Question about implementing yolov5

[18520506]

I followed this link: https://renesas.info/wiki/YoloV5_TVM_Implementation_Guide to implement Yolov5L. When I exported the onnx, I get the structure of the onnx like image below:

But after I translated the onnx for drp-ai. I got the structure of the onnx after translated like below:

It looks like that the onnx's input is different with the yolov5's onnx. Why does the onnx's input change after being translated? Is it a bug ?

[simobepu]

Thank you for your question. It's not a bug. The translated ONNX is the part of the original ONNX and it is the part that can be processed by DRP-AI. The part from the input (images) to the first Mul in the original ONNX is processed by the CPU. This part is not output as ONNX.

[18520506]

Thank you very much for your answer.

[18520506]

I have a question, @simobepu . What is the right DRP-AI input image information? I have done some experiments but I still didn't know which is right. Experiment 1: I have tried to define value as followed:

    constexpr static int32_t  TVM_MODEL_IN_W = (640);
    constexpr static int32_t  TVM_MODEL_IN_H = (640);
    constexpr static int32_t  TVM_MODEL_IN_C = (3); 
    num_grids = {80,40,20}
    constexpr static int32_t YOLOV5_NUM_BB= 3;
    constexpr static int32_t YOLOV5_NUM_INF_OUT_LAYER=3;

Note: I used the bounding box and class prediction as YOLOV3

If I defined the value right, why does the result look like below: Experiment 2: I have tried to do as followed:

    constexpr static int32_t  TVM_MODEL_IN_W = (320);
    constexpr static int32_t  TVM_MODEL_IN_H = (320);
    constexpr static int32_t  TVM_MODEL_IN_C = (32); 
    num_grids = {80,40,20}
    constexpr static int32_t YOLOV5_NUM_BB= 3;
    constexpr static int32_t YOLOV5_NUM_INF_OUT_LAYER=3;

I used the bounding box and class prediction as YOLOV3 The result is that it did not detect anything. I checked the calculation and found that tx,ty,tw,th,tc sometime has nan value.

If I defined the value right, then does the error happen because I did not change inf_pre_process function to take the input(images) to the first Mul as the previous answer? Thank you for reading

[wk-mnA]

Thank you for posting the issue.

All values you have tried to define depends on each YOLO model. The model defines the input size, output layer details (such as num_grids, _NUM_BB, _NUM_INF_OUT_LAYER) and anchor values.

Check your YOLO config file and model architecture to find the appropriate value for your model. Also, YOLOv5 CPU post-processing(processing for tx,ty, etc.) may differ from YOLOv3. Check your training environment and its post-processing match your application source code.