search

philipperemy / yolo-9000

YOLO9000: Better, Faster, Stronger - Real-Time Object Detection. 9000 classes!
Apache License 2.0
1.18k stars 309 forks source link

memory requirements #9

Closed mhnatiuk closed 6 years ago

mhnatiuk commented 6 years ago

Are memory requirements for running detection on yolo-9000 higher than yolov2 on coco? If so, can someone provide an estimate for memory req.?

philipperemy commented 6 years ago

I have no idea of yolov2. But on my experiments, yolo 9000 uses around 4206MiB on the GPU. So it should not be a problem if your GPUs have more than 6-8GB.

philipperemy commented 6 years ago

If the model is equivalent to yolo v2, then the memory footprint should be the same. (The memory allocated is not proportional to the size of the data).

So I guess yolo v2 and yolo 9000 should be pretty similar since they share almost the same architecture (from what I remember).

ThibaultGROUEIX commented 6 years ago

I'm running it on 6GB with 562MB already taken for the desktop and i get an out of memory. I guess i must be slightly under the requirements.

philipperemy commented 6 years ago

@ThibaultGROUEIX thanks for the feedback! As you said, I guess 6GB is the limit. I think it would work if you didn't have your Desktop app running on the same GPU.

I can confirm it works flawlessly on a 8GB GPU (GTX 1070).

philipperemy commented 6 years ago

I added a note in the README file: 73e54e951f7b5b87ee88a91ce93415f4f90f09bf

Despite the fact that it can work with 6GB, I prefer to recommend 8GB.