Clarifying DLC usage with TF2+

Hello all!

I just want to clarify the usage of TF2+ with DLC. The original blog post states, that only DLC 2.1.8.2 is up to date with TF2+, but that a project created in DLC 2.1.9 is fine. Would this scenario work then:

  1. Install DLC 2.1.9 with the supplied conda environment
  2. Within this environment install dlc-core: pip install git+https://github.com/DeepLabCut/DeepLabCut-core.git@tf2.2alpha
  3. and the latest tensorflow (I have CUDA11 installed): pip install tensorflow==2.4
  4. Create a project and label videos with DLC 2.1.9 (import deeplabcut as dlc)
  5. Create a training dataset, train, evaluate and analyze new videos with dlc-core (import deeplabcut-core as dlc)
  6. For refining and extracting outliers, go back to DLC 2.1.9 (import deeplabcut as dlc)
  7. Repeat

Additionally, the DLC website states that only CUDA10 or lower is supported. I have already been running DLC 2.1 and DLC 2.2b8 with tensorflow 1.13.1 and CUDA 11 and I seemed to have not experienced any issues. Am I missing something and should roll down my CUDA to 10? CUDA 11 was automatically installed as part of my NVIDIA GPU driver update.

I thank you kindly for your awesome support!

that works! :slight_smile: