Research-oriented — flexible interfaces and dynamic modules

Latency-oriented — efficient implementation and various types of timing support

Features:

• Training auto-resume after interruption
• Support for DataParellel & DistributedDataParallel
• Tensorboard integration

Supported tasks:

• Image classification
• Object detection

Notable models:

• ResNet-18 on CIFAR-10 (Top-1 95.62%)
• ResNet-50 on ImageNet (Top-1 76.56%, Top-5 93.02%)
• Faster R-CNN on COCO (AP 37.0, through torchvision)

Note that this repo is currently under active development and compatibility is not guaranteed. While there are many standard techniques for faster implementations, this repo mostly focuses on the research aspect of reducing training and inference cost.

Install PyTorch (>= 1.1) and run:

pip install llcv

If you prefer to install it in develop mode:

git clone https://github.com/mtli/llcv
cd llcv
pip install -e .

Check out the scripts in the samples/ for various tasks.

To use this framework for your own projects, you can either fork this repo, or using this repo as a dependency through dynamic modules.

You can find more details about the framework's design here.

Note that latency and throughput are measured on Geforce GTX 1080 Ti GPUs, unless otherwise stated.

In the above table, timings are measured on a single GPU for both training and testing.

In the above table, timings are measured on 4 GPUs for training and a single GPU for testing.

In the above table, timings are measured on 8 GPUs for training and a single GPU for testing.