Very 1st version

Forge

A front-end supported ML/DL tracking tool

This tool can be considered as 2 parts:

• Code API
• Interactive WebUI

So a troubled data scientist can inject some lines of code in his experiments/production. And view tracked result/ trained latent space vectors (Later version) related to the hyper-parameters he tried on the model(s).

Code API (forgebox)

Upon this version the code API has changed to the following features:

• Most of the training configs/metrics/weights records/ are stored in a SQL db (could be most other popular SQL db, the default is SQLite)
• The code API was designed to minimize the coding effort from user.
• For now, many enhanced functionalities are only for PyTorch
  • train: Managing iteration for PyTorch, while allow liberty on what happens within on step
  • tracker: Easily tracking the Hparams, metrics, weights paths
  • layers: Some self-defined PyTorch modules, like Unnormalize for unnormalize Imagenet color range, AttenionLSTM for attention augmented RNN.
  • prepro: PyTorch datasets for preprocessing data
  • metrics: Some frequently used metric functions

WebUI(forge)

The WebUI has CRUD(Create, Read, Updated, Delete) view on the following information

• Train Tasks
• Configs (HyperParameters)
• Metrics (Model Performance)
• Location of model weights

The WebUI was based on flask app-builder, Hence the following features are already given

• Simple CRUD view, associated with paths and menu
• User/Auth management available, and other api login methods available (OAuth, PEM, LDAP)
• Most SQL databases' support available, as long as you change the URI.