"gputouse" option for analyze_videos

[mschart]

Hi there,

I have two GPUs and I could train in parallel using the "gputouse" option in train_network. I'd like to also analyze videos in parallel (as I have many) so would it be possible to include the same "gputouse" option in the analyze_videos function? If I start two different ones in parallel now I get a tensorflow error:

InternalError: Failed to create session.

Cheers,

Michael

[mschart]

Ps: I guess it's just including this line:

os.environ['CUDA_VISIBLE_DEVICES'] = str(gputouse)

[MMathisLab]

good idea (edited*: which Alex evidently already had implemented - hehe); typically I just run separate docker containers and they are individually linked to a specific GPU, but for non-Docker users, this is indeed useful.

[AlexEMG]

Yes, and you can set it e.g. by:
deeplabcut.train_network(config,shuffle=1,trainingsetindex=0,gputouse=3)
then just run it for your other GPU in another terminal...
See:
https://github.com/AlexEMG/DeepLabCut/blob/efa95129061b1ba1535f7361fe76e9267568a156/deeplabcut/pose_estimation_tensorflow/training.py#L12

[mschart]

Yes, for training. The same option would be good to have for the "analyze_videos" function.

[AlexEMG]

Wait, doesn't that exist as well: https://github.com/AlexEMG/DeepLabCut/blob/master/deeplabcut/pose_estimation_tensorflow/predict_videos.py#L34

[mschart]

Mhm, indeed. Then all is good and I did something silly :)

[kenziew]

I'm having a similar issue trying to train a network and analyze videos on 2 separate GPUs. The analyze videos function seems to call a function that utilizes both GPU's. Any ideas? Thanks!

[kenziew]

this is with just running analyze videos function

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.43 Driver Version: 418.43 CUDA Version: 10.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce RTX 208... Off | 00000000:08:00.0 Off | N/A |
| 48% 65C P2 258W / 260W | 10856MiB / 10989MiB | 100% Default |
+-------------------------------+----------------------+----------------------+
| 1 GeForce RTX 208... Off | 00000000:42:00.0 On | N/A |
| 41% 41C P2 64W / 260W | 438MiB / 10981MiB | 0% Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 4108 C /home/kenzie/anaconda3/envs/DLC/bin/python 10845MiB |
| 1 1709 G /usr/lib/xorg/Xorg 18MiB |
| 1 1816 G /usr/bin/gnome-shell 58MiB |
| 1 2077 G /usr/lib/xorg/Xorg 108MiB |
| 1 2206 G /usr/bin/gnome-shell 85MiB |
| 1 4108 C /home/kenzie/anaconda3/envs/DLC/bin/python 155MiB |
+-----------------------------------------------------------------------------+

[AlexEMG]

Just pass which GPU you want to use, e.g.

deeplabcut.train_network(config,gputouse=1)

or

deeplabcut.analyze_videos(config,videos,videotype='avi',shuffle=1,trainingsetindex=0,gputouse=0):

[kenziew]

We tried that too. It's just with analyze_videos. I can train 2 networks simultaneously but when I have one network running and try to analyze_videos on the second GPU I get a device CUDA:0 not supported by XLA service error

When only one GPU is installed, analyze_videos will only utilize one GPU. But when the second GPU is installed, analyze_videos tries to utilize both even when I specify which GPU using gputuose so then I get the error

[AlexEMG]

Thanks for all the details. I looked into the code again and noticed that the environment variable is set after the TF session is initialized (for the predict code). I just swapped the order and updated the github repo (not pypi so far). Could you please download 2.0.6.3 and check if it works now for you?

[kenziew]

that worked! thank you!

[AlexEMG]

Version released: https://pypi.org/project/deeplabcut/2.0.6.3/