|
|
--- |
|
|
title: PinPoint |
|
|
emoji: 😻 |
|
|
colorFrom: gray |
|
|
colorTo: pink |
|
|
sdk: static |
|
|
pinned: false |
|
|
license: mit |
|
|
short_description: code for the submission 1386 |
|
|
--- |
|
|
|
|
|
# Pinpoint Counterfactuals: localized gender counterfactual generation (NeurIPS 2025 Datasets and Benchmarks track. Submission 1386) |
|
|
|
|
|
## Getting started |
|
|
|
|
|
To generate PinPoint Counterfactuals, take the following steps. |
|
|
|
|
|
### Download the data |
|
|
|
|
|
First, download the <a href="https://ai.meta.com/datasets/facet-downloads/">FACET</a> and <a href="https://ai.google.com/research/ConceptualCaptions/download">CC3M</a> dataset. Unpack them in the directory of your choice. |
|
|
|
|
|
### Generating PP masks |
|
|
|
|
|
Use the `Color-Invariant-Skin-Segmentation` module to generate masks, following the methodology outlined in the main submission manuscript. |
|
|
|
|
|
### In-paint the images |
|
|
|
|
|
Use the `BrushNet` module to in-paint the images from FACET and/or CC3M (see the respective scripts in `BrushNet/examples/brushnet/inapaint_*.py`. |
|
|
|
|
|
|
|
|
## Zero-shot classification |
|
|
|
|
|
Use the `zero_shot_classification.py` script to test the occupation classification accuracy of different CLIP models (for different in-painting setups, i.e. PP, PP*, WB, etc.). To run it, first install PyTorch and the following dependencies: |
|
|
`pip install open_clip` |
|
|
`pip install git+https://github.com/openai/CLIP.git` |
|
|
`pip install tqdm` |
|
|
`pip install numpy` |
|
|
`pip install pandas` |
|
|
`pip install pillow` |
|
|
|
