Training

1. Prepare training data : -- download CelebAMask-HQ dataset

2. Preprocessing (to stack all the mask images of a person into one)

python prepropess_data.py

2. Train

python train.py

• Current model (Unet)
• loss function: $Loss = cross entropy + dice loss$
• evaluation metric: mean IoU
• hyperparamters and global variables are in configs.py
• beaware of the arguments in the scheduler, unexpected result are produced if the scheduler is initialized unappropriately
• a jupyter notebook is also created for visualization convenience.

Inference

python test.py

• 60 comparison results are generated for better visualization of the model performance

Notes

• model weight is too large, can't push to gihub
• segmentation model pytorch library might be useful for building different model architecture and applying pretrained weight.
• this reference given by TAs shows lots of performance results by using different model architecture
• this library includes more architectures, but seems quite difficilt to use
• 10 samples from Unseen dataset
• FaceSynthetics dataset (optional)