cox package

Cox is a lightweight, serverless framework for designing and managing experiments. Inspired by our own struggles with ad-hoc filesystem-based experiment collection, and our inability to use heavy-duty frameworks, Cox aims to be a minimal burden while inducing more organization. Created by Logan Engstrom and Andrew Ilyas.

Cox works by helping you easily log, collect, and analyze experimental results.

Why “Cox”? (Aside): The name Cox draws both from Coxswain, the person in charge of steering the boat in a rowing crew, and from the name of Gertrude Cox, a pioneer of experimental design.

Quick Logging Overview

The cox logging system is designed for dealing with repeated experiments. The user defines schemas for Pandas dataframes that contain all the data necessary for each experiment instance. Each experiment ran corresponds to a data store, and each specified dataframe from above corresponds to a table within this store. The experiment stores are organized within the same directory. Cox has a number of utilities for running and collecting data from experiments of this nature.

Walkthroughs

• Walkthrough 1: Logging and Reading Data
  • Initializing stores
  • Creating tables
    • A note on serialization
  • Logging
    • Incremental updates with update_row
  • Reading data
  • CollectionReader: Reading many experiments at once
• Walkthrough 2: Using cox with tensorboardX
  • Viewing multiple tensorboards with cox.tensorboard_view

Submodules

• cox.readers module
• cox.store module
• cox.tensorboard_view module
• cox.utils module