search

ljsabc / MangaLineExtraction_PyTorch

The (Official) PyTorch Implementation of the paper "Deep Extraction of Manga Structural Lines"
MIT License
149 stars 12 forks source link

Training Code #5

Open giantstonex7 opened 2 years ago

giantstonex7 commented 2 years ago

Hello, as you mentioned the training code would be ready soon, what's the progress so far, may i help?

zccoder commented 2 years ago

同上，有训练代码嘛

ljsabc commented 2 years ago

The training and tone filling code was written like 6 years ago so there might be some problems when running. We are not maintaining it anymore so please try solving these issues yourself. For line drawings, you can gather your own dataset by searching "line illustrations" or something. We cannot share that to you so sorry. But any common line drawings can produce quite promising results. We only include the filling of screentones in the following link. It's badly written in python2 so I guess you may need some adoption if you need to generate on-the-fly.

DL: Erika.zip

BTW the secret santa dataset is from deviantart but I cannot find the credit.

To train the model you may need to create a PyTorch dataloader yourself with the model code (should be easy).

It includes the essential things for reproducing the model when you have the dataset ready. It's suggested to train with > 5000 tones line drawings. It would be even better if you can include some cherry-picking on samples (by removing the outliers and badly looked samples).

Many thanks.