Minimum GPU requirement for demo purposes

[Howarto]

Hi!

First of all, amazing job. I read the paper and it seems quite good 👍 I am a developer and this project seems to solve my problem, I am trying to count water fleas to help my wife with its research 😅, and to count a few objects from my garage (bars and tubes). I was trying to build something on my own, but it was hard without using the correct vocabulary (e.g: low-shot, prototype...) and I could get a lot of references from this.

I have only one question. I read the training methodology and you used two A100 GPUs. Also to run the demo? I'd like to know if you used a less powerful graphics card to run the demo. My 1060 6GB runs out of memory, and I want to see about renting an AWS instance or buying a new card. Obviously, I can't afford an A100

Thank you!,
Diego.

[jerpelhan]

Hi Diego,

Thanks a lot for your kind words! :). The two A100 GPUs were only needed for training. For inference (running the demo), you don’t need that kind of hardware. You should be able to run it on more modest GPUs if you enable mixed precision, i.e.,:

with torch.no_grad():
    with torch.cuda.amp.autocast(dtype=torch.float16):
        X = model(img, boxes)

This reduces memory usage significantly, it works on my local 8GB card. That said, the model was trained on FSC147, so its generalization is still limited. By the end of the year, we plan to release a stronger model with much better generalization.

Hope this helps, and best of luck!

[Howarto]

I tried and it seems that my gpu is not capable 😅 So, 6GB seems to be not enough. I'll check if some friend can run it with a better gpu or I'll get the AWS instance.

This is the memory before closing the matplot window to draw the rects:

[cripsgreen]

Hello! I’d like to ask how long it approximately takes to train one epoch. Is the total training set to 200 epochs? Also, during training, does each epoch reload the dataset? Thank you! @Howarto

[Howarto]

Hey @cripsgreen I only tried to run the existing demo, not training. My project (remember that I was not the repository owner!) was paused because I had more work to do, but probably this next week I will continue.

I got a g4dn.xlarge instance from AWS and the demo worked well. Now I am only changing the main script to create a ssh connection and redirecting the X server.

[jerpelhan]

I recommend trying out the recently released and less GPU-hungry GECO2: https://github.com/jerpelhan/GECO2