CUDA: memory error on darknet yolov1

Question

I am trying to run darknet code YOLOv1 on a Tesla K40 GPU.

Everything worked fine a few days ago with the "ARCH = -gpu-architecture = compute_20 -gpu-code = compute_20" parameter in the makefile.

But, Today, when I did again. "make" worked, but when running the YOLO demo the following error appears on the video

I've tried different ARCH options like compute_30, 35, 52 .. But nothing works.

What are the possible solutions?

Note: darknet yoloV2 works fine. I am implementing the following blog: http://guanghan.info/blog/en/my-works/train-yolo/

+3

Nidhi Apr 25. 17 at 15:16

1 answer

Roxy · Answer 1 · 2017-06-01T14:58:15+0000

Have you tried modifying cfg / tiny-yolo-voc.cfg ?

It should look like this:

[net]
batch=64
subdivisions=64
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
...

The trick is to make the package equal units .