eniat 0.0.3

ENIAT supports and boosts your machine learning experiments!

ENIAT

Eniat is a Python template for various ML packages including PyTorch, Scikit-learn.

Currently supports only PyTorch :moyai:

It provides several convenient features

• automated training: you don't have to re-write same epoch-wise loop everytime!
• automated distributed learning for PyTorch.
• Easy configurations for your experiments.
• Various package support (in the future)

Quick Start

1. Installation

Install eniat with the below command.

pip install eniat

eniat installation doesn't convey ML packages such as torch or sklearn, so install one you need before using eniat. :moyai:

2. CLI execution

You can initiate any tasks with only console commands. :moyai:

With below command, you can conduct ML experiemnts on console provided that you have proper resources for models and data. These configurations except basic arguments are based on omegaconf, so refer to the documentation and provide valid arguments.

eniat -p PACKAGE -t TASK --OTHER_CONFIGS_YOU_NEED

Or you can simply write every configurations you need into one yaml and just type the path of it like below. I strongly recommend this, because ML experiments usually require quite large amount of parameters. Make a base configuration yaml file, and modify slightly by typing additional configurations into console command, and it will be much more convenient.

eniat -c=PATH_TO_CONFIG_FILE

3. Custom code execution

eniat is designed for various environments from mere .py codes to jupyter notebooks.

The most fundamental component of eniat is Trainer. It is in charge of training your models. In the other hand, Grader only evaluates your model. Below is an example code of torch training wih eniat.

from eniat.utils.statelogger import DummyLogger
from eniat.torch import TorchTrainer, TorchGrader

logger = DummyLogger() # Default logger for Eniat
grader = TorchGrader(grader_cfg, logger = logger) # If none, train without evaluation
trainer = TorchTrainer(trainer_cfg, learner_cfg, data_cfg, logger, grader)

trainer.fit(device = 0) # training on dev 0