Self trained ONNX model

[NeilDale]

Hi,

I am a civil engineer who has some experience in GIS and very little programming experience and I came across your plug-in and I am trying to find a way to use a self trained model on it. I have trained a yolov7 model on +-4000 images of solar panels. I have exported it to ONNX but when I run the plugin in qgis I get the following error: "Error! Processing error: "Unhandled exception occured. See Python Console for detains"!

Details: Type Error - 'NonType' object is not subscriptable Last Traceback: File"---deepness\processing\models\detector.py", line 140, in get_number_of_output_channels
class_names = self.get_outputs_channel_names()[0]

could you please take a look and point me in the right direction to solve it please.

Thank you

Neil

[bartoszptak]

Hi @NeilDale!
It looks like a problem with class names in the model's metadata. The best way to help you is to debug your model. Can you share it? You can send it to me by mail; the address is in my profile description.

[NeilDale]

Thank you so much.

I have sent the model.

[bartoszptak]

For some reason unknown to me, your model does not contain metadata after export. You can add it manually using this guide.

In more detail, metadata allows you to define the behaviour and default parameters for the models. One such parameter is the class_names variable, which specifies what name the model outputs should use. I was sure that the YOLOv6-YOLOv9 model export provides such metadata, but I could be wrong. Deepness shall also automatically label the outputs output_1 in case of an undefined class_names variable. We need to look into this.

[NeilDale]

Bartosz, Thanks for the swift reply. I will give it a try and I will let you know of the outcome, i am a novice so it might take some time. If you are available for a call some time, it would be great just to meet you and have a chat. It's always nice to learn from experiences like yours. I am available almost any time Kind Regards Neil

…

On Tue, Jul 9, 2024 at 11:24 AM Bartosz Ptak ***@***.***> wrote: For some reason unknown to me, your model does not contain metadata after export. You can add it manually using this <https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_add_metadata_to_model.html> guide. In more detail, metadata allows you to define the behaviour and default parameters for the models. One such parameter is the class_names variable, which specifies what name the model outputs should use. I was sure that the YOLOv6-YOLOv9 model export provides such metadata, but I could be wrong. Deepness shall also automatically label the outputs output_1 in case of an undefined class_names variable. We need to look into this. — Reply to this email directly, view it on GitHub <#176 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BH2T7DQ6H7V2KTUIZFULCOTZLOT6PAVCNFSM6AAAAABKQNOSP6VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TSOJWGQ3DM> . You are receiving this because you were mentioned.Message ID: <PUTvision/qgis-plugin-deepness/repo-discussions/176/comments/9996466@ github.com>

[NeilDale]

Hi, Imanaged to follow the instructions and it works very well on pleiades neo imagery that I have. as long as I select pixel size at 10cm. Please feel free to try it out: solar-panel_detect.onnx <https://1drv.ms/u/s!AqNZ3aqm1jXeiMFNrGQmfmuHjPLC7A?e=sSznbL>. I will be happy to receive any feedback if you have any. Thanks again for your time. Kind Regards

…

On Tue, Jul 9, 2024 at 11:51 AM Dale ***@***.***> wrote: Bartosz, Thanks for the swift reply. I will give it a try and I will let you know of the outcome, i am a novice so it might take some time. If you are available for a call some time, it would be great just to meet you and have a chat. It's always nice to learn from experiences like yours. I am available almost any time Kind Regards Neil On Tue, Jul 9, 2024 at 11:24 AM Bartosz Ptak ***@***.***> wrote: > For some reason unknown to me, your model does not contain metadata after > export. You can add it manually using this > <https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_add_metadata_to_model.html> > guide. > > In more detail, metadata allows you to define the behaviour and default > parameters for the models. One such parameter is the class_names > variable, which specifies what name the model outputs should use. I was > sure that the YOLOv6-YOLOv9 model export provides such metadata, but I > could be wrong. Deepness shall also automatically label the outputs > output_1 in case of an undefined class_names variable. We need to look > into this. > > — > Reply to this email directly, view it on GitHub > <#176 (reply in thread)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/BH2T7DQ6H7V2KTUIZFULCOTZLOT6PAVCNFSM6AAAAABKQNOSP6VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4TSOJWGQ3DM> > . > You are receiving this because you were mentioned.Message ID: > <PUTvision/qgis-plugin-deepness/repo-discussions/176/comments/9996466@ > github.com> >

[AlainFELER]

Hello, I tried your model on two small places where I know there is at least one solar panel, but it did not find them. Maybe it is because the local orthophoto pixel size is 20 cm and not 10. Maybe I can find a 10 cm cover, but is that the probably main problem ?

Also, could you explain in some detail how you produced the model ? I would like to make a detector model myself, for which I can also provide thousands of sample images, but I don't know how to proceed. I am sure it would be useful to many people.

[NeilDale]

Hi Alain,

Small disclaimer, I lucked out on getting this model to work and i'm far from an expert in the field.

I tested it on Pleiades neo (30cm imagery)and it could not detect any solar panels, but when i lied to the plugin and said that the pixels are 10cm, it worked. Maybe try smaller pixels values and see what happens.

Regards