Train on my own dataset

[disturbed-mystic1]

Is there any implementation of this using an .ipynb notebook. An implementation of this would be very helpful.

[usert5432]

Hi @disturbed-mystic1,

Unfortunately, we have not developed any notebooks for this project.

However, if you wish to train uvcgan on your own dataset, you could try taking one of our scripts as a starting point, e.g. this one should work. And modify it to your needs. We will be happy to assist you if you run into any problems.

[Deepika1498]

What does align_train = false mean in the training script?

[usert5432]

Hi @Deepika1498,

What does align_train = false mean in the training script?

This option makes sense only when you have an aligned dataset. An aligned dataset, is the dataset where all images are paired. That is, for each source image you know how its translation should look like. An example of the aligned dataset is a dataset of satellite images (domain A) and the corresponding maps (domain B), as illustrated here.

If you have an aligned dataset, setting align_train = True will load the images in pairs. So, the network can see exactly how the source and translation should look like. If you set align_train = False, then the images will be shuffled, and so the network will not know how the translation of the image should look like.

If the dataset is not aligned (like CelebA), this option is largely irrelevant and can be set to any value.

[Deepika1498]

Can uvcgan be used to perform image fusion of satellite images to improve the spatial resolution of the landsat images by training on 30m landsat and 10m sentinel data?

[usert5432]

Hi @Deepika1498, I am not familiar with the problem of image fusion of satellite images, so I am not able to answer this question.

I may be able to provide more information on the feasibility of the UVCGAN usage for the problem of image fusion, if you could elaborate on your training setup?

[MajidHeydari]

I want to set up with satellite dataset with this:http://weegee.vision.ucmerced.edu/datasets/landuse.html