Open yaxingwang opened 6 years ago
what is the dataset ? can you provide this dataset ? did you crop only the animal face ?
Thank you @taki0112, I use the dataset online where you can download and crop the animal face based on the annotation and resize 128*128, and also collect from the other website so that the number of every domain is about 5000. Now I keep training until the iteration is more than 1M, then I will share the final result with you.
Besides, did you try another dataset? what is result during training?
I tried Selfie <-> Comic character
, But the image was not distorted like you.
Ok, let's wait for the final result from my experiment. If you can try, I can share the dataset with you later.
Hi, I have got the latest result which is as following: cat--->dog:
dog--->cat:
There are some reasons for this result is like:
the domain A and B must be similar. IN my experiment, I just crop a dog and cat face, but I do not select the face which should be aligned (similar pose, shape an face scale)
It is overfitting from dog to cat for the output are same whatever the input is when testing the model. so I don't know the author use 1M iterations.
The real parameter for the animal could be different from other, for which the author did share.
@yaxingwang can you share the hyper-parameter for the animal ?
Hi, I use the same parameter from your file for the animal. I guess the hyper-parameter could be different.
@yaxingwang hi, can you please share me a copy of dataset which has mentioned above, if it is convenient to you. I am very interested in it, I promise you I will never distribute it without your permission, and I will only use for academic research. My email is huangsh6@mail.sustc.edu.cn, I will be very appreciate it.
@yaxingwang Hi, can you please share me a copy of dataset which has mentioned above, if it is convenient to you. I am very interested in it, I promise you I will never distribute it without your permission, and I will only use for academic research. My email is c1a1o1@qq.com, I will be very appreciate it.
thank your for your sharing, i trained three models: cat2dog,summer2winter, cityscapes,but all failed after 1000000,1000000, 960000 iters,i didn't change any hyper-parameter :python main.py --phase train --dataset cityscapes --batch_size 1; dataset: cat2dog is 178/218, trainA have 771 imgs and B 1264, i do not think they are aligned;and summer2winter(summer2winter_yosemite) is 256/256,trainA have 1231 imgs and B 962, i do not think they are aligned too;cityscapes is std dataset download from cityscapes; my results: cat -> dog summer2winter cityscapes
i didn't know what's wrong? need more iters or some code changes,, any ideas? thank you very much @taki0112 @yaxingwang
Hi @xiang-zhe, I tried the official code of MUNIT, it works well. I also experienced a similar condition to you. The following translation is from cat2tiger Maybe there are tiny parts with the official project.
The cat2tiger is also tried on this project(taki0112), which fails.
@yaxingwang thank your very much, i will try it by torch MUNIT, but in fact i want to perform all my practice under tf, thank you again! addditionally, aligned and resize are must?
Hi I trained the code with the animal dataset(cat and dog) the following results are from 400000 iteration