nightwalker 1.0.1

Neural Network Supervision Platform

nightwalker

nightwalker is a Python based library, that helps Deep / Machine Learning specialists to track their models during training without accessing server, and getting notifications full of their desired information via nightwalker App.

PROS

• Easy to install and use

• Takes Steps, Epochs, Train & Test Loss, Acc as arguments so you can have all useful metrics in your app

• No registration required, just take the unique code after walker.UID generates it for you, and share it with any number of co-workers to add the model to their apps

Used Frameworks & Libraries

AFS is built totally on Python & Node.JS.

• Python 3 - for library building

• Node.JS - for Back-End

Installation

Python 3.6+ required to use.

Get the package from PyPi

$ pip install nightwalker

Usage

Import the nightwalker and reach 'walker' function.

Define the walker function inside the training loop, and pass the arguments.

$ from nightwalker import walker



$ walker(arg1, arg2)

After executing the training loop, the library will generate and print unique id for you, by which you'll then verify your session within the app.

It'll look like this:

$ Your unique ID is --- 1sy45kd9

Add new model to your app using the generated unique id.

Arguments

$iteration argument is for counting iterations. type = number.




$itercount argument is basically a divider, for every how many iterations do you need to send the metric update. type = number.




$maxiter is a maximum of iterations, after which the model finishes training. type = number.



$epoch counts epochs. type = number.




$maxepoch is a maximum number of epochs, after which nightwalker will decide the model had finished training.



$testloss takes test loss as a metric. type = number.




$trainloss takes validation loss as a metric. type = number.




$acc takes accuracy as a metric. type = number.

Suggestions

In case you have any suggestions how to improve the app and library functionalities for you, please feel free to open new Issue, or contact us at [e-mail]

License

---

BSD 3-Clause Licence