search

mit-han-lab / gan-compression

[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Other
1.1k stars 148 forks source link

why the test result of the compressed gaugan trained by me is not as expected? but the training result of supernet is right. #71

Closed yuanlunxi closed 3 years ago

yuanlunxi commented 3 years ago

_ 下载 (2)

lmxyy commented 3 years ago

Sorry, I do not quite understand your problem? Could you please elaborate more. Which dataset are you using?

yuanlunxi commented 3 years ago

I used the dataset of myself,the dataset is correct, because the result of training process is correct.

yuanlunxi commented 3 years ago

in the fast gan-compression process, when i export the compressed the model from the supernet,should i use the super_mobile_sapde or mobile_spade?

yuanlunxi commented 3 years ago

the test output of head0 of compressed gaugan is very huge, image

yuanlunxi commented 3 years ago

so is there wrong step cause by me in the export process?

dearkafka commented 3 years ago

I bet there is a problem with input data normalization in your code (preprocessing of your dataset for compression)

yuanlunxi commented 3 years ago

I bet there is a problem with input data normalization in your code (preprocessing of your dataset for compression)

the input is normalized

yuanlunxi commented 3 years ago

I find the output of image is very huge

yuanlunxi commented 3 years ago

i find the problem, when the self.training in the SuperSynchronizedBatchNorm2d is set True ,the result of test is right. but can anyone explain this ?

seekingdeep commented 3 years ago

@yuanlunxi