RASTA

The Rasta project aims at recognising art style automatically using pixel data only (i.e. no meta-data). For more details about the methodology and the results, see the publication.

Adrian Lecoutre, Benjamin Negrevergne and Florian Yger. Rasta: Recognizing art style automatically in painting with deep learning. In Asian Conference on Machine Learning, 2017.

How to use this code

Setup Rasta Python environment with Pip

You can install python requirements with

pip3 install -r python/requirements.txt

See python/requirements.txt for the complete list of requirements.

If you have access to GPUs we encourage you to use them, this will speedup both inference and training. To use GPU, install tensorflow-gpu in addition to previous packages.

pip3 install tensorflow-gpu

Download model files (mandatory)

cd models
wget www.lamsade.dauphine.fr/~bnegrevergne/webpage/software/rasta/rasta_models.tgz
tar xzvf rasta_models.tgz
cd ../

Download extra data files (optional)

If you want to download the full wikipaintings dataset (the one from WikiArt), execute the following commands. Warning: the file is about ~20GiB, we suggest that you first try with the small datasets provided in data/wikipaintaings_small.

cd data
wget www.lamsade.dauphine.fr/~bnegrevergne/webpage/software/rasta/wikipaintings_full.tgz
tar xzvf wikipaintings_full.tgz
cd ../

Predict the style of one image

Use

python3 python/evaluation.py -t pred  --data_path=PATH_TO_IMAGE

Where PATH_TO_IMAGE points toward a valid jpeg image file.

See python3 python/evaluation.py -h for more details

Evaluate Rasta models on a large batch of images

You can evaluate the accuracy of the default Rasta model using:

python3 python/evaluation.py

This will evaluate the accuracy on a the small test set available in wikipaintings_small/wikipaintings_test using the model in models/default.

You can evaluate the accuracy of other Rasta models, or using other datasets with:

python3 python/evaluation.py --model_path=MODEL_PATH --data_path=DATA_PATH

where MODEL_PATH is the path to a .h5 model file, and DATA_PATH is a path to a directory containing the test set. In the test set, there should be one sub-directory for each class containing all the images of this class. See wikipaintings_small/wikipaintings_test for an example.

See python3 python/evaluation.py -h for more details about the options.

Note: At the moment --isdecaf is necessary if you want to evaluate models based on decaf. Hopefully, this will be fixed soon.

License

See LICENSE file.

Authors

• Adrian Lecoutre
• Benjamin Negrevergne
• Florian Yger

Contact author

Benjamin Negrevergne: firstname.lastname @ dauphine.fr