Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Let's Do Training!!!! #18

Open
Ugness opened this issue Nov 12, 2019 · 15 comments
Open

Let's Do Training!!!! #18

Ugness opened this issue Nov 12, 2019 · 15 comments

Comments

@Ugness
Copy link
Owner

Ugness commented Nov 12, 2019

첫 실험 결과

image

  • 색 번짐.
@Ugness
Copy link
Owner Author

Ugness commented Nov 12, 2019

  • 고쳐야할 것
  • Line Image Interpolate 할 때 NEAREST로 해줘야 경계선이 유지되어서 잘된다고 함.

@Ugness
Copy link
Owner Author

Ugness commented Nov 13, 2019

image

  • nearest + Reflection Padding 했는데도 고쳐지지 않음.
  • VGG19가 imageNet pretrained인데 VGGLoss를 적용한 것이 문제 있을 수 있다고 생각함.
    • 그리고 VGG19 RGB 기준인데 Lab 이미지 넣어주고 있었음.
    • vgg loss term 끄고 학습해보기
  • SPADE Module이 Sketch + Hint에서 제대로 된 beta, gamma를 추정하고 있지 못하다고 생각함.
    • SPADE Module 앞 쪽에 Conv Layer 추가 + LadderNet

@Ugness
Copy link
Owner Author

Ugness commented Nov 13, 2019

  • Sketch min-max활용해서 힌트 뽑는 범위 / loss 주는 범위 등등 지정해서 흰색 바탕 Vanishing 해결하기.
  • Lab Output Clipping 하기.
  • Hint Patch Size 늘리기. (*2~4) -> Nearest 때문에 넣어 준 힌트가 대부분 죽을 것 같음.

@Ugness
Copy link
Owner Author

Ugness commented Nov 13, 2019

  • 돌릴 때 nThreads 16~64로 넣고 하면 더 빨리 돌아감.
  • SPADE Normalization module에 Conv Layer 추가하기.

@Ugness
Copy link
Owner Author

Ugness commented Nov 13, 2019

  • https://arxiv.org/pdf/1808.03240.pdf (AlacGAN)
  • 참고하기.
    • ImageNet Pretrained VGG16 이용해서 Content loss 줌.
    • Sketch + Hint + (Features extracted from sketch) 넣고 채색함.
    • penalty loss 라는 걸 추가해줬는데 공부해보면 좋을지도?

@Ugness
Copy link
Owner Author

Ugness commented Nov 13, 2019

image
그냥 기존꺼 오래 돌린 모델.
잘되는 듯?

@Ugness
Copy link
Owner Author

Ugness commented Nov 14, 2019

  • Sketch에서 Feature Extract하는 netF 네트워크 추가.
    image
    성능향상 잘 모르겠음. 캐릭터의 얼굴 / 몸 은 이전 모델보다 더 잘 구분하는 것 같기도 함.
  • 모델들이 Hint를 무시하고 채색하는 것 + 실제 같지 않게 색칠하는 것을 고쳐야할 것 같음.

@Ugness
Copy link
Owner Author

Ugness commented Nov 14, 2019

@Ugness
Copy link
Owner Author

Ugness commented Nov 14, 2019

@Ugness
Copy link
Owner Author

Ugness commented Nov 14, 2019

image
image
image

Hint Feeding 할 새로운 방법 찾기

@Ugness
Copy link
Owner Author

Ugness commented Nov 16, 2019

image
image
image
image

@Ugness
Copy link
Owner Author

Ugness commented Nov 16, 2019

NetL + BooruLoss + L2Loss + CosineAnnealingLR 학습 결과

image
image
image
image

@Ugness
Copy link
Owner Author

Ugness commented Nov 16, 2019

  • 가장 최신 옵션
python train.py --name Lnet_hsv_aug --dataset_mode safebooru --dataroot ../CS470_Project/data/safebooru/train_upper_body_768 --tf_log --no_html --batchSize 8 --gpu_ids 0 --sample_Ps 2 22 2 --use_vae --netG spadeladder --L2_loss --no_vgg_loss --booru_loss --lr 0.0002 --SGDR --hsv_aug 0.1

@Ugness
Copy link
Owner Author

Ugness commented Nov 17, 2019

image
image

  • hsv_aug 0.1 추가했더니 힌트가 잘 들어감.
    • 같은 스케치에 대해 여러 채색 정답이 있다는 것을 data aug로 알려줌.

@Ugness
Copy link
Owner Author

Ugness commented Nov 18, 2019

--high_sv, --hsv_tv loss 추가할 시 더 안좋아짐.
image

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant