Design style result not good enough on 10,000 step.

[johndzengpft]

Hi, I have trained the design style by the following steps with default config:

1. Run python generate_data.py --style design
2. Run python extract_align_faces.py and pick 200 style images
3. Train content calibration network
4. Generated content calibrated samples
5. Run geometry calibration for both photo and cartoon
6. Run python train_localtoon.py with total 19962 source images & 79832 style images

The result of 10,000 step is not as good as official pre-train:

Did I do any step wrong? Thank you very much!

[menyifang]

First, ensure stylized images fed to CCN are high-quality and style-faithfulness as possible, manual picking may help. Second, more steps may need for sufficient training, we use 30w steps on average.

[watertianyi]

@menyifang
Excuse me, I want to ask you questions:

1. If you want to train data with a resolution of 1024X1024, do you need 30w epochs, and how long does it take to train on a GPU?
2. If training is interrupted, how to resume training from the interruption and continue training?
   3.The following is the training loss. I don’t know if it meets expectations?
   

4.The training was interrupted, and the resume_epoch in the configuration file was modified to start from 42999. The loss training 20k has not changed. Is this normal?