acmi-lab / CHILS

Code and results accompanying our paper titled CHiLS: Zero-Shot Image Classification with Hierarchical Label Sets

CHiLS

This is the official implementation for CHiLS: Zero-shot Image Classification with Hierarchical Label Sets. If you find this repository useful or use this code in your research, please cite the following paper:

Zachary Novack, Julian McAuley, Zachary Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning (ICML), 2023.

@inproceedings{novack2023chils,
    title={CHiLS: Zero-Shot Image Classification with Hierarchical Label Sets},
    author={Novack, Zachary and McAuley, Julian and Lipton, Zachary and Garg, Saurabh},
    year={2023},
    booktitle={International Conference on Machine Learning (ICML)}, 
}

There are three main steps for recreating the paper results:

1. Setting up the environment and datasets
2. Caching the CLIP-extracted features for each dataset and model
3. Running zero-shot inference

Setting up the environment and datasets:

All requisite packages can be installed via the environment.yml file. For access to GPT-3 through OpenAI, you must have an account and save your access token in the environment variable OPENAI_API_KEY.

Besides ImageNet, CIFAR100 and Fashion-MNIST (which can be autoloaded through the torchvision API), each dataset can be downloaded through the standard websites for each: Office31, OfficeHome, Food-101, Fruits360, Fashion1M, LSUN-Scene, ObjectNet. Dataset Notes:

• Both LSUN-Scene and Fashion1M must be configured into the ImageFolder format, wherein the directory has named folders for each class, each containing all the images. Due to compute constraints, for LSUN-Scene we use the validation data only and for Fashion1M we use the first two large image folders (i.e. 0 and 1).

Caching the CLIP-extracted features for each dataset and model:

Running run.py will use the variables specified in config.yaml and extract the features of a given dataset and CLIP model. In order to run this, the variable data_loc must be changed to the directory where your datasets are held.

Running zero-shot inference:

Once the features are extracted, you may run zshot.py to generate the zero-shot inference results with CHiLS. For example, to generate the results with the GPT-generated label sets (which are provided for reproducibility) on Food-101, the command would be:

python zshot.py --dataset=food-101 --model=ClipViTL14 --experiment=gpt --label-set-size=10 --data-dir=[INSERT YOUR PATH HERE]

See the src/constants.py file for valid inputs for each argument in the command.