Open ak9250 opened 3 years ago
Hi, our method is more suitable to more close-domain adaptation since it leverages the pretrain features heavily, so settings like Cars -> Faces will not work.
@e-271 thanks, so for close-domain adaption, n=5 should work, like faces-> faces sketches?
Hi I tried finetuning from n=5 with a datasets of cars from faces but the results after 28 kming![fakes000028 (5)](https://user-images.githubusercontent.com/6818285/97200349-796ad380-1787-11eb-92fb-45ff6b9d5276.png)
i trained faces to sketches with n=84 and that seemed to work better![fakes000030 (2) (1)](https://user-images.githubusercontent.com/6818285/97200468-a15a3700-1787-11eb-8bb1-c085489016d4.png)