F1 normalization with applications in computer vision.

conda env create -f environment.yml
conda activate f1-cv
pip install "git+https://github.com/philferriere/cocoapi.git#egg=pycocotools&subdirectory=PythonAPI"
pip install slidingwindow

conda env create -f environment.yml
conda activate f1-cv
pip install pycocotools
pip install slidingwindow

(f1-cv) $ mv configs/_deprecated/full_coco_stuff/coco_stuff_official_eval.yml configs/
(f1-cv) $ python main.py configs/coco_stuff_official_eval.yml

Find the experiment you are interested in under experiments/archive, and then just read the config .yml file to replicate. For example, consider experiments/archive/coco_stuff/coco_stuff_full/coco_stuff_f1_deep_lab_amazon/coco_stuff_f1_deep_lab_amazon.yml:

agent: COCOStuffF1Trainer

# Dataset
dataset path: /home/ubuntu/data/filtered_datasets/full_stuff_amazon

...

This experiment uses the COCOStuffF1Trainer in file coco_stuff_f1_trainer.py with the full_stuff_amazon dataset. You can generate the dataset from https://github.com/abhay-venkatesh/coco-stuff-tools.

You can directly use the coco_stuff_f1_deep_lab_amazon.yml file as an argument into main.py to replicate.