GinZhu / RDST

Official implementation of RDST. A residual dense swin transformer for medical image super-resolution
https://arxiv.org/abs/2302.11184
17 stars 1 forks source link

Training Discrepancy #3

Closed yuque2023 closed 11 months ago

yuque2023 commented 1 year ago

Hello, I trained the RDST model using the approach you recommended, but the PSNR and SSIM values I obtained during testing differ significantly from the ones you provided. I'm wondering if I made a mistake during training, or if you have unique training techniques? Here are the metrics I obtained: Case IDs: ['OAS1_0009_MR1', 'OAS1_0033_MR1', 'OAS1_0023_MR1', 'OAS1_0004_MR1', 'OAS1_0019_MR1', 'OAS1_0032_MR1', 'OAS1_0029_MR1', 'OAS1_0010_MR1', 'OAS1_0003_MR1'] SR psnr ssim mse uqi ergas scc vifp fid


4 32.71(3.2) 0.9218(0.037) 0.0006475(0.00033) 0.6924(0.3) 7.013(9.9) 0.3758(0.16) 0.8148(0.065) 93.57(5.7)

I made modifications to the 'RDST_E1_OASIS_example_SRx4.ini' file you provided by changing the number of epochs to 100000 and 20000, and performing validation every 2000 epochs. I would like to ask if there are any other aspects that need to be changed.