ddpw 4.0.0

A utility package to scaffold PyTorch's DDP

DDPW

---

Distributed Data Parallel Wrapper (DDPW) is created as a utility package to encapsulate the scaffolding for PyTorch's Distributed Data Parallel.

This code is written in Python. The DDPW documentation contains details on how to use this package.

Overview

Installation

conda install -c tvsujal ddpw # with conda
pip install ddpw # with pip from PyPI

Usage

from ddpw.platform import Platform, PlatformConfig
from ddpw.artefacts import ArtefactsConfig
from ddpw.job import JobConfig, JobMode
from ddpw.wrapper import Wrapper

from src import MyDataset, MyModel, MyOptimiser, MyTrainer

# datasets
train_set = MyDataset(train=True)
test_set = MyDataset(train=False)

# configure the platform
p_config = PlatformConfig(platform=Platform.GPU, n_gpus=4, cpus_per_task=2)

# configure the artefacts (model, dataset, optimiser, etc.)
a_config = ArtefactsConfig(train_set=train_set, test_set=test_set,
    batch_size=64, model=MyModel(), optimiser_loader=MyOptimiser(lr=0.1))

# call the job
Wrapper(p_config, a_config).start(MyTrainer())