GPU - Out of Memory error - but GPU not fully utilised (1/2)

• face_recognition version: v1.2.2
• Python version:3.6
• Operating System:CentOS

Description

I am training a set of images on a GPU machine, but it fails with Out of Memory error - while checking on the GPU usage only one GPU is being used and it is full.

What I Did

Code:

face_bounding_boxes = face_recognition.face_locations(image, number_of_times_to_upsample=2, model="cnn")
X.append(face_recognition.face_encodings(image, known_face_locations=face_bounding_boxes, num_jitters=100)[0])

Images are less than 100kb and if I try to reduce upsample and num_jitters, it is impacting my accuracy. Why is it using only one GPU when I have 2 GPUs on my machine.

Please advice on this. Thank you.

[essam-salah]

If you trying this repo, here are some details, hope it will be helpful:

• PostgreSQL implementation of euclidean distance calculation works great - no need to train all the existing images for adding new registration - you can skip the entire training and classifier part - and PostgreSQL supports CUBE datatype.
• Accuracy of Indian faces was average - I was able to improve this to some extent - by cropping the face and applying gamma scaling - this will remove unnecessary pixels from the image and improve the brightness - getting decent accuracy now.
• if you are using PostgreSQL, you can reduce the error by comparing against specific user encodings, I have tried 1-N comparison on 25K stored encodings with good accuracy - however, for production deployments, I would prefer 1-1 comparison.

Thanks!

Hello rathishkumar,

I am trying to use this model to build an attendance system using face recognition but I have an issue with the accuracy. I am trying to reduce the distance threshold to ignore the false positive recognition. So can you please elaborate more how could you improve the accuracy of the model ?

[rathishkumar]

If you trying this repo, here are some details, hope it will be helpful:

• PostgreSQL implementation of euclidean distance calculation works great - no need to train all the existing images for adding new registration - you can skip the entire training and classifier part - and PostgreSQL supports CUBE datatype.
• Accuracy of Indian faces was average - I was able to improve this to some extent - by cropping the face and applying gamma scaling - this will remove unnecessary pixels from the image and improve the brightness - getting decent accuracy now.
• if you are using PostgreSQL, you can reduce the error by comparing against specific user encodings, I have tried 1-N comparison on 25K stored encodings with good accuracy - however, for production deployments, I would prefer 1-1 comparison.

Thanks!

Hello rathishkumar,

I am trying to use this model to build an attendance system using face recognition but I have an issue with the accuracy. I am trying to reduce the distance threshold to ignore the false positive recognition. So can you please elaborate more how could you improve the accuracy of the model ?

I have shared my thoughts to a similar question here . Apart from these, I have used PostgreSQL for saving the encodings and computing the euclidean distance, I was trying to do a 1 - N ( N number of photos of a same person), so I was able to get good results. I have shared details here . Hope it helps.