Installing deeplabcut in GPU

Hi I was trying to install DLC with gpu but couldn’t succeed yet. I am using Quadro 4000 GPU, Cuda 8, NVIDA driver 377.83 and default tensor flow version 1.13 in windows 10. Here are my few questions and would really appreciate for your help and suggestion.

  1. After installing driver and CUDA 8 I checked with nvcc-V it shows CUDA being installed but nvidia -smi it doesn’t show any CUDA. As I read some where the order of operations matters should I install first CUDA-8 followed by driver and then DLC or first the driver then CUDA 8 and the DLC.

  2. I came to realize for windows with cuda-8, I need to have tensor flow 1.4.0 however the default version is 1.13.0. Do I need to change it to 1.4.0?

This is my first time using CUDA and GPU so please excuse me if the questions are too naïve.

Hi,

nvidia-smi doesn’t show CUDA version until 410.72 driver. The default tensorflow-gpu in current GPU env is actually 1.15.5.

You can also use newer CUDA (10 for instance) and newer driver (I run the latest and update regularly on one of my computers since it’s also my gaming rig). Anyway it should also work with what you have.

Installing CUDA it asks you to have Visual Studio installed for CUDA code compilation - did you install it prior to installing CUDA?

You say you didn’t succeed in setting up GPU env, but report no error. What exactly isn’t working?

Hi Konrad thanks

The GPU I have (Quadro 4000) is quite old and the driver compatible with it says it only supports CUDA 8. I am not sure if CUDA 10 will work on it. I installed visual studio community 2015 before installing CUDA. The problem was I started training with DLC-GPU but it was too slow later in I checked my GPU usage in task manager and found GPU not being used at all.

I will try to reinstall everything and try once again. Meanwhile, do you think should I change the tensor flow to 1.4 for CUDA-8.

Sorry, thought you meant Quadro RTX 4000 (Nvidia and their naming schemes…). You are correct, CUDA 8 is the last one where Quadro 4000 was supported.

According to Which TensorFlow and CUDA version combinations are compatible? - Stack Overflow you should use python==3.6 and tensorflow-gpu==1.4.0

That being said, older version should work too. Did you run the testscript? Where there any errors?

Also, the only indication in task manager that you’ll see is in the gpu memory, not in “usage” or processors. For DLC training, all of the memory will likely get used, but otherwise the gpu will not be working that hard.

Thank you all. I will try my luck one more time. Meanwhile can you recommend any new GPU. We are planning to buy the new one. Mostly I see people using GTX 1080 mostly for but I don’t think that’s available now.

In the current market, if you don’t have a GPU, best way would be to use a cloud service. Due to recent issues with silicon supply/demand you’ll have a hard time getting a good GPU at non-absurd price.

Though if the price isn’t an issue, I’d wait for 3080Ti performance reviews. 3060 (non-Ti since it has more VRAM) is not a bad option, as to 3090 it depends how it will look compared to the new 3080Ti.
If you want an older generation, 2080 Super, 2080Ti are good. 1080 or 1080Ti aren’t available anymore as far as I know - means you’d have to get a used one which I do not recommend if you cannot properly evaluate it’s condition.

2 Likes