This is 3rd place solution in challenge Open Problems – Single-Cell Perturbations. The stage 1 doesn't use the pseudolabels, but the stage 2 does. The stage simple is related to fast, single model. The outputs are placed in submissions directory.
- A detailed solution https://www.kaggle.com/competitions/open-problems-single-cell-perturbations/discussion/458750.
- A required data https://www.kaggle.com/competitions/open-problems-single-cell-perturbations/data.
- de_train.parquet - a training data
- id_map.csv - a test data
- sample_submission.csv - a sample of the submission in kaggle challenge
Project is created with:
- python 3.9
- tensorflow 2.12.0
- numpy 1.23.5
- pandas 2.0.3
- scikit-learn 1.3.2
Ubuntu 18.04.6 LTS (256 GB boot disk) Intel(R) Core(TM) i5-9300H CPU @ 2.40GHz (8 cores) 8GB RAM
To train simple model
python -m prepare_data --stage simple && python -m train --stage simple
This model (model_7_0.keras) is serialized and placed in models/simple/model_7, so training can be omitted.
To train stage 1 models
python -m prepare_data --stage stage_1 && python -m train --stage stage_1
To train stage 2 models
python -m repro --stage stage_1 && python -m prepare_data --stage stage_2 && python -m train --stage stage_2
In the stage 2 the id_map.csv has to be replaced by a new test set, due to creating the pseudolabels from the stage 1.
Replace id_map.csv with new test set. To predict execute
python -m predict --stage {one of simple, stage_1, stage_2}
Clear all directories with models, except of models/simple directory, which contains a serialized model.
- Download from challenge website and copy de_train.parquet, id_map.csv, sample_submission.csv files to data directory.
- Build docker image using
sudo docker build -t env . - Run docker image
sudo docker run --name container env - Copy predicted file to host
sudo docker cp container:/app/submissions/stage_2_submission.csv ./submission.csv.