The conda environment we use is provided in env.yml. You can clone it using:
conda env create --name msgr-emma -f env.yml
After doing this, you will still need to install the messenger package by following the instructions here
We provide scripts to submit jobs using SLURM, you may need to modify / adapt these for your own systems or if you want to run these interactively. For example to submit jobs on stage 1:
cd training/stage_1
conda activate msgr-emma
bash stage_1.sh
Note: We run EMMA on stage 1 for 12 hours, but this is to make comparisons with baselines that take this long to converge. In general EMMA should converge much faster than this.
You can see the training curves generated using the provided scripts at this wandb project