krasserm / super-resolution

Tensorflow 2.x based implementation of EDSR, WDSR and SRGAN for single image super-resolution
Apache License 2.0
1.48k stars 357 forks source link

EDSR Model PSNR not improving #69

Open purijs opened 3 years ago

purijs commented 3 years ago

I'm trying to train a model on custom data, however the PSNR is stuck around 19 while loss is constantly going down, what can I infer from this?

In what sense is the model improving or not improving? My data is around 300 images, scaled at 4x

70000/400000: loss = 10.043, PSNR = 19.924332 (608.94s)
80000/400000: loss = 9.812, PSNR = 19.933113 (608.67s)
90000/400000: loss = 9.628, PSNR = 19.861080 (608.45s)
100000/400000: loss = 9.466, PSNR = 19.953009 (608.42s)
110000/400000: loss = 9.336, PSNR = 19.993717 (608.45s)
120000/400000: loss = 9.225, PSNR = 19.842371 (608.44s)
tuvovan commented 3 years ago

any updates?