A bridge for YOLOv4: Darknet weights to TF, Lite, and TensorRT

What it does This repository re-implements YOLOv4, YOLOv4-tiny, YOLOv3, and YOLOv3-tiny in TensorFlow 2. Its primary purpose is ingesting the original Darknet .weights format and exporting it to TensorFlow SavedModel, quantized TensorFlow Lite, and TensorRT engines. It also includes training scripts and COCO 2017 evaluation, though the author notes that reproducing full training performance is still a work in progress.

The interesting bit The real value is the plumbing, not the port. The project treats Darknet as the training source of truth and provides a one-way pipeline to modern deployment targets, complete with int8 quantization for mobile and FP16 acceleration for NVIDIA GPUs. It even publishes throughput tables across Tesla cards and TensorRT batch sizes so you can see the trade-offs before flashing a device.

Key highlights

• Converts YOLOv4, YOLOv4-tiny, YOLOv3, and YOLOv3-tiny .weights to TensorFlow, TFLite, and TensorRT formats.
• Supports TFLite quantization in float16 and int8 for edge and mobile deployment.
• Includes COCO 2017 evaluation and mAP50 benchmarks comparing YOLOv3 and YOLOv4.
• Publishes GPU throughput numbers for Tesla P100, K80, T4, and P4, plus TensorRT batch benchmarks.

Caveats

• YOLOv4 and YOLOv4-tiny int8 quantization currently have issues; the author suggests using YOLOv3 variants for int8 instead.
• Training performance is not fully reproduced, and the README explicitly recommends training in AlexeyAB’s Darknet and converting the weights afterward.
• The Macbook Pro CPU benchmark table is present but empty.

Verdict Grab this if you need to deploy existing Darknet-trained YOLO models to TensorFlow ecosystems or edge hardware. Skip it if you want a fully self-contained training suite or flawless int8 quantization on YOLOv4.