about the discriminator stucture
mashumaro opened this issue · comments
According to the paper, after encoder down-sampling, the shape of x should be h/8, w/8, or h/32, w/32. So I use the pre-trained model and 96x96 size, but the shape of x became 11,11, and 2,2, which ideally should be 12 and 3. Maybe a silly question, but I really curious about this.
By the way, the pre-trained model seems failed to translate the selfie2anime dataset, right?