INSA Lyon Deeplearning Course - exercice 1 - Ball detection and forecasting with deep learning on a synthetic dataset
By Paul-Emmanuel Sotir
############## Installation ##############
git clone https://github.com/PaulEmmanuelSotir/BallDetectionAndForecasting.git
conda env create -f ./environement.yml
conda activate pytorch_5if
# Downloads datasets ('curl' and 'tar' packages needed on a Linux distro (or WSL - Linux subsystem on Windows)):
bash ./download_datasets.sh
############## Usage examples ##############
# Trains ball detection model
python -O ./src/train.py --model detect
# Performs an hyperparameter search for ball detection model (hyperopt)
python -O ./src/hp.py --model detect | tee ./docs/hp_search_logs/hp_detect4.log
# Trains ball position forecasting model
python -O ./src/train.py --model forecast
# Performs an hyperparameter search for ball position forecasting model (hyperopt)
python -O ./src/hp.py --model forecast | tee ./docs/hp_search_logs/hp_forecast2.logOnce hyperparameter searchs are done (or still running), you can use balldetect.torch_utils.extract_from_hp_search_log() and balldetect.torch_utils.summarize_hp_search() function to parse and summarize hyperparameter search results. See ./notebooks/hp_search_results.ipynb (or ./docs/2_annexe_hp_search_results.html / ./docs/2_annexe_hp_search_results.pdf) for some hyperparameter search results of our own.
For (much) more details on this deeplearning course project see report located here: ./docs/rapport.md (or ./docs/0_rapport.html / ./docs/_rapport_export.pdf) (unfortunately, this report is in French.)