Making my own face recogniser - same results independent of input image

Thomas Wood

unread,

Feb 22, 2016, 5:45:38 AM2/22/16

to Caffe Users

Hi,

I am working on training a face recogniser. The recogniser I am planning will be trained on a few thousand faces.

As an easy starting point, I am trying to make a neural net that can differentiate 2 faces: Princess Diana and David Beckham (these were faces that it was easy to get a lot of images for!)

I have preprocessed the images with OpenCV so my training set is a set of 227x227 images each with only the face in. I've put 300 images of each. I know this isn't much, but it's more a practice run than anything.

My net is a modified CaffeNet file (attached), and I've also taken the CaffeNet solver.

I've been able to train it for 9000 iterations.

What happens when I then run my trained model is that every time, I get the same output values, no matter what input image I give it:

David Beckham 0 0.432829

Princess Diana 1 0.567171

The numbers are exact every time to all visible decimal places.

This looks odd to me, as I would not expect excellent performance, but I would at least have expected a change in output for different input images.

Has anyone seen this effect before? Is there something obvious that I'm missing?

I know CaffeNet isn't the most adequate net architecture for the problem, also I've been looking at the VGG net (http://www.robots.ox.ac.uk/~vgg/software/vgg_face/), but they only supply the deploy.prototxt file and I was not sure how to convert this to a train_test.prototxt.

I would be grateful for any advice, especially if someone has tried making their own face recogniser from scratch as I am doing.

Thanks

Tom

tw-caffenet-solver2.prototxt

tw-caffenet-train_val.prototxt

Nam Vo

unread,

Feb 22, 2016, 8:56:43 PM2/22/16

to Caffe Users

Did the loss drop and accuracy increase during training.
I suggest to include accuracy layer for training phase to monitor the accuracy.

Thomas Wood

unread,

Feb 23, 2016, 7:29:10 AM2/23/16

to Caffe Users

Thank you Nam. I will check that, it looks like at some points the loss has got to this:

loss = -nan

So maybe that will be the problem. I will try adding the accuracy layer

Nam Vo

unread,

Feb 23, 2016, 4:46:44 PM2/23/16

to Caffe Users

Yeah you want to make sure the training is working first.
Look like it isn't here, I suggest lowering the learning rate till it does, that might be the problem.

Thomas Wood

unread,

Feb 25, 2016, 6:42:21 AM2/25/16

to Caffe Users

OK thank you. If I get any problems I will post again.

Dhruv Bhandari

unread,

Oct 23, 2016, 5:34:10 AM10/23/16

to Caffe Users

Hi Thomas,

I'm facing the same error. My net is producing the same feature vector (1x4096) for all inputs. I checked the euclidean distance between feature vectors of two different images and it was found to be ZERO.