search

soapisnotfat / super-resolution

collection of super-resolution models & algorithms
Apache License 2.0
744 stars 188 forks source link

about 'fsrcnn' #19

Closed T5eng closed 4 years ago

T5eng commented 5 years ago

optimizer should be SGD instead of Adam learning rate of fsrcnn should be 0.001 weight init is not proper. Hyper parameter in Transposed Conv layer is not correct, causing error when changing upscale factor.

fangangping commented 5 years ago

have you fixed this problem ? I find fsrcnn can not work when i set the upscale factor to 8.

T5eng commented 5 years ago

I think we need to adjust the "outpadding" argument according to different upscale factor. you can check out the pytorch document about ConvTranspose2d.