OoD-Diffusion

Environments

Python 3.8, PyTorch 1.10, CUDA 11.3

Docker image

Docker image for this project is available.

$ docker pull hushon/ood-diffusion

I/O format

JSON format is used to store the inputs/outputs of this module. Input includes the training images and the test images. Ouput includes the per-sample OoD evaluation scores.

sample_input.json

{
    "train_images": [
        "./dataset/train/0.jpg",
        "./dataset/train/1.jpg",
        "./dataset/train/2.jpg",
        "./dataset/train/3.jpg",
        "./dataset/train/4.jpg"
    ],

    "test_images": [
        "./dataset/test/0.jpg",
        "./dataset/test/1.jpg"
    ]
}

sample_output.json

{
    "test_ood_score": {
        "./dataset/test/0.jpg": 0.1111,
        "./dataset/test/1.jpg": 0.7999
    }
}

Pre-trained weights

Pre-trained PyTorch checkpoints are available at the releases page. Currently we have a model trained on DramaQA dataset available.

Training

python main.py --phase train --log_dir ./results/ --input_json_path ./sample_io/sample_input.json

The checkpoint file will be stored in the log_dir directory.

Testing

python main.py --phase test --ckpt_path <checkpoint path> --input_json_path ./sample_io/sample_input.json --ouput_json_path ./sample_io/sample_output.json

The evaluation results will be stored to ouput_json_path file.