convert a (pytorch/onnx)model to NMSIS-NN about nmsis HOT 6 CLOSED

Not yet any plan to convert pytorch/onnx model to NMSIS-NN, if you want to take a try, you can try the following things:

1. tflite-micro: https://github.com/Nuclei-Software/tflite-micro/tree/nuclei_main/tensorflow/lite/micro/nuclei_demosoc
2. tengine(need some port using nmsis-nn): https://github.com/OAID/Tengine
3. any other opensource ai model deploy model tool, and you can adapt the tool to support NMSIS-NN api, if the tool already support CMSIS-NN, it will be easy to achieve it.

from nmsis.

Not yet any plan to convert pytorch/onnx model to NMSIS-NN, if you want to take a try, you can try the following things:

1. tflite-micro: https://github.com/Nuclei-Software/tflite-micro/tree/nuclei_main/tensorflow/lite/micro/nuclei_demosoc
2. tengine(need some port using nmsis-nn): https://github.com/OAID/Tengine
3. any other opensource ai model deploy model tool, and you can adapt the tool to support NMSIS-NN api, if the tool already support CMSIS-NN, it will be easy to achieve it.

Thanks very much for your help! And I did some investigation regarding to your advice.

1. It seems tflite-micro is the most suitable tool I could use right now (for tensorflow model).
2. I am aware tengine could deploy model to general RISCV device. However, I am wondering how is its efficiency, comparing to NMSIS-NN?
3. This is the most proper way if I need to convert pytorch/onnx models. But developing new toolchain is exhausting. We may not be able to do it immediately.

Here are my new questions, I hope you can give me some feedback, Thanks very much!!!

1. Is there any misunderstanding about my above findings?
2. regarding to [1], did tflite-micro use NMSIS-NN library? if not, how is its efficiency, , comparing to NMSIS-NN?
3. regarding to [2], did tengine use NMSIS-NN library? if not, how is its efficiency, , comparing to NMSIS-NN?

from nmsis.

Currently tflite-micro support is using NMSIS-NN, you can take a try. For tengine, it may no use any of NMSIS-NN code, but you can port to it, if it has cmsis-nn support, it can be easily ported.

Thanks

from nmsis.

Currently tflite-micro support is using NMSIS-NN, you can take a try. For tengine, it may no use any of NMSIS-NN code, but you can port to it, if it has cmsis-nn support, it can be easily ported.

Thanks

1. What do you meaning by porting NMSIS-NN to tengine? Do you mean adding new CODEGEN to Tengine for NMSIS-NN? Could you explain it in detail?
2. I did see some comments about converting cmsis-nn to nmsis-nn somewhere else, but I did not understand what meaning is that. Here is my naive guess: nmsis-nn a variant of cmsis-nn, which is made for RISCV. NMSIS-NN is roughly simultaneously developed. Generally speaking, their header files are synchronized (similar API). That is saying any CMSIS-NN api call can be replaced by a corresponding NMSIS-NN api. I have no idea about the correnctness of my guess. Can you please give me more information about NMSIS-NN & CMSIS-NN and the "port" things in a Toolchain

from nmsis.

1. yes
2. NMSIS-NN are designed to be api compatiable with CMSIS-NN, just replace arm to riscv, see https://doc.nucleisys.com/nmsis/nn/index.html

from nmsis.

1. yes
2. NMSIS-NN are designed to be api compatiable with CMSIS-NN, just replace arm to riscv, see https://doc.nucleisys.com/nmsis/nn/index.html

1. Is there an existing opensource NMSIS-NN codegen I could use to migrate? Or I should implement it on my own?

from nmsis.