Weijie Bao requested to merge Weijie/sadl-extension:ScatterND into dev_v5

Deer Franck, in this MR, I add a new operation called ScatterND, which can be found at Currently, this operation is compatible with PyTorch and TF2, supporting both float and integer inference.

To demonstrate the functionality of ScatterND, I have provided four ONNX files. Two are simple demos, while the other two are complex demos that include Conv2D operations. It's worth noting that the second parameter of ScatterND, called "indices," is expected to be in an integer format. Therefore, the quantization parameter of indices needs to be set to 0. Here's an example command for performing this quantization:

echo "0 15 1 15 2 0" | ./naive_quantization scatternd_c_pytorch.sadl scatternd_c_pytorch_int16.sadl;

Thank you for your time.

Merge request reports