Skip to content

Latest commit

 

History

History
65 lines (41 loc) · 3.16 KB

pose_on_rtx30x.md

File metadata and controls

65 lines (41 loc) · 3.16 KB

Pose estimation on newer GPUs

With the newer GPUs (e.g., NVIDEA Rtx3080 ) you can no longer use Tensorflow (TF) version 1.x, which is what earlier versions of DeepLabCut depend on. Instead, you have to use TF version 2.x, and this requires a different set of package versions than previously. Here is our protocol for getting TF and DLC to work on our Microsoft Windows NVIDEA Rtx3080 supported machines.

1. Download and install CUDA version 11.0 and CUDA version 11.1 for Windows
2. Download cudnn version 8.0.4.30 for windows
3. Install cudnn in the CUDA 11.0 version

Note: The CUDA default installations directories on Windows is C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0 and C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.1, respectively.

4. Add the CUDA bin folder to the windows environment paths.

Note: The Windows environment paths may look something like this:

5. Create a new conda environment – python 3.6 should work. In your conda environment, install the following python packages:

  • tensorflow-gpu==2.4

  • keras==2.3.1

  • scipy==1.2

  • protobuf==3.6.0

  • tf_slim==1.1.0

  • tf-nightly==2.5.0.dev20210128

Alternatively copy past this in the terminal:

pip install tensorflow-gpu==2.4 keras==2.3.1 scipy==1.2 protobuf==3.6.0 tf_slim==1.1.0 tf-nightly== 2.5.0.dev20210128

6. In your conda environment, install deeplabcutcore with the below command. For more information, see the DeepLabCut documentation:

pip install git+https://github.com/DeepLabCut/[email protected]

7. In your CUDA 11.1 folder, locate ptxas.exe. Copy this file and use it to replace ptxas.exe in your CUDA 11.0 version.
8. At the time of writing, I don’t think deeplabcutcore has a graphical interface, so we have to run it by the command line. We run the DLC model with the script like this. It can be necessary to specify the per_process_gpu_memory_fraction argument as less than 1, as the default settings may consume all memory and kill the program.
import deeplabcutcore as deeplabcut
import tensorflow as tf
import sys

config = tf.compat.v1.ConfigProto(gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=0.8))
config.gpu_options.allow_growth = True
session = tf.compat.v1.Session(config=config)
tf.compat.v1.keras.backend.set_session(session)

def perform_pose(project_path, videos_path):
    deeplabcut.analyze_videos(project_path, videos_path, videotype='mp4', shuffle=1, save_as_csv=True)

if __name__ == "__main__":
    project_path = DLC PROJECT CONFIG YAML PATH E.G r"C:\Users\Windows\Desktop\Project1-SN-2020-04-07\config.yaml"
    videos_path = LIST CONTAINING STRING PATH TO THE FOLDER WHERE THE VIDEOS TO ANALYZE ARE LOCATED  E.G [“C:\Users\Windows\Desktop\MyVideos”]
    perform_pose(project_path, videos_path)
    print('Video analysis complete')

Author Simon N, JJ Choong