To get started with Atlas, activate the virtual environment (if any) that was used to install Atlas server.
Next, let's make sure Atlas Server is running by doing:
By default Atlas runs without using your GPUs.
To run GPU jobs start Atlas with
atlas-server start -g
This will bring up all of the Atlas services, including the GUI which is available on your instances 5555 port, (e.g.
<ipv4_of_your_remote_instance>:5555) by default.
Modes of Operation¶
Atlas runs in two modes.
In execution mode, code is run in an existing local environment that you already have setup. Atlas provides experiment version control, tracks your experiments and any associated metadata (e.g. hyper-parameters, metrics, artifacts) that you choose to track.
Use this mode for single runs, manual parameter tuning, and small experiments.
In scheduling mode, Atlas queues your experiment with the local Atlas scheduler which runs it in a containerized environment. This mode gives you the ability to queue a large number of experiments as well as leverage common containerized tools such as NVIDIA Rapids.
Use this mode if you want to schedule many jobs and do a large hyperparameter search