Train 128 model / 512 model

[Ram-Godavarthi]

HI,

Is it possible to train for 512 facenet model ??

What is the procedure for it ??

Btw, how to get the dataset for the training ?

[swghosh]

Are you referencing the dataset that was used in the original FaceNet paper? If yes, that won't be possible because FaceNet paper uses a proprietary dataset created by researchers at Google which was never brought it into the public, however you can find helpful information about the dataset on the paper and most likely it resembles to other publicly available datasets like MS-Celeb-1K or MegaFace IIRC.

[swghosh]

What do you mean by 512 / 128 models, I am unable to recall the exact details in the paper actually cause this repository is quite old. If you could point me something specific on the original FaceNet paper, I might be able to recall and offer some help. AFAIR the CNN architectures in this repo are accurate and exactly resembles the one in the paper. The paper had two variants ZF-Net and Inception based FaceNet, my implementation is the one based on Inception (origin Inception network modified for some changes to adapt it to the FaceNet face verification task).

Also, this repository is quite unmaintained and the implementation of Focal Loss does not exactly resemble the one in the paper which required resampling of individual classes at train time (unimplemented). If you're interested in digging deeper you can look into this wonderful repo by David Sandberg: https://github.com/davidsandberg/facenet (though it's quite old now).

[Ram-Godavarthi]

Thank you for your detailed explanation.
I want to do Face verification task.

I use either 128 / 512 embeddings model to achieve that.

I need a method to train a model with 128 embeddings or 512 embeddings using custom data or some available data.

Do you know how to train the network ?? What are the requirements for it in terms of dataset format and labelling format?