This fixes metric aggregation for NMSE and PSNR in the validation loop of the Lightning modules, which likely caused Issue #129. Previously, NMSE was being normalized with the slice norm and PSNR was using the maximum value of the slice. Both of these should have been using volume metrics to match the methodology in the fastMRI paper. This PR fixes the volume metrics.
This fixes metric aggregation for NMSE and PSNR in the validation loop of the Lightning modules, which likely caused Issue #129. Previously, NMSE was being normalized with the slice norm and PSNR was using the maximum value of the slice. Both of these should have been using volume metrics to match the methodology in the fastMRI paper. This PR fixes the volume metrics.
Previous U-Net Knee Single-Coil NMSE: 0.04213 PSNR: 27.02
This PR (epoch 43) NMSE: 0.0323 PSNR: 32.34