Invariance to image shapes about csgm-mri-langevin HOT 2 CLOSED

Hello and thank you for the question, sorry for the late reply.

First, when you say it can sample knees data, please note that this refers to conditional sampling given k-space measurements, not pure unconditional sampling. One can think of the k-space measurements as "guiding" the sampling process for a simplification.

Second, regarding the sizes you mentioned, the network architecture we used only contains convolutional and self-attention blocks and can thus perform forward passes on any height and width shapes (up to some limitations, such as sizes being divisible by 16). What is not invariant is the number of input/output channels, which is 2 in our case.

The exact model is here: https://github.com/ajiljalal/ncsnv2/blob/ef1eae385fd2b4ec9d31f13be3640b91bc621d26/models/ncsnv2.py#L198

To change the spatial dimensions you would need to change the last two shapes of the initial samples as shown below. I'd recommend keeping them a multiple of 16 to avoid other errors:

Or directly via the config file, e.g., like here for the knees experiment:

from csgm-mri-langevin.

Closing due to inactivity, please feel free to reopen if there are further questions.

from csgm-mri-langevin.