Skip to main content

NeptuneML wrapper to quickly record data about experiments

Project description


Quickly track your Jupyter Notebook experiments with NeptuneML

fast-neptune is a library that helps you quickly record all the information you need to launch your experiments, when you are using Jupyter Notebooks. Reproducibility has become a crucial issue in Machine Learning, not only for research, but also for real world applications, where we want to have robust results, and track every set of parameters tested, along with their results.

fast-neptune spirit is highly inspired from nbdev in the user experience, so it is recommended to have a look first at it.

Furthermore, fast-neptune is built upon the solution neptune-ml which allows users to quickly record on a public or private repository. Following the quick introduction is a plus to understand fast-neptune, but as we will see it's quite intuitive.


fast-neptune is available on pypi so you can simply run :

pip install fast-neptune

How to use

fast-neptune has implemend several features to help you when you run ML experiments. These include :

  • metadata about the machine where the code is run, including OS, and OS version
  • requirements of the notebook where the experiments are run
  • parameters used during the experience, which means the names of the values of the variables you want to track
  • code you used during the run that you want to record

Note that code and parameters are not mandatorily tracked, while the two former are, though the whole purpose of this library is to facilitate the tracking of parameters and code.


Metadata is tracked about the requirements used and information about the Python version and OS used.

They are added automatically when an experiment is created, through the functions get_metadata and create_requirements.


Properties refer to variables you want to record.

To record properties, simply add #property to each cell containing the variables you want to record.

Note : if one variable in a property is a file encapsulated in a Path object, it is automatically tracked, and this file will be sent to NeptuneML. You can disable this option when creating the experiment.


Code cells can also be tracked.

To do so, simply add #code to each cell you want to record. You have the possibility to specify under what name the cell code will be registered. By default, all cells with #code will be added chronogically to the file "", but if you can put them in specific files by adding the name of the module you want them to be put in.

Example : add #code to the cells that take care of the data loading.

Example with fastai vision to train a MNIST classifier

To understand more how all of this fit in, we will follow a simple example, using fastai library.

First we will log in using NeptuneML

# Neptune login
from neptune.sessions import Session
import getpass

# First we get the token
api_token = getpass.getpass("Please enter your NeptuneML API token : ")
session = Session(api_token=api_token)

# Then we enter user name and project
user_name = "danywin"
project_name = "fast-neptune"

project = session.get_project(project_qualified_name=f'{user_name}/{project_name}')
Please enter your NeptuneML API token : 

WARNING: Instantiating Session without specifying a backend is deprecated and will be removed in future versions. For current behaviour use `neptune.init(...)` or `Session.with_default_backend(...)
WARNING: It is not secure to place API token in your source code. You should treat it as a password to your account. It is strongly recommended to use NEPTUNE_API_TOKEN environment variable instead. Remember not to upload source file with API token to any public repository.

Now we will start using fast-neptune.

First we will record the code used to load the data and learner. To do so, we simply add #code to the cell used to load the data.

from fast_neptune.core import *
from import *

path = untar_data(URLs.MNIST_SAMPLE)
data = ImageDataBunch.from_folder(path)
learn = cnn_learner(data, models.resnet18, metrics=accuracy)

Then we want to keep track of the number of epochs during training, and the learning rate we used. To do so we add #property to the cell we want to track.

n_epoch = 1
lr = 1e-2

Finally we use fast_experiment to create our experiment. Note that you must pass it the name of the notebook you are using for your experiments, and the global variables you are currently using to record each variable in the properties cells.

Here we will create an experiment, then record the validation accuracy.

nb_name = "index.ipynb"
globs = globals()

with fast_experiment(project,nb_name,globs) as exp:
    pred,y = learn.get_preds()

    score = accuracy(pred,y).item()
epoch train_loss valid_loss accuracy time
0 0.052504 0.014153 0.994603 00:04

Now if we go on the link provided by NeptuneML we can have a closer look at this experiment :

Here we can see the parameters that were registered, including epoch, learning rate, OS, and Python version.


In the source code tab, we can have a look at the bits of code we registered, here the dataloading. Code

In the artifacts we also have the modules used in this notebook with their version :


Finally, we logged the validation accuracy, which can be found in the Logs tab :


Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

fast_neptune-0.0.3.tar.gz (13.3 kB view hashes)

Uploaded source

Built Distribution

fast_neptune-0.0.3-py3-none-any.whl (7.8 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page