HI,
We are trying using tflite framework with npu to do machince learing. Here is our step:
1）tranin a model by pytorch
2) export model with onnx format
3) covnert model by eiq tools
and we meet this issue :
And we need you to help us to answer these questions：
a) Dose eIQ(version 2.7.12 ) support convert onnx model to tflte format ?(file see in attach)
b) we can not found quantization option for float16 , do you know whether tf-lite framework can support use npu (nnapi) to inference with float16 precision ？ （PS ：original model is onnx）
c) Any recommend for use NPU to do inference with float 16 ? ONNX runtme 、 Tensor flow 、DeepViewRT ?（PS ：original model is onnx , bsp version we use 5.10.72-2.2.2）
Waiting for your response.
Best Regards!
Hi @wuhuangcangg,
Please find the answers to your questions below:
a) I have reviewed this case and I tried to replicate it, but I got the same issue with conversion ONNX to TFLite in eIQ Model Tool 2.7.5.
Then I used the eIQ Model Tool 2.6.1 with a succesful conversion.
I suggest using eIQ Model Tool 2.6.1, you will find at this link: eIQ Download
b) Yes, actually there is no support for float16 for quantization using the eIQ Model Tools.
c) I'm not totally sure about the full support of float16 NPU inference but you could try with a quantized model with this datatype and using NNAPI as execution provider.
I hope this information will be helpful.
Have a great day.
