Tamar Rott Shaham
Tamar Rott Shaham
Hi, The built-in torch interpolation function does not perform any anti-aliasing filtering when down-sampling the image, and this is the reason for using this function. This is especially important when...
Another easy option to try is maybe downsampling with the current image resize function and then upsampling with the build in torch function.
Since generation is done with random walk around z_opt, which has the same spacial dimensions as the training image, you can't change the spacial dimension.
you should put the image you want in the folder "Input/Images" and the call the training command, but replace "" with the file name
In the paper we compared a *single* image sample with it's corresponding real image, for 50 different images (you can find all the data in the Download folder, we took...
I just checked the code again: I get exactly the same numbers as in the paper without changing anything in the code.
seems like a problem with your CUDA drivers, not with the code
--padd_size control the padding of each conv-layer. because we ended up using an initial padding and not layer padding (see fig. 1 in the [SM](https://tomer.net.technion.ac.il/files/2019/09/SingleImageGan_SM.pdf)), with the current model --padd_size...
you can only increase the batch size of the fake samples, because for real samples you only have a *single image*. This will significantly affect the min-max optimization equilibrium and...
min_size and max_size effect the training. These are the minimal/maximal image dimensions over the model's levels. The number of level is adjusted accordingly.