error when try with undersampled data

[linshuijin123]

hi, I had successfully reappeared your erxperience with fastmri train data, and get the inference with fastmri test data{-r09.tfrecords}, as far as I know, -r09 means 512512, so I try to inference with -r08, which is 256256, I got this error: tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.

(0) Invalid argument: required broadcastable shapes at loc(unknown) [[{{node while/cond/add}}]] [[Abs/_1463]] (1) Invalid argument: required broadcastable shapes at loc(unknown) [[{{node while/cond/add}}]] 0 successful operations. 0 derived errors ignored.

1. maybe I need to pad the images before I test?
2. Additionally, I find images in -r09.tfrecords all have two channels, like 2512512, I don't get why two channels are needed? and is there any relation between these two channels?
3. I noticed that you mentioned Mapper gets two inputs: time index and extracted information about data (like image contrast, R), how did you get the latter? auto by attention or man-made by adding a condition label? Looking forward to your reply! I can't be more honored to reappear your method.

[yilmazkorkmaz1]

Hi,

Thanks for your interest.

1) You should try with 512x512 images and it automatically crops images in fastmri according to the label (i.e., contrast).

2) FastMRI gave 2 channels because it has complex images, first channel corresponds to real second channel imaginary component.

3) R can be get easily if you take fourier transform of undersampled image and count the zeros on fourier domain. We did not automate it and labeled by hand since I composed that dataset by using various contrasts/acc. rates I knew the sources.