Awesome

Source Code for Trade-offs in Fine-tuned Diffusion Models Between Accuracy and Interpretability

UPDATE: The paper got accepted as oral to AAAI as part of the Main Conference Safe, Robust, and Responsible AI (SRRAI) track 2024!

A preprint is available at: https://arxiv.org/abs/2303.17908

Training

Go to main directory (with ./src) and add the package to the python directory:

export PYTHONPATH=$PWD
pip install -e . 

Train baseline model. Requieres a pre-trained Stable diffusion v2 checkpoint We use the 512x512 model for our experiments.

As a preliminary task you have to prepare a directory with the trainings dataset. The folder with the data has to contain a .csv list with the relative path to all files name train2017_meta.csv or mimic_metadata_preprocessed.csv Examples can be found in ./experiments/train2017_meta.csv

You need to set your own paths in src/experiments/default_cfg_mscoco.py:

config.data_dir # path to .csv containing paths to images and their correspoingng text label
config.work_dir # path to this repo (where ./src is located)
config.ckpt # path to Stable diffusion v2 512x512 img geneation ema model

Start the finetuning with:

python scripts/train_baseline.py src/experiments/default_cfg_mscoco.py mscoco

Generative Results

To sample the model prepare csv file sample.csv

python scripts/sample_model.py src/experiments/default_cfg.py sample_baseline --ckpt=path/to/finetuned/model.ckpt --N=10 --label_list_path=experiments/p19_test.csv 

where experiments/p19_test.csv has the same structure as mimic_metadata_preprocessed.csv.

Localization

To reproduce localization results from Table 1 and Table 2 and Table 3 (requires MS_CXR_Local_Alignment_v1.0.0.csv from MS-CXR in data_dir):

Mimic

python scripts/compute_bbox_iou.py src/experiments/default_cfg.py mimic --ckpt=path/to/finetuned/ckpt.ckpt --filter_bad_impressions

add "--filter_bad_impressions" to reproduce results from Table 7.

MS-COCO

python scripts/compute_bbox_iou_multi_caption.py src/experiments/default_cfg_mscoco.py mscoco_singlegpu --ckpt=path/to/ckpt/512-base-ema.ckpt --phrase_grounding_mode --mask_dir=output/save/dir

Reference

@misc{dombrowski2023tradeoffs,
  title={Trade-offs in Fine-tuned Diffusion Models Between Accuracy and Interpretability}, 
  author={Mischa Dombrowski and Hadrien Reynaud and Johanna P. Müller and Matthew Baugh and Bernhard Kainz},
  year={2023},
  eprint={2303.17908},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}