Hello I am having a problem using the YOLOv5-->blob converter: https://tools.luxonis.com/ I have a YOLOv5 model trained at 640x640 I've checked this using Netron I upload the .pt file to https://tools.luxonis.com/ and it converts to blob with no problems However, when I try to run this with DepthAI, I get a message "Mask is not defined for output layer with width '12'" I've checked the json & it looks ok- there is no anchor_masks entry for "side12"- there is a correct entry for "side20" i.e 640/32 Happy to share the data files if someone can help me, thanks Andrew

problem with YOLOv5 using https://tools.luxonis.com/

barney2074

Hello

I am having a problem using the YOLOv5-->blob converter: https://tools.luxonis.com/

I have a YOLOv5 model trained at 640x640
I've checked this using Netron
I upload the .pt file to https://tools.luxonis.com/ and it converts to blob with no problems
However, when I try to run this with DepthAI, I get a message "Mask is not defined for output layer with width '12'"

I've checked the json & it looks ok- there is no anchor_masks entry for "side12"- there is a correct entry for "side20" i.e 640/32

Happy to share the data files if someone can help me,

thanks

Andrew

Matija

Hi @barney2074 , that's strange. Are you using official ultralytics repository?

Could you please share with me which config did you use for training? What flags did you specify when calling the train.py?

From the error it seems that the model contains an output node that should be here. Could you also share the zip that the tools.luxonis.com generate with me please?

Thanks,
Matija

barney2074

Hi Matija

I've been experimenting with Ultralytics HUB- which I'd assume uses the latest YOLOv5.

I did the training locally (basically uses YOLOv5 running in a Docker container) so unfortunately don't have the logs or Wanb record with exact train.py flags

Ultralytics HUB shows basic info:

I believe this is the same as --img 640 --epochs 300 etc

I've put the luxonis tools output and original .pt and YOLOv5 .onnx export here

I did notice in Netron the image size is clear in the first node in the .onnx format, but not in the original .pt (that the onnx was produced from)

thanks in advance

Andrew

Matija

Hi @barney2074 ,

I downloaded the ZIP you've shared with me and it seems to work fine. I tested the model by running python3 main_api.py -m path/to/blob -c path/to/json from gen2-yolo/device-decoding, and providing path to the .blob and .json from the zip.

Which version of depthai are you using? can you make sure to upgrade to the latest version 2.16.0.0 ?

barney2074

thanks Matija

I'll give it another try.
I only cloned it a few days ago- so should be the latest version- but I will check

I did use main.py (rather than main_api.py)- it was not clear to me which to use

Andrew