aimrun

aimrun

A simple interface for integrating aim into MLOps frameworks.

Installation

pip install aimrun

Features

1. Multiple runs. Simply initialize multiple aimruns using aimrun.init(), and track to multiple repositories at once.
2. No need for a main-process wrapper. You do not need to make sure that only the main-process calls aimrun functions - we take care of that for you.
3. Sync project-specific (local) repositories to larger (remote) repositories. See python -m aimrun sync --help for guidance.

Usage (Recommended)

1. Initialize one or more aimruns using aimrun.init()
2. Use aimrun.track() to track values. Parse a dictionary.
3. Use aimrun.close() to finalize the experiments.

Example usage

import aimrun

# initalize 
aimrun.init(repo='aim://172.3.66.145:53800', experiment='my_experiment', description='description of run' args={"arg": 1}) # args=vars(args) if you use argsparse

# track 
aimrun.track({"value_0": A, "value_1": B})
# or 
aimrun.track(A, name="value_0")

# close 
aimrun.close() 

Synchronizing on-going runs

aimrun.init(repo=".", sync_repo='aim://172.3.66.145:53800', sync_args={"repeat": 60}, experiment='my_experiment', description='description of run' args={"arg": 1})

This starts a thread that incrementally synchronizes the current on-going run to a remote repo while using the current directory as the local repository.

To profit from mass updates (faster synchronization), consider installing an improved aim version:

pip install git+https://github.com/schneiderkamplab/aim

Drop-in replacement Wandb (Experimental)

We experimentally offer aimrun as a drop-in replacement for wandb, making a seamless integration in your framework even easier.

1. Replaceimport wandb with from aimrun import wandb
2. Set default repository before init (e.g. right after import) wandb.set_default_repo('aim://172.3.66.145:53800')
3. Supported functions .init(), .log(), .finish()