Cellseg-models-pytorch

Features

- Image encoders are imported now only from timm models.
- Add `enc_out_indices` to model classes, to enable selecting which layers to use as the encoder outputs.

Removed
- Removed SAM and DINOv2 original implementation image-encoders from this repo. These can be found from timm models these days.
- Removed `cellseg_models_pytorch.training` module which was left unused after example notebooks were updated.

Examples

- Updated example notebooks.
- Added new example notebooks utilizing UNI foundation model from the MahmoodLab.
- Added new example notebooks utilizing the Prov-GigaPath foundation model from the Microsoft Research.
- **NOTE:** These examples use the huggingface model hub to load the weights. Permission to use the model weights is required to run these examples.

Chore

- Update timm version to above 1.0.0.

Breaking changes

- Lose support for python 3.9
- The `self.encoder` in each model is new, thus, models with trained weights from previous versions of the package will not work with this version.

<a id='changelog-0.1.24'></a>

Style

- Update the `Ìnferer.infer()` -method api to accept arguments related to saving the model outputs.

Features

- Add `CPP-Net`. https://arxiv.org/abs/2102.06867

- Add option for mixed precision inference

- Add option to interpolate model outputs to a given size to all of the segmentation models.

- Add DINOv2 Backbone

- Add support for `.geojson`, `.feather`, `.parquet` file formats when running inference.

Docs

- Add `CPP-Net` example trainng with Pannuke dataset.

Fixes

- Fix resize transformation bug.

<a id='changelog-0.1.23'></a>

Features

- add a stem-skip module. (Long skip for the input image resolution feature map)

- add UnetTR transformer encoder wrapper class
- add a new Encoder wrapper for timm and unetTR based encoders

- Add stem skip support and upsampling block options to all current model architectures

- Add masking option to all the criterions
- Add `MAELoss`
- Add `BCELoss`

- Add base class for transformer based backbones
- Add SAM-VitDet image encoder with support to load pre-trained SAM weights

- Add `CellVIT-SAM` model.

Docs

- Add notebook example on training Hover-Net with lightning from scratch.

- Add notebook example on training StarDist with lightning from scratch.
- Add notebook example on training CellPose with accelerate from scratch.
- Add notebook example on training OmniPose with accelerate from scratch.

- Add notebook example on finetuning CellVIT-SAM with accelerate.

Fixes

- Fix current TimmEncoder to store feature info

- Fix Up block to support transconv and bilinear upsampling and fix data flow issues.

- Fix StardistUnet class to output all the decoder features.

- Fix Decoder, DecoderStage and long-skip modules to work with up scale factors instead of output dimensions.

<a id='changelog-0.1.22'></a>

Features

- Add mps (Mac) support for inference
- Add cell class probabilities to saved geojson files

<a id='changelog-0.1.21'></a>

Features

- Add StrongAugment data augmentation to data-loading pipeline: https://arxiv.org/abs/2206.15274

Fixes

- Minor bug fixes
<a id='changelog-0.1.20'></a>

Fixes

- Enable writing folder & hdf5 datasets with only images (previously needed image-mask pairs)
- Enable writing datasets without patching.

- Add long missing h5 reading utility function to `FileHandler`

Features

- Add hdf5 input file reading to `Inferer` classes.

- Add option to write pannuke dataset to h5 db in `PannukeDataModule` and `LizardDataModule`.

- Add a generic model builder function `get_model` to `models.__init__.py`

- Rewrite segmentation benchmarker. Now it can take in hdf5 datasets.

<a id='changelog-0.1.19'></a>