Wgan on custom data

i tried to use my own dataset for lesson-7-wgan notebook. when i used image size of 64, i encountered no error, but when the image size was increased to 120 am having some error

when pass through to the leaner i was getting this error

please your help

I believe the image size needs to be a power of 2 (2,4,8,16, etc)

The problem is due to some conv layer. It is difficult to tell the exact soln but you can easily check. The idea is the the kernel_size is (4x4) is less than that.

Thanks. I think 120 is a factor of 2 ?

Is there a way to change the kernel size

Changing kernel size is not possible as it means you are changing your model arch i.e. creating a new model.

He meant powers of 2. But it is somewhat wrong. The image size should be 7*2x, if you are to look at various imagenet models like Resnet, VGG, and mostly all others.

But size 64 is working but am getting a noisy image

i get the idea now. in the notebook 64 is use which is 8^2. and if am to use a more large size, i willneed to use 16^2 which is too large.

They need to be multiples of 16. So 128 should work.

1 Like

size 128 works. Thanks
Apart from the image size, does the wgan need a very large dataset. am using about 1456 images. Am getting a very noisy image has output