The experiment still has a lot of flaws and needs to be perfected in the future.
A complete(almost) PyTorch Implementation of StyleBank: An Explicit Representation for Neural Image Style Transfer
Because there is not enough time, so this is still a bug version.
Now I face the following two questions:
-
The parameters in the paper can't run out of obvious results. I find the style loss is too low when I used paper's parameters (see the picture below).But you can change the
$\alpha / \beta$ 's value from 100 to 1000000, and the style loss is big enough to generate stylized image.
fig1 low style loss result
-
A high-resolution style dataset need to be collected. I personally think that the resolution of the style image has a great effect on the result. But because it's the end of the semester, I have too little time.QAQ
BTW, I've ran the results of the experiment, but it's not satisfying. See it below.
The dataset is COCO_2017val(my computer's hard disk space is not enough for a lager dataset), I chose the first 1000 pictures from this dataset for training.
And you can see some result in generate_image folder, the codes is in codes folder.
QZQTechMonster gave me this great help in this experiment.