About MS-Celeb-1M dataset about learn-to-cluster

[yl-1993]

@luofan18 The face dataset we used is face_emore. But it seems that the filename is a little different now, which has been reported by another researcher:

0_5822656
0_5822657
0_5822658
...

The format of our provided list is as follows:

part6_test.list:85741/85741_5822650.jpg
part6_test.list:85741/85741_5822651.jpg
...

As each image is named as 'faceid/faceid_imageid.jpg', you can extract the imageid to match the corresponding image.

[luofan18]

Thank you for providing the link. I have downloaded the dataset.

Besides, I found that you also provided pretrained model and extracted features for YoutubeFace at here https://drive.google.com/file/d/1hg3PQTOwyduLVyfgJ7qrN52o9QE35XM4/view?usp=sharing.

However, the F score is only 0.6004 for the pretrained model. The program also says

The model and loaded state dict do not match exactly 

unexpected key in source state_dict: base.layers.0.gc.weight, base.layers.1.gc.weight, base.classifier.weight, base.classifier.bias 

missing keys in source state_dict: classifier.weight, layers.1.gc.weight, classifier.bias, layers.0.gc.weight 

[yl-1993]

@luofan18 It seems that the pretrained model does not match the built model, thus the output is basically from a random initialized model. I will check it in these two days.

[yl-1993]

Hi @luofan18, I find it is my mistakes that the pretrained model has unexpected prefix module.base.. I have corrected it through commit 09c7262.

Feel free to download the latest YouTube-Faces in DATASET.md or use python tools/download_data.py --data youtube_face, to have a try.