Supported neural networks and runtimes
Supported Neural Networks and formats
Here is a list of neural networks and runtimes that run on the devices DSP that provides adequate performance for real time inferencing.
| Network Topology | Verified Formats | Support for DSP |
|---|---|---|
| AlexNet | Caffe2 | Y |
| DeepLabV3+ | Tensorflow | Partial |
| GoogleNet | Caffe2 | Y |
| Inception-V3 | Tensorflow, Caffe | Y |
| Inception-V3 2016 | Tensorflow | Y |
| Inception-ResNet v2 | Tensorflow | Y |
| Inception-V4 | Caffe | Y |
| MobileNet | TensorFlow | * |
| MobileNet-v2 | TensorFlow, Caffe | * |
| MobileNet-SSD | TensorFlow, Caffe | * |
| ResNet-50 | TensorFlow | Y |
| ResNet-101 | Caffe | Y |
| ShuffleNet | TensorFlow | Y |
| SqueezeNet | TensorFlow | Y |
| VGG-16 | TensorFlow | Y |
| VGG-19 | Caffe | Y |
*this network topology requires changes to preserve accuracy when quantized. Instructions for topology modification are available from Qualcomm.
ONNX support
ONNX support is currently experimental and most models will run on the CPU, rather than be offloaded to the DSP.
ONNX models from the following networks can be converted for use with the Vision AI DevKit:
- bvlc_alexnet
- bvlc_googlenet
- bvlc_reference_caffenet
- bvlc_reference_rcnn_ilsvrc13
- densenet121
- inception_v1
- inception_v2
- resnet_50
- vgg16
- vgg19