Inference on images with different sizes

[Naxter]

Hi!
Thank you for the amazing work. I appreciate it and hope to contribute in the future.
So far, I worked with official implementations of the authors that mostly wrote "pure" pytorch/tensorflow code. I never had problems infering the model with any arbitrary size after training (for those networks that were not restricted).
I always trained on lower image sizes (cropped images) due to memory requirements and then infered with much higher image size (image after image). I never had problems with that (I never scaled my images).

I wanted to do the same for my mmdetection trained network. I tried to find documentation or related issues/discussions but everyone just used the same image sizes, which is not possible for me.

Do I miss something? How can I infer the model with any image size after training? Does the mmdetection framework support this? Or is the whole train/test pipeline thing restricting this?

I hope to find some answers here. I would really appreciate it!
Thank you!