Training Openface

[Siddha Ganju]

I was hoping you could share the dataset or the identities that you used for training Openface. The documentation says that you used 500k images for 10k people/classes from the CASIA webface + facescrub datasets. However, joint training on both these datasets gives around 595,594 (webface = 455,594 and facescrub = 140,000) images from 11106 people/classes. Applying a threshold of 10 images per person (for better training; as was mentioned somewhere in the google group) gives a total of 603,850 images with 10989 people/classes. 

I've tried training with various thresholds of 10, 20, 25 on the joint webface+facescrub datasets but I am not able to get more than 82% on the LFW dataset (according to the testing lfw.py script provided in the demo)

Secondly, I'm not able to understand how the embeddings are mapped to the triplets, and is there any way to calculate how many approximate triplets will be produced for each batch (because of the random choice of the anchor-positive pair only those that satisfy the alpha condition out of the total number of people * number of photos per person will be accepted)

Thank you very much!