Kalray Neural Network Models

The KaNN™ Model Zoo repository provides a list of neural networks models ready to compile & run on MPPA® manycore processor. This comes on top of KaNN™ tool for model generation and enhance AI solutions onto Kalray processor.

SDK Kalray Neural Network (KaNN)

Kalray Neural Network (KaNN) is a SDK included in AccessCore Embedded (ACE) compute offer to optimize AI inference on our dedicated processor called MPPA® (last generation, the 3rd, is named Coolidge 2). It is composed by:

• generator : a python wheel to parse, optimize and paralellize an intermediate representation of a neural networks. Thanks to the runtime, it gives you then the opportunity to run the algorithm directly on the MPPA®
• runtime : optimized libraries (in ASM/C/C++) to execute each operation nodes.

Important notes on ACE-5.4

• Since ACE 5.4.0, the extension file of the serialized binary file (serialized_params_<my_network>.bin) generated by kann is now renamed to <my_network>.kann file. The use of the file is exactly the same in ACE 5.3.0 (and older version). Please refer to ACE's README documentation.
• Tensorflow and Tensorflow-lite are now deprecated in ACE-5.4.0 version and would be dropped in next ACE release. All TF networks of the repository can be converted into ONNX, using tf2onnx for example link

Contents

To quickly deploy a neural network on the MPPA®, a WIKI note is available here:

• Kalray neural networks (KaNN) framework description
• Pre-requisites: SW environment & configuration
• How models are packaged
• Generate a model to run on the processor (MPPA®)
• Evaluate the neural network inference on the MPPA®
• Run neural network as a demo
• Custom Layers for extended neural networks
• Jupyter Notebooks

CNN Models are divided into 3 types of Machine Vision applications:

• classification
• object-detection
• segmentation

The examples below illustrates the kind of predictions you must have:

Classification (SqueezeNet)	Object-detection (Yolov8n)	Segmentation (Deeplabv3+)

*images has been realized using model from this repository and KaNN™ SDK solution (ACE-5.4.0)

List of Neural Networks

All networks are proposed into selected Neural Network architectures, such as:

Classifiers : complete list can be found here

• DenseNet
• EfficientNet
• Inception
• ResNet
• RegNet
• MobileNet
• NasNet
• SqueezeNet
• VGG

Object-detection : complete list can be found here

• EfficientDet
• Faster-RCNN
• FCN
• RetinatNet
• SSD
• YOLO

Segmentation : complete list can be found here

• DeeplabV3+
• Mask-RCNN
• UNet
• YOLO

Requirements

Hardware requirements

Host machine(s):

• x86_64 CPU
• DDR RAM 8Go min
• HDD disk 32 Go min
• PCIe Gen3 min, Gen4 recommended

Acceleration card(s):

• 
• 

Software requirements

• 
• 
•