search

isaaccorley / pytorch-enhance

Open-source Library of Image Super-Resolution Models, Datasets, and Metrics for Benchmarking or Pretrained Use
Apache License 2.0
80 stars 12 forks source link

Using bicubic model results in ValueError: optimizer got an empty parameter list #12

Closed robmarkcole closed 3 years ago

robmarkcole commented 3 years ago

Using model = Bicubic(scale_factor, channels) results in:

image

I realise this model is not for training, but wish to get the PSNR on the test dataset for use as baseline. Is it valid to just do trainer.test(module, test_dataloader) ?

isaaccorley commented 3 years ago

Yeah you can't actually train the Bicubic model because it has no learnable parameters so you would just do inference or test.

robmarkcole commented 3 years ago

OK thanks. Interestingly my test_psnr result was better for bicubic than for srcnn & espcnn in my quick trial..!

isaaccorley commented 3 years ago

Yeah I've experienced the same with SRCNN but not ESPCN. I've had better results by using a small learning rate with SGD and training for longer. Sometimes PSNR can be relatively the same but visually the models produce better results compared to interpolation.