Dataset sizes

[bob80333]

Do your datasets have to be the same size? Or do they just have to be divisible by batch size? I was trying this repo out but ran into some issues where randomly the losses would become NaN.

My trainA has 33344 images and my trainB has 22933 images. All hyperparams were default except dataset paths/names.

[WonwoongCho]

Hi bob80333,
The number of data between different domains (trainA and trainB) is not necessarily the same.

Actually, I'm not sure it is due to the same issue,
but empirically, I observed NaN during training when my dataset is too noisy to learn a translation. For example, if trainA data is vector illustrations, it rarely has consistent patterns across the dataset, thus it fails to learn. Therefore, I recommend you to carefully watch your dataset first in order to verify that a domain has a coherent pattern.

Please report again if the issue is not settled.