A YOLO that learns one representation for multiple tasks

What it does

YOLOR is a PyTorch implementation of the paper You Only Learn One Representation: Unified Network for Multiple Tasks. It provides a family of scaled object detectors—from lightweight CSP variants up to the heavy D6—along with pretrained weights and exhaustive COCO benchmark tables covering AP scores and batch throughput.

The interesting bit

The project’s central thesis is that a single learned representation can serve multiple vision tasks, though the README only demonstrates object detection on COCO. It treats the idea as a natural evolution of the YOLOv4-CSP lineage, offering six scaled variants that trade resolution and capacity for speed.

Key highlights

• Six model scales (CSP, CSP-X, P6, W6, E6, D6) with input resolutions of 640 or 1280.
• COCO test-dev AP up to 58.2% on YOLOR-D6, with batch-1 throughput ranging from 106 fps (CSP) down to 34 fps (D6).
• Pretrained weights released for CSP and CSP-X variants; a dedicated paper branch preserves the exact checkpoints used for the published tables.
• Training schedules include 300-epoch from-scratch runs and a three-stage fine-tuning pipeline demonstrated on the P6 variant.

Caveats

• The README is rich with benchmark tables and setup instructions but offers almost no architectural explanation of how the “unified representation” works or what the additional tasks are.
• Sections labeled “Developing…” and “Train from scratch…” suggest some results are still in flux rather than finalized.
• A typo in the figure path (unifued_network.png) and sparse documentation signal that this is a research artifact first and a polished product second.

Verdict

Grab it if you want a well-benchmarked YOLO detector with clear speed/accuracy trade-offs and released CSP weights. Look elsewhere if you need a documented multi-task framework or an actively maintained general-purpose vision library.