FineTuning with similar format dataset as coco but different classes

[Naivepro1990]

Hi,

I want to fine DETR with data similar in format as mscoco but the number of classes as different. I have only 6 classes.

Firstly, Where I need minimum changes in DETR code such as the change in the last layer to predict the number of classes and use the pretrained weights.

Secondly, the number of classes will be 6 or 7 i.e. 6 classes and 1 background.

Regards,

[m-klasen]

Hello, please take a look at the following links for information.
#9
https://gist.github.com/mlk1337/651297e28199b4bb7907fc413c49f58f
https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr
These links should include plenty of examples for training on custom data.
Regards

[alcinos]

Hello @Naivepro1990
In addition to @mlk1337 's ressources, I encourage you to checkout #9, #152 and #125
Feel free to ask questions there if you are still stuck.
Best of luck.