Face_recognition: GPU running out of memory

[gavanpang]

• face_recognition version:
• Python version: 3.6.5
• Operating System: Windows 10

Description

I've installed dlib with CUDA support on a GTX1060 3GB.

I've loaded 2 images, and have called

face_recognition.face_locations(img, number_of_times_to_upsample=0, model="cnn")

on each image separately. The first image is parsed correctly, and I receive the coordinates of the face in the first image. But trying to detect faces in the second image, I receive the error

RuntimeError: Error while calling cudaMalloc(&data, n) in file C:\Users\Gavan\dlib-master\dlib\cuda\cuda_data_ptr.cpp:28. code: 2, reason: out of memory

I've run nvidia-smi and confirmed that the program fills up ~2.7GB of VRAM upon running face_locations on the second image, then the crash happens. Is it possible to force VRAM to be flushed? Or is there an easier solution I'm not aware of?

[lolstatsguy]

Maybe try to resize the image. cv2.resize or scipy.misc.imresize. Try different values 1/2 or 1/4 etc

[gavanpang]

Thanks lolstatsguy, you've solved this particular issue for me. I'm going to push the face_recognition api further and see how far I get. Cheers!