Ziwen Liu

BERT-style layer-wise LR decay following ConvNext v2 during fine-tuning.

Test on python 3.12

1

PyTorch 2.2.1 added python 3.12 support.

Shows how to subclass the trainer with a placeholder method so the CLI parser can be reused. Adapted from https://github.com/Lightning-AI/lightning/issues/12302#issuecomment-1110425635. What we actually want is probably https://github.com/Lightning-AI/lightning/pull/18105.

The training loop lacks automated tests. Although unit-testing DL code is not as straightforward as other software, there are some strategies to improve the coverage. See these blog posts for...

Augmentation strategy for generalization across magnification

1

Different magnification of the microscope alters the sampling of all 3 spatial dimensions. And the changes in Z is different from that in XY. If we want to train a...

Use `ModuleList` instead of the repetitive [custom method](https://github.com/mehta-lab/microDL/blob/1e936662dc708487692d2036d6c050e307cea20f/micro_dl/torch_unet/networks/layers/ConvBlock3D.py#L328-L339). Originally posted in https://github.com/mehta-lab/microDL/issues/214#issuecomment-1548150343

Add far red color code

2

Currently the `CHANNEL_COLORS` look up table used for OME-Zarr metadata does not contain far red, e.g. Y5 (~700 nm emission). What will be a good color for those? Currently ~600...

Pydantic 2.0

2

Pydantic released 2.0 and has many breaking changes. Since other software (e.g. napari) does not work with the new version yet, the upgrade will need to be postponed. For now...