nebulae 0.6.27

A deep learning framework based on PyTorch and prevalent image processing libraries. It aims to offer a set of useful tools and functions.

Nebulae

A deep learning framework based on PyTorch and prevalent image processing libraries
It aims to offer a set of useful tools and functions.

---

🚀 Spotlight

Several frequently used features are integrated with simple interfaces.

🛠️ Integrated Feats - EMA module, multi-source datasets, training log plotting and timer etc.

🎯 Simplified API - Unified API for distributed and single-GPU training.

⚡️ Efficiency - Data augmentations are reimplemented using Numpy which is faster than PIL.

🧩 High Compatibility - Users are able to build networks using Nebulae with PyTorch seamlessly.

---

⚡ Quick Start

📸 Utility

Obtain GPU stats after some training epochs.

import nebulae as neb
from nebulae import *

gu = kit.GPUtil()
gu.monitor()
for epoch in range(10):
  # --- training code --- #
gu.status()

Automatically select unoccupied GPUs. It is useful for a shared machine.

import nebulae as neb
from nebulae import *
# select 4 GPUs with 2GB or more memory left
engine = npe(device=power.GPU, ngpu=4, least_mem=2048)

Find entire distributed training and test code in ./examples/demo_core.py

---

📦 Installation

Users can install nebulae from pip

pip install nebulae

For better development, building from Dockerfile is also available. Modifying the libs version and have nvidia-docker on your machine is recommended.

sudo docker build -t nebulae:std -f Dockerfile.std .
sudo docker run -it --gpus all --ipc=host --ulimit memlock=-1 nebulae:std

The latest version supports PyTorch1.6 and above

---

❤️ Support

If you find Nebulae helpful, consider giving it a ⭐ on GitHub! ▶️ https://github.com/SeriaQ/Nebulae