RV1106 model deployment troubleshooting

maskmoo

RV1106 model deployment troubleshooting [Copy link]

#AI Challenge Camp Terminal#RV1106 Model Deployment - Embedded Systems - Electronic Engineering World - Forum (eeworld.com.cn)

The recognition rate of the final test on the model deployment at the AI Challenge Camp terminal #RV1106 was relatively poor. Previously, we suspected that problems occurred in the model conversion quantization process and the engineering deployment process.

This project deployment used the project of luyism , so the focus of the investigation was mainly on the model itself and RKNN conversion quantization.

The first thing to do is to check the format of ONNX and the converted RKNN. Here is a comparison with luyism and qiao--- .

In this comparison, the data dimensions and types of the input and output are consistent.

Then confirm whether there is any problem with the RKNN format after conversion. Although the internal structure of the output rknn model cannot be parsed by the Netron tool, the basic information of input and output can be viewed for preliminary judgment. Because RV1106 only supports int8 type input and output.. If the input and output tensor structure of the `float` type appears after the model conversion, the data cannot be obtained correctly, and the model structure needs to be readjusted.

At this point, the conversion problem of RKNN has been eliminated.

Finally, I found that there was something wrong with the test method. The method I used before was to identify through computer pictures. I tried to deploy other big guys' engineering models and also sampled computer pictures to test the recognition rate. The initial speculation is that there may be deviations in the camera's acquisition of screen photos (those who know the reason can help to explain it), and the specific reason is still uncertain.

So in the end, I also used handwriting on white paper for testing. The final recognition effect is shown in the video. The overall recognition effect of 0 to 8 is good, and the number 9 is easily recognized as 7. This problem has also been seen in other big guys, but the probability of this happening in my model is a little higher. I'm not sure if it's because of the model itself or my own handwriting habits.

6月10日