Reproducibility study for "Explaining in Style: Training a GAN to explain a classifier in StyleSpace"
This repository contains our experiments in reproducing Lang et.al's "Explaining in Style: Training a GAN to explain a classifier in StyleSpace".
Original Paper Material: Website GitHub
If case you want to run our code locally, you need to activate the following environment in anaconda.
We strongly recommend running both our notebooks in Colab, since this is the environment we used.
conda env create -f environment.yml
To train our implementation of the StylEx model in the PyTorch, run the StylEx.ipynb
(inside train folder) notebook:
To save time we provide our pre-trained models and .py
files needed in a .zip
file in Google Drive and load them in the following Colab notebook to reproduce our results.
To reproduce the results shown in our paper, run the Att_Find.ipynb
notebook:
Top 4 attributes for the perceived age classifier
Independent selection strategy - Top-5 detected attributes
Classification study results
Verbal description study results
The files legacy_modified.py
and mobilenet_pytorch.py
can be used to transfer the original paper's pre-trained models from TensorFlow to Pytorch. For this, the pre-trained models provided in the authors' notebook (GitHub) are needed.
All contributions are welcome. All content in this repository is licensed under the MIT license.