nn4.small2.v1 model released: LFW accuracy improved from 91.5% to 93.6% and better performance.

[Brandon Amos]

Hi OpenFace users,

This is a smaller announcement of the latest model, nn4.small2.v1, that improves

the LFW accuracy from 91.5% to 93.6%.

It uses a smaller model and improves the runtime performance

compared to nn4.v2 from 679.75 ms to 460.89 ms on an 8-core 3.70 GHz CPU

and from 21.96 ms to 13.72 ms on a Tesla K40 GPU.

The same notes from https://groups.google.com/forum/#!topic/cmu-openface/wsUp0p0F960 apply:

these embeddings aren't compatible with our other models.

I've updated the "Models and Accuracies" page at http://cmusatyalab.github.io/openface/models-and-accuracies/

with more information about the available models.

The model definition is available at

https://github.com/cmusatyalab/openface/blob/master/models/openface/nn4.small2.def.lua

and the model is available at

http://openface-models.storage.cmusatyalab.org/nn4.small2.v1.t7.

This improvement is from manually making a smaller neural network

than FaceNet's original nn4 network with the (naive) intuition that a small

model will work better with less data.

I think further exploring model architectures will result in better

performance and accuracies.

I think the best approach to this is to randomly sample hyper-parameter

choices, train for a day while saving the best result, then repeat.

I won't start implementing this for a while,

but I'll track progress at this GitHub issue:

https://github.com/cmusatyalab/openface/issues/78.

I've labeled it with the 'help wanted' tag for now if anybody wants to contribute.

-Brandon.