LSENS-DeepLabCut

LSENS repository for deeplabcut model training in SCITAS clusters and preprocessing scripts Get SCITAS access: Talk to Carl Set up cluster:

1. Open Anaconda Prompt. Type: ssh <gaspar id>@<cluster>.epfl.ch

• Cluster can be izar, jed, or helvetios for scitas clusters
• You’ll be prompted to type your gaspar password
• You will be dropped in your /home folder. This is to have “permanent” files with info or .sbatch files for running your routines.

2. Mount LSENS server:

Create servers folder -> type:

mkdir servers

Open bashrc file:

vim .bashrc

Press i to be able to edit bashrc file In the last line of the bashrc file copy paste (paste in command line is done by right clicking):

gio mount smb://intranet\;<gaspar_id>@sv-nas1.rcp.epfl.ch/Petersen-lab
ln -s  /run/user/$(id -u)/gvfs/smb-share\:domain\=intranet\,server\=sv-nas1.rcp.epfl.ch\,share\=petersen-lab\,user\=<gaspar_id>/* ~/servers

and replace <gaspar_id> appropriately.

Press ESC and then type “:wq” to save and quit (you can use just “:q” to quit without saving)

3. Load necessary modules: Type:

module load gcc python openmpi py-tensorflow

4. Create virtual environment: Type:

mkdir venvs
virtualenv --system-site-packages venvs/DLC

5. Activate DLC virtualenv and install deeplabcut: Type:

source venvs/DLC/bin/activate
python -m pip install --no-cache-dir deeplabcut==2.3.9
python -m pip install --no-cache-dir typing-extensions==4.6
python -m pip install --no-cache-dir keras==2.10

6. Test DLC installation a. In command line open python: type: python, then

import deeplabcut

7. Set up file transfer: a. Install WinSCP in your local computer b. Create folder in your /home/<gaspar_id> directory for the files you want to run: you will need a folder with your code (i.e this repo), and a logs folder. c. Create folder in your /scratch/izar/<gaspar_id> directory for the videos (videos_to_anly) and for the results (dlc_results). Copy your network on the scratch folder too.

To train your network:

1. Create and label your project in your local computer (no gpu needed but make sure the deeplabcut version is the same).
2. Change the path in your config.yaml to the correct cluster path and transfer your network to the cluster.
3. In terminal:

source venvs/DLC/bin/activate

a. Open a python instance typing python and import deeplabcut. Pointing the network to your config path (somehting like /scratch/izar/<gaspar_id>//config.yaml), create training dataset: dlc.create_training_dataset(config_path). This is important to do already on the cluster since otherwise you will have path issues. Close python. b. Edit the train.sbatch file in the repo with the path to your network config.yaml file. You can type in terminal: vim /home/<gaspar_id>/LSENS-DeepLabCut/train.sbatch, press "i" to insert and change the name of the job, the max time of the job (12 h should be more than enough) and the email so you receive notifications if your code finishes or crashes. Exit vim by pressing ESC, typing :wq to save or :q! to not save c. Run your training by typing in terminal:

sbatch train.sbatch

To analyze videos:

1. Create a .json file with the make_json_config.py file. You can do this on your local computer. Note that you may have to tweak a few parameters like your config_path (in cluster).
2. Copy the json file to the json_files folder in your server.
3. Activate environment:

source venvs/DLC/bin/activate

5. Run analysis:

python /home/$(whoami)/LSENS-DeepLabCut/data_transfer_and_run.py $(whoami) <name of json file>

6. After everything ran, transfer data and delete videos from scratch folder by using the command:

python /home/$(whoami)/LSENS-DeepLabCut/transfer_results.py $(whoami) <name of json file>

Enjoy and may the pose estimation gods be with you!

Troubleshooting tips

• When installing DLC in your virtual env, you get the following error: TypeError: canonicalize_version() got an unexpected keyword argument 'strip_trailing_zero' This means you have to either downgrade or upgrade the packaging module e.g. >22.0 or higher.