Cluster proposal configuration for YouTube Faces about learn-to-cluster

[yl-1993]

Hi @jaantoots, thanks for checking out our project. I add the related configs in commit 09c7262 to reproduce the result on YouTube-Faces. Besides, I find the pretrained model has unexpected prefix module.base. and it is corrected now. You can reproduce the result by the following two steps:

1. Download the latest YouTube-Faces in DATASET.md or use python tools/download_data.py --data youtube_face.
2. Run scripts/dsgcn/test_cluster_det_ytb.sh

The result is expected to be ave_pre: 0.9674, ave_rec: 0.9225, fscore: 0.9444, which is close to the result reported in the paper.

Feel free to have a try after PR #32 is merged.

[yl-1993]

Hi @jaantoots, PR #32 is merged. Feel free to have a try and let me know if there is any problem.

[jaantoots]

Many thanks for the update. This worked as expected.

As an aside, it wasn't clear from the paper that the YouTube Faces result was GCN-D only and no outlier segmentation. This is implied in multiple places in this repo so no problem but it would be great to mention this more explicitly whether GCN-S didn't improve results or you simply haven't had the time to test it. I could (re)train the models myself to see but have not got around to it either 😄