/CharlieC30/IsoScopeXX
- (forked from changlabntu/IsoScopeXX)

Testing/Inference code:
- https://github.com/CharlieC30/TestingMicroscopy
- (forked from changlabntu/TestingMicroscopy)

### Experiments

Three main parameters were tuned:

| Parameter | Tested Values |
|-----------|---------------|
| lamb | 1, 3, 4, 5, 6, 7, 8, 9, 10 |
| skipl1 | 4, 8 |
| l1how | max, dsp |

### Results


- The best parameters so far are: lamb=5, skipl1=4, l1how=max.
- lamb controls how much the output shrinks. Too small values cause more background noise or empty outputs; too large values also cause more background noise. lamb=5 provides a good balance.
- skipl1=4 produces better results than skipl1=8. With skipl1=8, extreme parameter values lead to empty outputs.
- The best choice between l1how=max and l1how=dsp depends on the data and needs to be tested.