YOLOv4 breaks out of Darknet into PyTorch

What it does

This is a minimal PyTorch re-implementation of the YOLOv4 object detector. It handles training—with Mosaic augmentation—and inference, and it bundles tools to port original Darknet weights into PyTorch, then export them to ONNX and TensorRT. You can also feed it non-square input resolutions, provided they follow the 96-pixel stride formula starting at 320.

The interesting bit

The real value is the conversion fidelity. The README shows COCO benchmarks where this PyTorch port lands within a hair’s breadth of the original Darknet model—0.466 AP versus 0.471 at 416×416—and the TensorRT exports actually edge it out at 0.472. That means you are not sacrificing accuracy to escape the Darknet ecosystem.

Key highlights

• Near-parity COCO accuracy: PyTorch AP 0.466 vs Darknet 0.471 at 416×416; TensorRT FP32 hits 0.472.
• Full conversion chain: Darknet weights → PyTorch → ONNX → TensorRT, plus an ONNX-to-TensorFlow path.
• Supports non-square inference sizes via a 96-pixel stride formula starting at 320.
• Training includes Mosaic augmentation.
• Dynamic or static batch-size ONNX exports.

Caveats

• Post-processing NMS is still required after inference, including for TensorRT outputs.
• The TensorRT demo script only handles batch size 1 or dynamic ranges including 1 without modification.
• The ONNX-to-TensorFlow path is reported to be brittle with standard package installs and requires manual source compilation.

Verdict

Worth a look if you need YOLOv4 inside the PyTorch or ONNX deployment stack. Skip it if you are already comfortable in Darknet or need a turn-key pipeline with built-in NMS.