custom dataset

[lzzlxxlsz]

Hello，I set the same structure of the data ,and the settings:

    self.parser.add_argument('--dataset', default='dfh', help='folder | cifar10 | mnist ')
    self.parser.add_argument('--dataroot', default='.data/dfh/', help='path to dataset')

my dataset name is dfh,and the file of dfh contains two folders : train and test.the train file contains one file of 0 , 0 file include the normal pictures .

However,when I run the train.py ,it turns the error: FileNotFoundError: [WinError 3] 系统找不到指定的路径。: '.data/dfh/train'

thanks

[Coderxiaobai1]

you do not need to set the default option --dataroot to '.data/dfh' , just let it be ' '.

[Coderxiaobai1]

of course, it depends the generate's input setting. if not , it will never come out of the while loop

[Coderxiaobai1]

In fact, i think it must be a series like [4,8,16,32,64,128,256,512...]

[lzzlxxlsz]

thanks ,another question.If my dataset size is not the multiple of 16,before I train the modle ,should I adjust the dataset size to be the multiple of 16?

In fact, i think it must be a series like [4,8,16,32,64,128,256,512...] 你好，你在自己的数据集上跑通 了吗

[Coderxiaobai1]

there is no need to do that, you can just set --isize to 128 as long as it is smaller than 200.(your dataset)

[lzzlxxlsz]

thanks，how about the result for your dataset?

[Coderxiaobai1]

i haven't done much since i'm still exploring it.

[lzzlxxlsz]

OK,thanks a lot

[Coderxiaobai1]

please let me know if you can run the demo well on your dataset cause i am haunted much these days.😭

[lzzlxxlsz]

I run the train.py but there is nothing display.only the output folder has some loss_log.txt. does it mean runs succussfully?

[Coderxiaobai1]

you can start the visdom server to have a look when in the training process

[Viki-researcher]

you do not need to set the default option --dataroot to '.data/dfh' , just let it be ' '. hi，why in my train the AUC is always nan,but the train generate picture is not too bad

[Linranran]

@17695996717 because there are only OK or NG samples in the test.

[Linranran]

@17695996717 the test should have the NG and OK samples simultaneously