Inpainting project done on https://paperswithcode.com/dataset/ffhq with custom inpainting. ( right now done on 128x128 - for testing purposes, find the thumbnails on: https://drive.google.com/drive/folders/1tg-Ur7d4vk1T8Bn0pPpUSQPxlPGBlGfv)
Table of end points:
| project feature | points awared | status |
|---|---|---|
| Project type: inpainting | 3 | ✅ |
| Additional problem: optimization of generator and discriminator training | 1 | ✅ |
| Model: pre-trained model on the different problem | 1 | ✅ |
| Model: non-trivial solution | 1 | ✅ |
| Dataset: > 10000 photos | 1 | ✅ |
| Dataset: own part > 500 | 1 | ❌ |
| Training: hyperparameter tuning | 1 | ✅ |
| Training: architecture tuning (at least 3 architecture) | 1 | ✅ |
| Training: overfitting some examples from the training set | 1 | ✅ |
| Training: data augmentation ( different mask on training) | 1 | ✅ |
| Training: cross-validation | 1 | ❌ |
| Training: testing a few optimizers | 1 | ❌ |
| Training: testing various loss functions | 1 | ✅ |
| Additional: Neptune | 1 | ✅ |
| Additional: Docker | 1 | ❌ |
| Sum | 18 | ❌ |
9.01.2023 Task 1: ✅
- try basc inpainting model (true model will be developed later in time)
- download ułomny dataset 128x128
11.01.2023 Task 2:
- Construct basic architecture for experiment ( for example model type + optimizer)
- Do overfitting experiment (architecture)
- Discuss and research possible models + look into using a pretrained one
- discuss git for large data...
11.01.2023 Task 3:
- Masking specific parts of face architecture (using pretrained object detector, objects being parts of face)
- See if masking specific face parts (mouth , eyes nose etc) makes sense
- MAT: Mask Aware Transformer for Large Hole Image Inpainting | CVPR 2022