BasicSR
BasicSR copied to clipboard
XPixelGroup
Custom Size ESRGAN Training error
Trying to train ESRGAN with custom lr/hr size, get the following error:
2020-11-12 19:13:17,807 INFO: Model [ESRGANModel] is created.
2020-11-12 19:13:18,060 INFO: Start training from epoch: 0, iter: 0
Traceback (most recent call last):
File "train.py", line 252, in <module>
main()
File "train.py", line 211, in main
model.optimize_parameters(current_iter)
File "../basicsr/models/esrgan_model.py", line 41, in optimize_parameters
real_d_pred = self.net_d(self.gt).detach()
File "/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 532, in __call__
result = self.forward(*input, **kwargs)
File "../basicsr/models/archs/discriminator_arch.py", line 57, in forward
f'Input spatial size must be 128x128, '
AssertionError: Input spatial size must be 128x128, but received torch.Size([16, 3, 32, 32]).
I've already specified 32 size in yaml file, what's the work-around for this?
I have the same problem.The solution is to adjust the number of linear mapping nodes of the discriminator.
hello! I have the same problem. How did you modify it?
Currently, discriminator used in ESRGAN takes as input only images of size 128 x 128. So you have to modify basicsr/models/archs/discriminator_arch.py file so that it could be able to handle images of the size you need. To do that, remove assert on line 52 and change number of units in linear layer on line 45. It can also be useful to add some more Conv2D blocks.
Thank you very much!
------------------ 原始邮件 ------------------ 发件人: "xinntao/BasicSR" @.>; 发送时间: 2021年5月17日(星期一) 凌晨2:54 @.>; @.@.>; 主题: Re: [xinntao/BasicSR] Custom Size ESRGAN Training error (#329)
hello! I have the same problem. How did you modify it?
Currently, discriminator used in ESRGAN takes as input only images of size 128 x 128. So you have to modify basicsr/models/archs/discriminator_arch.py file so that it could be able to handle images of the size you need. To do that, remove assert on line 52 and change number of units in linear layer on line 45. It can also be useful to add some more Conv2D blocks.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.