Acuity Model Zoo

Acuity model zoo contains a set of popular neural-network models created or converted (from Caffe, Tensorflow, TFLite, DarkNet or ONNX) by Acuity toolset.

Model Viewer

Acuity uses JSON format to describe a neural-network model, and we provide an online model viewer to help visualized data flow graphs. The model viewer is inspired by netscope.

Classification

• Alexnet
• Inception-v1
• Inception-v2
• Inception-v3
• Inception-v4
• Mobilenet-v1
• Mobilenet-v2
• Nasnet-Large
• Nasnet-Mobile
• Resnet-50
• Resnext-50
• Senet-50
• Squeezenet
• VGG-16
• Xception

Detection

• Faster-RCNN-ZF
• Mobilenet-SSD
• Mobilenet-SSD-FPN
• MTCNN PNet RNet ONet LNet
• SSD
• Tiny-YOLO
• YOLO-v1
• YOLO-v2
• YOLO-v3

Segmentation

• ENET
• SegNet
• DeepLab-v3

Pixel Processing

• Denoiser
• Super Resolution
• Fast Style Transfer
• Pix2Pix

Pose Estimation

• Open Pose
• CPM Mobile

Recurrent Net

• LSTM - Command Recognition
• LSTM - Speech Recognition

About Acuity

Acuity is a python based neural-network framework built on top of Tensorflow, it provides a set of easy to use high level layer API as well as infrastructure for optimizing neural networks for deployment on Vivante Vision IP powered hardware platforms. Going from a pre-trained model to hardware inferencing can be as simple as 3 automated steps.

• Importing from popular frameworks such as Caffe and Tensorflow

  AcuityNet natively supports Caffe, Tensorflow, TFLite, DarkNet and ONNX imports, it can also be expanded to support other NN frameworks.

• Fixed Point Quantization

  AcuityNet provides accurate Fixed Point Quantization from floating point 32 with a calibration dataset and produces accuracy numbers before and after quantization for comparison

• Graph Optimization

  Neural-network graph optimization is performed to reduce graph complexity for inference, such as Layer Merging, Layer Removal and Layer Swapping

  • Merge consective layers into dense layers, such as ConvolutionReluPool, FullyConnectedRelu, etc.
  • Fold BatchNrom layers into Convolution
  • Swap layer ordering when suitable to reduce output size
  • Remove Concatenation and Split layers
  • Horizontal layer fusion
  • Intelligent layer optimization when mathamatically equivalent

• Tensor Pruning

  Pruning neural networks tensors to remove ineffective synapses and neurons to create sparse matrix

• Training and Validation

  Acuitynet provides capability to train and validate Neural Networks

• Inference Code Generator

  Generates OpenVX Neural Network inference code which can run on any OpenVX enabled platforms