Using detectron2 for gradient matching

[R-Sadeghi]

Hi,

I'm working with detectron2 for a while, for object detection. Now I want to work on domain generalization using the gradient matching approach. In gradient matching, we train the network on different batches from different domains and finally update the weights using the weights we calculated in the iteration and use the new network for the next iteration.

Is it possible to implement this approach using detectron2 due to weight updating in each iteration?
I will be thankful if you help me with this problem.
The arxiv link of the related paper is https://arxiv.org/abs/2104.09937.

Thank you
Reyhaneh