AndesAIRE(TM) Neural Network Software Development Kit (hereafter NN SDK) is a collection of software tools for facilitating the execution of pretrained neural network (NN) models on Andes development platforms.
This repository provides a collection of widely used neural-network models and the resources required by the NNPilot(R) submodule in AndesAIRE NN SDK for quantizing FP32 models.
Each model entry includes the deployment workflow using AndesAIRE NN SDK, covering data preparation, calibration, evaluation, and the associated evaluation methods.
Each model directory also stores the generated quantized TFL model, the corresponding test inputs, and the resulting test metrics generated from either the TFL model or the fake-quantized model in PyTorch format, enabling rapid reproduction of the subsequent hardware-deployment workflow.
| Network Architecture | Task | Network |
|---|---|---|
| CNN | Image Classification | MobileNetV1 |
| MobileNetV2 | ||
| ResNet-8 | ||
| ResNet-50 | ||
| ShuffleNet-v2 | ||
| SqueezeNet-v1_1 | ||
| EfficientNet-Lite0 | ||
| Inception-v2 | ||
| Object Detection | YOLOv1-tiny | |
| YOLOv2-tiny | ||
| MobileNetV1-SSD | ||
| Keyword Spotting | BC-ResNet-8 | |
| DS-CNN-L | ||
| Visual Wake Words | MCUNet-VWW2 | |
| Speech Recogntion | Tiny-Wav2letter | |
| Face Detection | BlazeFace | |
| RetinaFace | ||
| Face Recognitionn | GhostFaceNet | |
| RNN | Noise Suppresion | RNNoise |
| DeepFilterNet | ||
| DTLN |
Contact your Andes representative to obtain access to the AndesAIRE NN SDK toolkit.
Clone this repository to the desired path following the instructions in the AndesAIRE NNPilot User Manual (UM267).pdf.
Prepare the dataset according to the instructions in each model directory's README. For running the quantization workflow of a selected model, see the AndesAIRE NN SDK Starter Manual (UM266).pdf.
AndesAIRE NN SDK Model Zoo is a project maintained by Andes Technology Corporation, Inc.
See the LICENSE file for details.