Mattt29/TER_Art

The pipeline

1. Train the model with train_resnet50.py. This will save the best epoch in checkpoint.torch.
2. Export features with predict_resnet50.py. This will save an attributes.csv containing infos about all the tasks of the current dataset used and if the image was used in train or test for each features and a matrix features.npy.
3. Reduce feature dimension with dim_reduction.py. The result will be csv files for every reduction dimension chosen. They will have this form {dataset}_{task}_{dim_reduction}.csv . Each of them will be consisting in attributes.csv completed with the two selected features from features.npy after dimension reduction applied on it.
4. Plot the features with selected colors with feature_plots.py.

Utilitaries

The file train_util.py contains utilitaries functions mostly used to manipulate the neural network.

Constructing the dataset

The file construct_splits permits to construct train and val folder for each task containing subfolder for each class. This enables the use of ImageFolder from PyTorch. The function does not use extraspace as it makes use of hard links.

Wikiart

• Download Wikiart here

• Put your Wikiart images in

static/images/wikiart/data