This is the code for Self-supervised learning of class embeddings from video in ICCV workshop in 2019.

Note that this code has not been 'cleaned' and so is only given for further explanation of the paper.

It's run by calling

python train_attention_hierarchy.py --use_cyclic --use_confidence.`

Training yourself

In order to use this training code, it is necessary to download a dataset. To use the VoxCeleb2 dataset, download the dataset. It should be put into folders as follows and the environment variables updated appropriately: VOX_CELEB_LOCATION should be set to the location of VoxCeleb2.

For our datasets we organised the directories as:

IDENTITY
-- VIDEO
-- -- TRACK
-- -- -- frame0001.jpg
-- -- -- frame0002.jpg
-- -- -- ...
-- -- -- frameXXXX.jpg

If you arrange the folders/files as illustrated above, then you can generate np split files using Datasets/generate_large_voxceleb.py and use our dataloader. Otherwise, you may have to write your own.

Then you need to update where the model/runs are stored to by setting BASE_LOCATION.