evaluation/README.md

Style transfer quaintitative evaluation using Deception Score

How to calculate Deception Score:

1. Run ./download_evaluation_data.py to download the weights for artist classification model.
2. Set results_dir variable in eval_deception_score.py:92 to point to the directory with stylized images. All images generated by one method must be in one directory. Image filenames must be in format "{content_name}_stylized_{artist_name}.jpg", for example: "Places366_val_00000510_stylized_vincent-van-gogh.jpg".
3. Run ./run_deception_score_vgg_16_wikiart.sh
4. Read results in the log file in ./logs directory.

How to evaluate your own model:

• Download validation sets from MSCOCO (val2017.zip) and Places365 (val_large.tar).
• To compare with deception score reported in the paper run your stylization model on the content images listed in eval_paths_700_val.json.