mrunner is tested on python2.7 and python3.5, but newer versions of python3 shall also work. Additionally we recommend to install and use mrunner in virtualenv.

1. To install it use following commands:

shell pip install neptune-cli==1.6 pip install git+ssh://git@pascal-tower01.intra.codilime.com/ml-robotics/mrunner.git@develop

Above sequence is related with neptune issues.

1. Set some remote contexts and select active one
2. Configure clients for at least one system:
3. slurm
4. kubernetes

Remote context

To avoid passing details on configuration of computation system, it is possible to store predefined configuration parameters. Each configuration is named and can be selected during experiment start or set by default:

mrunner run foo.py -- --param1 2    # run with context defined in current_context config key
mrunner --context plgrid.agents run foo.py -- --param1 2
mrunner --context plgrid.agents --config mrunner.yaml run foo.py -- --param1 2   # loads configuration from local file (instead of user configuration directory)

Set of keys depends on type of remote context. For description of available keys go to proper sections (ex. slurm, kubernetes).

To manage contexts use context command. Example calls:

mrunner context
mrunner context add --name gke.sandbox --backend_type kubernetes \
                    --registry_url https://gcr.io --storage /storage
                    --resources "tpu=1 cpu=4"
mrunner context edit gke.sandbox               # opens editor with context parameters
mrunner context set-active gke.sandbox
mrunner context delete gke.sandbox
mrunner context copy gke.sandbox gke.new
mrunner --config mrunner.yaml context set-active gke.sandbox

Example remote contexts':

name: gke.sandbox
type: kubernetes
registry_url: https://gcr.io
resources:
  cpu: 4
  tpu: 1
neptune: true
storage: /storage