The model for trained on kaggle GPU kernel for 300 epochs
- tensorflow==1.8.0
- opencv-python==3.4.0.12
- Keras==2.1.6
- keras-contrib==2.0.8
- Pillow==4.4.0
- numpy==1.14.0
-
Prepare data: run
prepare_data.ipynb
-
Build Model: run
OutPaint.ipynb
- Data is not generalized since in some of the images the missing parts lack relation with the maksed part. Can be improved by changing
MASK_PERCENTAGE
in code but will also need to change theimg_size = 256
- Hence, after many epochs it tends to overfit as we see that model learns those representations which it should not like hills in below image which is missing in case of masked image.