Houston, We've Got a Problem (I mean.. I have)!

Antonio Paes

unread,

May 26, 2015, 4:13:31 PM5/26/15

to caffe...@googlegroups.com

Hi guys, I'm trying train my own net with LFW dataset: http://vis-www.cs.umass.edu/lfw/

I'm using 9.000 images for train and 4.233 for validation, size of images is 58x58, and I labeled my images with values between 0......N. I create the databases, based on imagenet tutorial: http://caffe.berkeleyvision.org/gathered/examples/imagenet.html

but my accuracy is always 0 and loss between 6.16 and 7.90. I will attach the network definition file and train and val files.

I designed a small architecture defined in paper of Guosheng Hu 2015 named: When Face recognition meets Deep Learning.

I don't no what I'm doing wrong, anybody help me?

Thanks guys!!

test.prototxt

train_alig.txt

val_alig.txt

Antonio Paes

unread,

May 27, 2015, 7:09:25 AM5/27/15

to caffe...@googlegroups.com

Anybody helps ?

Antonio Paes

unread,

May 27, 2015, 6:13:32 PM5/27/15

to caffe...@googlegroups.com

Sameone?

Axel Angel

unread,

May 30, 2015, 3:50:58 AM5/30/15

to caffe...@googlegroups.com

Can you post your solver and train_test prototxt? How is your loss behaving? Have you tried to decrease the learning rate by 10/100x? You can try to add more units in your layer to make sure it has enough capacity to learn (try to overfit first, then simplify). Considering there is so few samples per class, maybe you should try a siamese network instead. You may be interested into other topics here like DeepFace.

Antonio Paes

unread,

May 30, 2015, 9:07:34 AM5/30/15

to caffe...@googlegroups.com

Hi thanks for answer, follow my solver and train_test.prototxt, I'm using leraning rate in 0.001, change output at last fully conected layer the behavior of my net change a little, I'm get 0.11 of accuracy, but loss is very strange, variyng between 0,000234 and 9.8767687e-05, this is strange correct?