search

pfnet / PaintsChainer

line drawing colorization using chainer
https://paintschainer.preferred.tech/
MIT License
3.77k stars 556 forks source link

re-train model #13

Closed masato-io closed 7 years ago

masato-io commented 7 years ago

Hi taizan-san,

First of all, thank you for sharing your great project. I had a success running local server using cpu and played around with it with random images so far.

Now I'm wondering how I can start training additional image data as I have no knowledge of chainer...your description says " to train 1st layer using GPU 0 python train_128.py -g 0 to train 2nd layer using GPU 0 python train_x2.py -g 0", yet it requires "models/liner_f" which of course doesn't exist in this repo.

Could you elaborate little bit more on how to start training additional image data by any chance? I'm inspired by your project and now I would love to getting started with chainer journey 👍

liamjones commented 7 years ago

I'd also be interested to know how big the training data set was for the model you're using

masato-io commented 7 years ago

@liamjones according to this author's article, he mentioned that this model was trained by around six hundred thousand images. http://qiita.com/taizan/items/cf77fd37ec3a0bef5d9d

taizan commented 7 years ago

Okay thanks for using I uploaded liner_f to downloads but you can train models without liner_f . ( liner_f is src image to line drawing model. ) For re-training, you need line image folder color image folder image name list file image name in 2 folder should be same. and set it in line 67 in train_128.py dataset = Image2ImageDataset("dat/images_color_train.dat",root+"line/",root+"color/", train=True)

masato-io commented 7 years ago

@taizan Thank you so much for your quick response. I will definitely experiment re-training with your suggestion. Really appreciated :)

masato-io commented 7 years ago

As my question is clearly & kindly answered by author, I'm closing this ticket.