README.md

Supervisely

AI for everyone! Neural networks, tools and a library we use in Supervisely.

Introduction • Agent • DTL • Neural Networks • Library • Related

Introduction

Supervisely is a web platform where you can find everything you need to build Deep Learning solutions within a single environment.

We learn a lot from our awesome comunity and want to give something back. Here you can find our python code we use to develop models and tools like DTL and also a source code for agent you deploy on your PC.

Agent

Supervisely Agent is a simple open-sourced task manager available as a Docker image.

Agent connects to Supervisely API and so you can run tasks like import, DTL, training and inference on a connected computer — host.

Internally, we use protobuf for communication with server. Check out documentation on how to deploy a new agent.

DTL

Data Transformation Language allows to automate complicated pipelines of data transformation. Different actions determined by DTL layers may be applied to images and annotations. In details it is described here.

Neural Networks

A number of different Neural Networks (NNs) is provided in Supervisely. NN architectures are available as separate Docker images.

• U-Net
• DeepLab
• Mask R-CNN
• YOLO
• SSD MobileNet
• Faster R-CNN
• ICNet (based on this implementation)
• PSPNet (based on this implementation)

Read here how to run training or inference with this models.

For all source implementations of NNs authors are retaining their original rights.

Library

Supervisely Lib contains Python code which is useful to process data in Supervisely format and to integrate new NNs with Supervisely.

Key features:

• Read, modify and write Supervisely projects on your disk;
• Work with figures (annotations);
• Modify existing implementations of NNs or to create new ones which are compatible with Supervisely;

Reference may be found here.

Related

• Supervise.ly - Website
• Medium - Recent tutorials on how to use SotA models
• Tutorials - Repo with tutorials sources and link to a related blog posts