LMobileNet, what does 'L' mean?

[HaoLiuHust]

No description provided.

[nttstar]

Means using stride11 in its first conv layer.

[HaoLiuHust]

@nttstar Thanks for you reply, I am training now. another question,

LMobileNetE	0.99633+-0.00314

is this model use sphereloss? and seems vggface2 have some overlap with lfw, have you do some clean?

[nttstar]

Softmax only, if not specified. Overlap exists between vggface2 and lfw but in my current experiments they're not removed. You can compare the relative number if the performance gap is larger than 0.1.
You can also check https://github.com/happynear/FaceDatasets for overlaps. I have 3 more runtime verification testing other than LFW and will be released soon.

[HaoLiuHust]

Thanks for your work, so sphereloss did not get better than softmax in your experiment?

[nttstar]

It depends on your network and dataset. Now I can not get sphere working for se-resnet50@vggface2, but it works for another dataset. If you can make it work on vggface2, please let me know.

[HaoLiuHust]

I am training LMobileNetE@vggface2 with sphereloss, still finetuning the network for now, I can only get about 98.8% acc on lfw, I will try more

[nttstar]

I can get 99.6% on LMobileNetE@vggface2@sphere and better acc of other metrics than softmax.

[HaoLiuHust]

what's your training params? I have clean the vggface2 at first, and for now I can only get 98.8 use "adam" optimizer, beta = 1000, margin = 4, scale = 0.9993,
beta_min = 4., verbose = 1000

[nttstar]

You can try yourself as I did not clean the data. Commonly I use SGD optimizer with mom=0.9 and lr=0.1, wd=5.e-4. beta = 1000, beta_min = 5. lr decrease to 0.01 and 0.001 after some epochs.

[HaoLiuHust]

for sphereloss, the evaluation metric is still euclid distance?

[nttstar]

Yes. And there should be no difference between euc distance and cosine sim if you normalize the vector firstly,

[HaoLiuHust]

Thanks, your input shape is 112x112? have you try 112x96？ since you mentioned in the description

[nttstar]

112x112 is slightly better than 112x96 on large pose variation

[nerddd]

@nttstar What's the reference points of 112*112 when aligning face?Thanks

[nttstar]

@nerddd +8 on x-axis from the points of 112*96