Giters

SerialLain3170 / adeleine

Automatic line art colorization using various types of hint or without hint

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Training atari_userhint_v2

HanzVivatma opened this issue · comments

commented

Thank you for sharing such a great model.

I am about to train atari_userhint_v2 model from scratch but being confused with dataset.
Following the instruction, there is need of 3 types of datasets. I dont know how to obtain 2 types of datasets below:

  • SKETCH_PATH: The name of the directory that contains line arts obtained by SketchKeras
  • SS_PATH: The name of the directory that contains quantized color images obtained by superpixel

Could you please share a model to get sketch and ss dataset as well as give me a detail of instruction for training model from scratch?

I will be grateful for your assistance.

commented

Thank you for the issue, and I am sorry for insufficient instructions.

  • When it comes to SS_PATH, you do not have to care about that because the path is not actually used in the code. So, you can set any kind of path as SS_PATH. I recommend that you can set the same path as SKETCH_PATH or DATA_PATH.
  • As for SKETCH_PATH, the path contains line arts corresponding to the images in DATA_PATH, and the line arts are obtained by running SketchKeras. You can use the scripts in the repository and modify aligning with your purpose.
commented

Sorry for the late reply.

The dataset I mainly used for training is this. If you share the result, it might give me more ideas about the suitable comments.

commented

Sorry for the late response.

I would like to know whether the model is overfitting or not. Could you check the colorized result of training dataset? If the result is appropriate, I think that the model is overfitting. In that case, you might have to increase the number of dataset. If not, I would like to know the modifications that you made in the code to run the script (e.g. hyperparameter settings in param.yaml).

commented

Thanks for your valuable assistance.

Currently, I would like to train model for Flatten hints (colorization based on scribble hints in GUI application).
Could you please suggest that what type of model I should train: atari_diverse, atari_spade, atari_userhint or atari_whitebox?

I am looking forward to your reply

commented

Sorry for the late response. I think that atari_whitebox is the reasonable choice. atari_userhint is the straightforward way to utilize scribble hints, but the code would not be fully organized.