Libtorch_YOLOv3_train_demo

A libtorch implementation of YOLOv3, supports training on custom dataset, evaluation and detection.

Requirements

1. LibTorch v1.5.0
2. CUDA v10.1
3. OpenCV v3.4.10

Build

cd build
cmake -DCMAKE_PREFIX_PATH=path/to/libtorch -DCMAKE_ECLIPSE_VERSION=4.15 -G "Eclipse CDT4 - Unix Makefiles" ../src

Recommend

Use more "standard" method: torch.jit(TorchScript), and do pre & post process in LibTorch. Note that this implementation does not support multi scale training yet, poor performance on difficult tasks.