search

WolframRhodium / Super-Resolution-Zoo

Collection of pre-trained super-resolution models in MXNet.
131 stars 17 forks source link

MoePhoto official dehazing result have some white line with mx.Predict. #4

Closed Ken1256 closed 5 years ago

Ken1256 commented 6 years ago

http://www.imagebam.com/image/4d46c41015403564

Also How to Converting pre-trained pytorch .pkl model to MXNet?(Try MMdnn few hours but no luck)

WolframRhodium commented 6 years ago

Hi,

I have checked the dehazing result produced by official MoePhoto in PyTorch and it does produce white line at the left edge of the image. Considering that the network is much lightweight, I suggest to feed the whole image to the network to reduce such artifacts. Adding borders to the image before prediction might also help.

Note: I find that I forgot to preprocess the image in the symbol file. You can try this to test the model. The params file is fine.

I convert PyTorch models to MXNet through ONNX. You can check these two tutorials: "PyTorch to ONNX", "ONNX to MXNet", and please feel free to contact me if you need any help.