Deep4Life project: Pigeons team

All the results visualisation are in the Presentation.

Trained models with results are in Best results.

Installation

We recommend creating venv or conda environment with python>=3.9. You can use environment.yaml.

Conda and requirements.txt

conda create -n stellar python=3.9
source activate stellar

And then:

pip3 install -r requirements.txt

Overview

When you want to run any experiment, run:

cd src

and then

python3 train_and_validate.py [ARGUMENTS]

with possible options:

• -h, --help: show this help message and exit
• --dataset-path (default="data/train"): dataset path
• --method {stellar,torch_mlp,sklearn_mlp,xgboost} (default="stellar"):
• --config (default="standard"): Name of a configuration in src/config/{method} directory.
• --cv-seed (default=42): Seed used to make k folds for cross validation.
• --n-folds (default=5): Number of folds in cross validation.
• --retrain (default=True): Retrain a model using the whole dataset.

We recommend using --config flag. Sample configs are given in src/config/{method} folders.

Notebooks

All the notebooks are in folder notebooks.

For each notebook run git config --local core.hooksPath .githooks/ to set up git hooks for the project.

EDA

EDA is in file notebooks/exploratory-data-analysis.ipynb.

Baselines

We have 3 (4) baselines in total:

SVM

XGBoost

Model is available in src/models/xgboost.py. Experiments were ran using configs from src/config/xgboost/standard.yaml and notebook notebook/xgboost_tryout.ipynb.

MLP

There are two available models: one using sklearn, and one using torch. Models are implemented in src/models/{torch, sklearn}_mlp.py. Experiments were ran using configs from src/config/{torch, sklearn}_mlp/standard.yaml and notebooks notebook/MLP_tryout.ipynb.

CellSighter

Stellar

Originally STELLAR was developed by Snap Stanford.

PyTorch implementation of STELLAR, a geometric deep learning tool for cell-type discovery and identification in spatially resolved single-cell datasets. STELLAR takes as input annotated reference spatial single-cell dataset in which cells are assigned to their cell types, and unannotated spatial dataset in which cell types are unknown. STELLAR then generates annotations for the unannotated dataset. For a detailed description of the algorithm, please see our manuscript Annotation of Spatially Resolved Single-cell Data with STELLAR.