#!/usr/bin/zsh ############################################## ##### Batch script for the MRCNN training #### ############################################## #### CREATE SBATCH ENTRIES #### #### Paths and parameters must be adapted accordingly. #### job name #SBATCH --job-name=<JobName> #### Path and name of the output file of the job execution #SBATCH --output=/home/<UserID>/.../<JobOutputFolderName>/%x_%J_output.txt #### Job runtime determined by testing jobs on the GPU node (see manual). #### Multiply the computing time per epoch resulting from the test by the number of epochs to be trained. #### Add a safety factor, e.g. multiply with 1.2 #SBATCH --time=0-00:00:00 #### Memory requirement per GPU determined by testing jobs on the GPU node (see manual). #### Add a safety factor, e.g. multiply with 1.2. #### For example: resulting value is 5GB --> --mem-per-gpu=5G #SBATCH --mem-per-gpu=5G #### E-mail address #SBATCH --mail-user=<EmailAdress> #### E-mails to be received #SBATCH --mail-type=ALL #### Number of tasks to be performed #SBATCH --ntasks=1 #### Number of GPUs required per node #SBATCH --gres=gpu:1 #### Definition of the job array starting at 0. ### #### This parameter is only required if you want to perform several jobs in parallel #### from one job script, e.g. grid search via Weights and Biases sweep. #### In this example we perform a grid search with 6 jobs --> array=0-5 #SBATCH --array=0-5 #### CREATE TERMINAL ENTRIES #### #### Paths and parameters must be adapted accordingly #### Loading the Cuda module module load cuda/10.0 #### Export path in which Anaconda is located export PATH=$PATH:/home/<UserID>/anaconda3/bin #### Activate environment source activate env_mrcnn_gpu #### Navigate to the path where the droplet.py script is located cd /home/<UserID>/.../samples/droplet/ #### Run MRCNN via Weights and Biases. #### The <SweepCode> is generated after a sweep is created at the Weights and Biases homepage. #### All training parameters are specified in the sweep configuration. wandb agent --count 1 avt-droplet-detection/paper/<SweepCode>