I have a personnal convolutionnal spiking NN, could Hailo accelerate my model ?
Welcome to the Hailo Community!
Whether your personal convolutional spiking neural network can be accelerated on Hailo devices mainly depends on how the spiking behavior is implemented.
Hailo devices are designed to accelerate standard deep learning operators such as convolutions, activations, pooling, and fully connected layers. The Hailo Dataflow Compiler User Guide contains a detailed list of all supported layers and operators. If your network is mostly composed of supported operations, then at least part of the model may be accelerated on Hailo hardware.
The Hailo Dataflow Compiler accepts ONNX and TFLite models as input. During conversion, unsupported layers or operations (often custom pre-processing, post-processing, or non-standard neural operations) may need to be excluded from the Hailo graph and executed on the host CPU instead.
I am not very familiar with spike-based execution models specifically, so I cannot say with certainty how well native spiking behavior maps to the Hailo architecture. If the spiking portions of the model are represented using standard ONNX/TFLite operators, it may still be possible to accelerate parts of the network.
In practice, a common approach is to:
- run the supported convolutional backbone on the Hailo accelerator,
- and execute unsupported operations on the host CPU.
I would recommend first getting familiar with the model conversion workflow by working through the tutorials included in the Hailo AI Software Suite Docker. After that, you can try converting your own model and inspect which layers are supported by the compiler and which are not.
1 Like