Skip to content

Latest commit

 

History

History
161 lines (133 loc) · 5.16 KB

README.md

File metadata and controls

161 lines (133 loc) · 5.16 KB
Raw

utils

Introduction

This is a small collection of utilities authored by @joseph-zhong, for the purpose of making Python modules and scripts more fun and simpler to write and work with.

In particular, cmd_line.py parses the arguments of a given function, and uses argparse under the hood and converts the python arguments into cmdline arguments, neatly and programatically determining whether the arguments should be optional/required, their types, and default values. TODO would be to parse the docstring for the following help message!

Table of Contents

Philosophy

Python is a great language for rapidly proto-typing ideas and getting them off the ground. When it comes to building apps, or running scientific experiments, it's hard to find another language with a comparable fluid experimental workflow thanks to the combination of the scientific-library stack (numpy, scipy, torch, tensorflow, jupyter, ...), as well as the simplistic, dynamically typed flexibility offered by python, allowing for rapid iteration for scripts and app-development.

As a result, what ends up happening is that developers will build a multitude of tools to support their own needs, from custom argparse-boilerplate code for spinning up new scripts, to creating their own directory organization systems for interfacing with their experimental datasets, models, results, visualizations, etc.

utils aims to unify this world of productivity

  • We want it to be easy to add code, and create new results.
    • We only have two files, because we want to Keep It Simple and Stupid.
    • utils.py only interfaces with what directories need to exist to organize which results a developer cares about. See more in #organization.
  • We want to simultaneously support a Python and cmd-line API.
    • We want to think as though python is the developer's world in which they develop new algorithms, models, applications, ...
    • But independent of code, one should be able to apply said algorithms, models, applications, ...
      • cmd_line.py takes an existing python API and automatically turns it into a cmd_line interfacing API.

Getting Started

  1. To generalize the useage of utils, we use an environment variable WS_PATH to specify the workspace path housing your project directory, as well as the relevant data.

    Add the following to your ~/.bashrc

    export WS_PATH=/path/to/your/ws/dir

  2. Use parseArgsForClassOrScript(...) to turn a python function into the head for a cmdline script

    For example, a script called train.py for the purpose of training models would be able to then be invoked with the following command

    ./train.py --dataset=mnist --num_epochs=50 --batch_size=128

#!/usr/bin/env python
import src.utils.utility as _util
import src.utils.cmd_line as _cmd

def train(dataset: str, num_epochs: int, batch_size=16, ...):
    ...
    for epoch in range(num_epochs):
        train_step(...)
    ...

def main():
    global _logger
    args = _cmd.parseArgsForClassOrScript(train)
    varsArgs = vars(args)
    verbosity = varsArgs.pop('verbosity', _util.DEFAULT_VERBOSITY)
    _logger.info("Passed arguments: '{}'".format(varsArgs))
    train(**varsArgs)

if __name__ == '__main__':
    main()

TODOs

  • Parse docstring to also produce a help-message
  • Correctly pass the verbosity flag recursively through the pipeline

Organization

Here is an example of how utils could be used in your python project

./src: ..... Root python project source code
  - scripts: Relevant scripts to run particular jobs
    - train.py: Training executable to run a training job
    - demo.py: Demo executable to run a demo 
  - utils: General Utilities
    - utility.py
    - cmd_line.py
  - model
    - object_detection
      - yolo
        - v1
        - v2
        - v3 
    - rnn
    - rnn+attn
    - transformer
      - bert
      - gpt2
    - cnn
      - alexnet
      - resnet50
./data: .... Root data directory to place all relevant datasets, model weights,
    logs, ...
  - datasets: pre-processed training datasets
    - imagenet:
      - 1000
      - 5000
      - 10000
      - 100000
      - all
    - ...
  - raw: pure, un-touched datasets in their raw form
    - imagenet:
      - 1000
      - ...
  - weights: model checkpoints
    - imagenet:
      - model=resnet50
        - 0000
          - weights.pth
  - tb: tensorboard logs
    - imagenet:
      - model=resnet50
        - 0000
          - ...

Acknowledgements

  • Shon Katezenberger for being such an immensely impactful and inspirational mentor over the years, and always driving for succinct, high-functioning, purposeful, and correct code.
  • Ryan Rowe (@rfrowe) for helping me test and debug cmd_line.py over our projects