.pt to .onnx conversion using u5 branch export.py

[efedmr1r]

I converted a custom yolov7 pt file to onnx but when I try to do inference with cv2.dnn module I keep getting the error
Can't create layer "onnx_node!ScatterND_243" of type "ScatterND" in function 'getLayerInstance'

Full error:
[ERROR:0@0.055] global /home/efedmrlr/Desktop/opencv/modules/dnn/src/onnx/onnx_importer.cpp (1041) handleNode DNN/ONNX: ERROR during processing node with 3 inputs and 1 outputs: [ScatterND]:(onnx_node!ScatterND_243) from domain='ai.onnx'
Traceback (most recent call last):
File "main.py", line 161, in
yolov7_detector = YOLOv7(args.modelpath, conf_thres=args.confThreshold, iou_thres=args.nmsThreshold)
File "main.py", line 13, in init
self.net = cv2.dnn.readNet(path)
cv2.error: OpenCV(4.6.0-dev) /home/efedmrlr/Desktop/opencv/modules/dnn/src/onnx/onnx_importer.cpp:1063: error: (-2:Unspecified error) in function 'handleNode'

Node [ScatterND@ai.onnx]:(onnx_node!ScatterND_243) parse error: OpenCV(4.6.0-dev) /home/efedmrlr/Desktop/opencv/modules/dnn/src/layer_internals.hpp:110: error: (-2:Unspecified error) Can't create layer "onnx_node!ScatterND_243" of type "ScatterND" in function 'getLayerInstance'

Note: I can export and use the example yolov7.pt coco weights but cant use my custom model.

[WongKinYiu]

You have to do reparameterization before convert the model.

[efedmr1r]

How should I reparameterize a yolov7-tiny model? There is no yolov7-tiny in reparameterization.ipynb

[holger-prause]

Just use the reparam script.
Dont forget to change it so it will work with tiny model too.

#514

[efedmr1r]

Just use the reparam script. Dont forget to change it so it will work with tiny model too.

#514

Hi @holger-prause,
I changed all model.105's to model.77 like you had done but I got another error now
RuntimeError: The size of tensor a (255) must match the size of tensor b (36) at non-singleton dimension 0

[holger-prause]

Hmm i don't know about this error - this whole procedure can be a bit error prone - i can only tell you how i do it:
Notice all the mentioned steps are referring to the master branch

• train a tiny model - make sure the training reached the end (i did by supplying --epochs argument as observing the logs showed no more improvements for like 25 epochs - and yolov7 does not support early stopping point). If i took intermediate model checkpoints - the whole process of of reparam and exporting did not worked for me
• perform the reparametrizations via the reparamscript and do the changes mentioned in Reparametrize tiny model model error #514 Don't forget to also adjust the number of classes and and pick the right model from deploy as mentioned in the linked issue
• export it to onnx and perform your detections- i am using the onnx runtime instead of the dnn module as i struggled to get it working with opencv https://colab.research.google.com/github/WongKinYiu/yolov7/blob/main/tools/YOLOv7onnx.ipynb#scrollTo=07vh3pUa1ccC

So in the end i got it working with my custom dataset and tiny model config and it is possible. There is nothing more i can say or do - good luck and you will get it working at some point in time. If you are still struggling - try starting from beginning and revert all you changes.

[efedmr1r]

Thank you so much, I converted the the model as you did and now I can use it with the onnx runtime. But it seems like I also have a problem with opencv dnn module, it keeps giving errors.

[efedmr1r]

Hmm i don't know about this error - this whole procedure can be a bit error prone - i can only tell you how i do it:
Notice all the mentioned steps are referring to the master branch

• train a tiny model - make sure the training reached the end (i did by supplying --epochs argument as observing the logs showed no more improvements for like 25 epochs - and yolov7 does not support early stopping point). If i took intermediate model checkpoints - the whole process of of reparam and exporting did not worked for me
• perform the reparametrizations via the reparamscript and do the changes mentioned in Reparametrize tiny model model error #514 Don't forget to also adjust the number of classes and and pick the right model from deploy as mentioned in the linked issue
• export it to onnx and perform your detections- i am using the onnx runtime instead of the dnn module as i struggled to get it working with opencv https://colab.research.google.com/github/WongKinYiu/yolov7/blob/main/tools/YOLOv7onnx.ipynb#scrollTo=07vh3pUa1ccC

So in the end i got it working with my custom dataset and tiny model config and it is possible. There is nothing more i can say or do - good luck and you will get it working at some point in time. If you are still struggling - try starting from beginning and revert all you changes.

Hi again
You said in the end you got it working with opencv dnn module. Can you give me some information about how did you fixed the issue with opencv?

[holger-prause]

Hmm maybe my statement was ambigious when i stated

So in the end i got it working with my custom dataset and tiny model config and it is possible

But i didnt stated it was opencv, actually i mentioned the opencv dnn module problem and i that i am using the onnx runtime, i still cannot use the opencv dnn module

If you export without the nms layer on the end - at least you can read the model with opencv, however any existing yolov7 script still do not work and my brain struggles too much understanding the datastructure returned from the output layer.
Maybe you have more luck - however this is a different discussion - i would advise closing this ticket.

[akashAD98]

You have to do reparameterization before convert the model.

#552