conda create -n hierarcaps python=3.9
conda activate hierarcaps
pip install -r requirements.txt
python train.py
Run with --help / -h to see all arguments and default values.
You may also download fine-tuned CLIP-B and CLIP-L checkpoints.
To run inference on HierarCaps:
python eval.py (-bc ...) (-w ...) # quantitative evaluation (1K-item test set)
python qual.py -i imgs_dir (-bc ...) (-w ...) # qualitative evaluation (expanded test candidate set)
Use optional -bc and -w flags to change which model is loaded (base pretrained model and fine-tuned weights respectively). For qualitative tests, imgs_dir is a directory containing the image files to evaluate on. Run with --help / -h to see all arguments and default values.