Welcome to Hailo's Developer Zone

User Models

Bring your own model, either built on If your model is built on one of the supported machine learning frameworks as listed above or built on any other network. In that case, we recommend converting it to ONNX or TF-Lite using the Dataflow Compiler. 
You can always use our Hailo Model Zoo to download ready-to-use models. 
The Hailo Dataflow Compiler can receive models in 32bit float precision. 
The Hailo parser which is the first step of the Dataflow Compiler process will translate the original model into a Hailo internal format. 
The second step will run the optimizer which will quantize the weights into a 4/8/16-bit integer format (8-bit as default). 

Hailo Model Zoo

The Hailo Model Zoo provides deep learning models for various computer vision tasks. The pre-trained models can be used to create fast prototypes on Hailo devices. Main features include: 

• A variety of common and state-of-the-art pre-trained models and tasks in TensorFlow and ONNX 
• Model details, including full precision accuracy vs. quantized model accuracy measured on Hailo-8™ 
• Each model also includes a binary HEF (Hailo Executable Format) file that is fully supported in the Hailo toolchain and Application suite (for registered users only) 
• Retraining Dockers (for selected models) 
• Ability to go through the whole process – parsing, model optimizer, compilation, inference, accuracy evaluation – either on the pretrained model or on a retrained model 

Additionally, the Hailo Model Zoo Github repository provides users with the capability to quickly and easily reproduce the Hailo-8’s published performance on the common models and architectures included in our Model Zoo. 

Hailo Dataflow Compiler

The Hailo Dataflow Compiler is responsible for converting the model from the original format all the way to a binary format that can run inference on the Hailo device. 
There are three major steps in the compilation process: 

1. Model Parser - translating the original model into Hailo format 
2. Model Optimizer - quantizing the model and optimizing the accuracy performance 
3. Compiler - allocating the needed resources and compiling it into a binary format that can run on the Hailo device. 

On top of these mandatory steps in compilation there are two very important tools to help with the process: 

• Profiler - profiling the expected model behavior while running on the Hailo hardware and predicting how many hardware resources are required to run the model
• Emulator - emulating the accuracy of the model and supporting in running the optimization process.

TAPPAS

Use the TAPPAS  demo suite to see a set of full application examples by running a single command line or via TAPPAS GUI. 
The TAPPAS demonstrates a variety of application examples ranging from classification, detection, and segmentation, all the way up to more complicated pipelines including several models such as LPR and VMS applications.
The TAPPAS code is open and can be viewed to learn how to use the different APIs of the HailoRT for your application.
Other runtime frameworks can be supported and used such as ONNX Runtime, GStreamer and others. 

OS IP Stack

Hailo supports PCIe interface as well as Ethernet.  
For the PCIe, a driver needs to be installed on your system - either pre-compiled or specifically compiled for your platform. 
For the Ethernet interface you have to use the OS IP Stack on your device.

Hailo RT

The HailoRT library is your means to communicate with the Hailo device. 
From the basic operations like loading models, sending and receiving data, to more complex pipeline management such as scheduling tasks, automatically postprocessing functionality such as NMS support on and off the Hailo device, and more. 
The HailoRT supports Python, C and C++ APIs.

Hailo-8™ Device

To run inference at top performance using Hailo-8™, you have to connect the Hailo device to a board. 

Hailo offers both M.2 and mPCIe modules, as well as an M.2 starter kit to help you get started with Hailo.