Suggestions requested for training a LAMA model on higher resolution data

[rachit221195]

Currently all LAMA models have been trained on 256x256 crops of 512x512 images.
I would like to understand what changes should be made to train a LAMA model on a bigger image resolution - maybe 512x512 crops from 1024x1024 images.

I want suggestions which can guide me on what changes in network architecture (number of upsampling, downsampling steps, number of resnet modules) can be experimented with. Apart from network architecture, are there any other changes that might be worth experimenting with.

[Abbsalehi]

There are two datasets with the below sizes mentioned in the Readme file.

Places dataset: 512 by 512 images
CelebA dataset: 256 by 256 images

Once you generate masks, you can crop your 1024x1024 images to 512 or cropping using a Python code by yourself. Please let me know if you have any questions.