Celldino

[coupriec]

Summary

This PR introduces code to reproduce main results described in the papers "Scaling Channel-Adaptive Self-Supervised Learning", published in TMLR a few months ago https://openreview.net/forum?id=pT8sgtRVAf , and to the paper "Cell-DINO: Self-Supervised Image-based Embeddings for Cell Fluorescent Microscopy" , in revision.
The code contains a main README_CELL-DINO_AND_CHANNEL-DINO.md, that points to two separate readme corresponding to the two papers.

We list the changes in 3 categories:

1- New non .py files or documentation: License, readmes, scripts to launch code to reproduce some results, and configuration files.

LICENSE_CELLDINO
LICENSE_CELLDINO_WEIGHTS
README.md
README_CELL-DINO_AND_CHANNEL-DINO.md

docs/Cell-DINO.png
docs/README_CELLDINO.md
docs/README_CHANNEL_ADAPTIVE_DINO.md
docs/launcher_CHAMMI_eval.sh
docs/launcher_knn_eval_on_chammi.sh
docs/test_inference_celldino.py

dinov2/configs/eval/celldino_hpaone.yaml
dinov2/configs/eval/channeldino_ext_chammi.yaml
dinov2/configs/train/hpafov_vitl16.yaml
dinov2/configs/train/hpafov_vitl16_boc.yaml
dinov2/configs/train/hpaone_vitl16.yaml

2- New .py files: data and evaluation files

These new files contain data or checkpointing helpers (dinov2/data/[accumulators.py), dinov2/utils/checkpoint.py : necessary function to the linear_celldino eval), specific transforms used for pretraining (cell_augmentations.py), for evaluations (dinov2/data/transforms_cells.py ), new data loaders (dinov2/data/datasets/chammi_cp.py dinov2/data/datasets/chammi_hpa.py
dinov2/data/datasets/chammi_wtc.py dinov2/data/datasets/hpafov.py dinov2/data/datasets/hpaone.py)

Specific knn and linear evaluation files have been created with new options:

dinov2/eval/utils_celldino.py
dinov2/eval/knn_celldino.py
dinov2/eval/linear_celldino.py
dinov2/run/eval/knn_celldino.py
dinov2/run/eval/linear_celldino.py

3- Modified files

dinov2/configs/ssl_default_config.yaml

dinov2/data/init.py : added necessary imports
dinov2/data/adapters.py : implemented padding in DatasetWithEnumeratedTargets
dinov2/data/datasets/init.py: added necessary imports
dinov2/data/datasets/decoders.py : introduced new decoders necessary to load multi-channel images
dinov2/data/datasets/extended.py : introduced options to use different decoders
dinov2/data/loaders.py : added necessary imports to use the new datasets

dinov2/eval/metrics.py : added multi-label and multi-class F1 score metrics

dinov2/hub/backbones.py: introduced Celldino models

dinov2/models/init.py: propagate the in_chans (number of channels) and channel_dataptive parameters in build_model
dinov2/models/vision_transformer.py: implemented the channel adaptive option

dinov2/train/train.py: picking the correct pretraining augmentation from the config option train.cell_augmentation
dinov2/utils/cluster.py : useful modification to test this PR.

[jccaicedo]

Can we change this for the new paper title, please?
"Cell-DINO: Self-Supervised Image-based Embeddings for Cell Fluorescent Microscopy"

[jccaicedo]

Please, change the title here for:
"Cell-DINO: Self-Supervised Image-based Embeddings for Cell Fluorescent Microscopy"
Change Unbiased for Self-Supervised.

Thank you!