Releasing a New YOLOv3 Implementation in PyTorch

**Note: YOLOv4 was released recently. You may also want to consider visiting our post on how to train YOLOv4 in Darknet.

At Roboflow, we're constantly adapting our product to make it as easy as possible for users to create custom computer vision models on high quality data. While we have an implementation of YOLOv3 in TensorFlow (and Keras), we continuously heard requests for support of YOLOv3 in PyTorch.

Today, we're introducing support for a PyTorch implementation of YOLOv3, originally introduced by the talented team at Ultralytics.

To find our implementation, navigate to our model library via https://docs.roboflow.ai, or direct Colab link here.

In our initial tests, we found this YOLOv3 implementation to be even more performant than our last. In fact, in 300 training epochs on Google Colab lasting slightly over an hour the model achieved 0.93+ recall and 0.978 mAP@50 on our (challenging) 12-class chess piece identification task.

If you are currently using our YOLOv3 TensorFlow implementation, we recommend attempting your results with the new notebook.

Results from our YOLOv3 testing on our chess dataset.

User Guide: Training a Custom PyTorch Model

If you're an existing Roboflow user with an uploaded object detection dataset, training a model with our PyTorch implementation is trivial.

Navigate to the dataset version you seek to use. When exporting, select "YOLO Darknet Weights," copy your download code Roboflow generates for you, and paste them into the PyTorch notebook at the cell marked # REPLACE THIS LINK WITH YOUR OWN.

If you're new to Roboflow but want to make use of this implementation, getting stated is easy. Roboflow is free for small datasets. Follow our Getting Started Guide or jump right into https://app.roboflow.ai and follow the guided onboarding tutorial.


Want to be the first to know about new computer vision tutorials and content like our synthetic dataset creation guide? Subscribe to our updates 📬.

Roboflow accelerates your computer vision workflow through automated annotation quality assurance, universal annotation format conversion (like PASCAL VOC XML to COCO JSON and creating TFRecords ), team sharing and versioning, and easy integration with popular open source computer vision models. Getting started with your first 1GB of data is completely free.

Show Comments