Could you share YOLOv3.onnx model in the sample from renesas?

Hi AnhOi,
My apologies, we cannot provide the requested model.
Thank you.
Kind Regards.

The AI SDK Demo are based on open source trained models. For yolov3 the models and the dataset used to create the trained model can be found.

YOLOv3: Darknet
Dataset: COCO

Example of how to translate the YoloV3 model to run on the RZV2L DRP please refer here.

The Models used for AI SDK are based on open source trained models. The YoloV3 model can be found below

YOLOv3: Darknet
Dataset: COCO

For information on how to traslate the model to run on the RZV2L DRP-AI please refer to here.

Hi PT_Renesas 

I only want to check the shape of 3 outputs from renesas's yolov3, because I could not find 3 nodes with same shape of YOLOv3 as Renesas mentioned.

In this link https://github.com/renesas-rz/rzv_ai_sdk/tree/main/R01_object_detection you said that

Input size: 1x3x416x416
Output1 size: 1x13x13x255
Output2 size: 1x26x26x255
Output3 size: 1x52x52x255

I follow this repo to convert from darknet to yolov3 https://github.com/zldrobit/onnx_tflite_yolov3 and visualize onnx model by Neutron but I did not see 3 nodes correspond with your above shape.

From my point of view, it is hard for user to follow guideline and customize for other problem. I think that if renesas can not provide model for specific dataset, you can provide general guide to create model (reproduce architecture). I think that it is better for users.

Thank you.

Hi michael kosinski 

Thank you for response. I also found this link today. It is better to link this guide https://github.com/renesas-rz/rzv_drp-ai_tvm/blob/main/docs/model_list/how_to_convert/How_to_convert_yolov3_onnx_model_V2L_V2M_V2MA.md to  https://github.com/renesas-rz/rzv_ai_sdk/tree/main/R01_object_detection 

Thank you again.

Hi michael kosinski 

I convert from yolov3.pt to yolov3.onnx, here are 3 outputs with shape. It seems that they are different than shapes in this link https://github.com/renesas-rz/rzv_ai_sdk/tree/bba59d0de1b6a6997eada351bc80e005fca76498/R01_object_detection#ai-model Could you please tell me specific output nodes of YOLOv3 you get for this application https://github.com/renesas-rz/rzv_ai_sdk/tree/bba59d0de1b6a6997eada351bc80e005fca76498/R01_object_detection#ai-model 

Thank you.

Hi AnhOi,
Please chek this header file.
The output shape is defined in lines 148-165.

There is a typo in the size description of README.
Kind Regards.

Hi PT_Renesas 

Thank you. Let me check it more.

Hi PT_Renesas 

Thank you.

I checked this link that you mentioned. In the line of header file https://github.com/renesas-rz/rzv_ai_sdk/blob/bba59d0de1b6a6997eada351bc80e005fca76498/R01_object_detection/src/define.h#L163 there is output inference size, not shape.

Because with different shapes, we can have same output inference size.

The shape of the tensor floatarr is determined by the number of output layers (NUM_INF_OUT_LAYER), the number of bounding boxes per grid cell (NUM_BB), the number of classes (NUM_CLASS), and the grid sizes (num_grids[]).

Each output layer has a grid of size num_grids[n] × num_grids[n], where n ranges from 0 to 2, meaning the grids are:

• 13 × 13
• 26 × 26
• 52 × 52

Each grid cell contains NUM_BB bounding boxes. Each bounding box has (NUM_CLASS + 5) values:

• 4 for tx, ty, tw, th (bounding box parameters)
• 1 for objectness
• NUM_CLASS for class scores

So, the total number of elements in floatarr is:

∑n=02NUMBB×(NUMCLASS+5)×num_grids[n]×num_grids[n]\sum_{n=0}^{2} NUM_BB \times (NUM_CLASS + 5) \times num\_grids[n] \times num\_grids[n]n=0∑2​NUMB​B×(NUMC​LASS+5)×num_grids[n]×num_grids[n]

Substituting values:

$$3\times (80+5)\times (13\times 13+26\times 26+52\times 52)$$ $$3\times 85\times (169+676+2704)$$ $$3\times 85\times 3549=904905$$

Thus, the shape of floatarr is (904905, ), a 1D tensor

The 1D tensor floatarr can be split into three separate tensors, corresponding to each grid size in num_grids[]:

1. First tensor (for 13×13 grid)

   $$(3,85,13,13)$$
   • 3 = NUM_BB (bounding boxes per grid cell)
   • 85 = NUM_CLASS + 5 (attributes per bounding box)
   • 13 × 13 = grid size

2. Second tensor (for 26×26 grid)

   $$(3,85,26,26)$$

3. Third tensor (for 52×52 grid)

   $$(3,85,52,52)$$

Hi michael kosinski 

Thank you for detailed information. Yes, I understood total number elements of floatarr.

As my understanding, we will use 3 default outputs of YOLOv3, because there is not note as for YOLOX https://confluence.renesas.com/display/REN/Translation+procedure+for+YOLOX . Sorry for miscummunication, I only want to confirm specific output nodes for YOLOv3.

Thank you again.