Pretrained Model

[Zheng222]

@Vious Hello, can you provide the pretrained model? Thank you!

[sadbhawnathakur]

@Vious Pretrained model link is not mentioned in the description.

[Vious]

Oh, sorry for my fault, I forgot to update the link of pretrained model into the README file, for the pretrained model, please kindly refer to the link of google drive: https://drive.google.com/drive/folders/17DJEp0APDcZoWPIHwMFs78ak59pYlDPw, or Baidu Cloud: https://pan.baidu.com/share/init?surl=z9hNEpbhsK2r-TEJbYkuQw, extract code is: mvzh

[hastaluegoph]

@Vious I download the pretrained model , but I found it go wrong when load the model. It may be something with the BN weights key, should I change the code?

[Vious]

@hastaluegoph I made a slightly change of the model by removing the batchnorm layer (setting bn to false) and the last tanh layer has been changed from absolute value to (tanh() + 1) / 2. I've updated the model files now or you could use the model files in the given link, I think that would be right.

[hastaluegoph]

@Vious Thanks a lot! I modified code under your instruction and test on places2 , but the result turn out to be worse than the Paper's,.Should I change some parameter and retrain the model?

[Vious]

@hastaluegoph For places2 model, due to the computational limitation, I only use 10 classes for training, including apartment_building-outdoor, beach, house, ocean, sky, throne_room, tower, tundra, valley and wheat_field, you could use the above 10 validation sets for evaluating.