/neptune_logger

Sample logger class to introduce Neptune.ai

Primary LanguagePython

Simple neptune.ai Logger

This repository provides simple but useful python class to manage experiments using neptune.ai. neptune.ai is a simple but powerful experiment management service. In this repository, I provide a simple logger class to log metrics, images, files, artifacts and hyper-parameters for any experiment.
A blog post I wrote about this can be found here. (Written in Korean)

How-to-Use

NeptuneLogger is implemented with python and neptune-client. Therefore, it can be intagrated with any project or experiment you would like to do. Detailed and step-by-step starter can be found here.

Install neptune-client

You can install neptune using pip. (it's not neptune, but neptune-client)
pip install neptune-client

If you want for neptune to inspect your hardware resource, also install psutil
pip install neptune-client psutil

If you want to run the sample MNIST codes with neptune, you have to install:

  • torch
  • torchvision
  • matplotlib

Prepare your NEPTUNE_API_TOKEN

Once you sign and log in to neptune, you will get the unique token for API.
If you don't know or remember, try to follow 'Getting Started' in upper right corner of the neptune page.

You can specify API token directly in the code, but it's unsafe.
I recommend setting API token as an environment variable.
For example, on linux you can add it in ~/.bashrc or use command as below:
export NEPTUNE_API_TOKEN = YOUR_API_TOKEN

Log whatever you like!

Define logger using NeptuneLogger and log! To initiate NeptuneLogger, there are many arguments to be specified.

  • api_key: Your API token
  • project_name: Project name with the format of YOUR_ID/PROJECT_NAME. This can be found in your main neptune page when you log in.
  • experiment_name: Arbitrary name of the experiment (e.g. SampleMNIST)
  • description: Short description of the experiment
  • tags: List of tags you want to specify. Tags are useful to filter your experiments of the project!
  • hparams: Hyper-parameters as python dictionary.
  • upload_source_files: Any source code files to upload.
  • hostname: Hostname. Useful when you experiment with different devices or servers.
  • offline: True/False. Log online neptune page if True. Do not log online otherwise.

Run sample MNIST experiment

For those who want to try out how neptune works, I provide simple MNIST classification experiment. In main.py, modify arguments for NeptuneLogger with your own. Then, run it and go neptune page to see what happens.

You will see metrics, plots, hyper-parameters are logged succesfully!