UNet

Re-implementation of U-Net[Ronneberger, O.(MICCAI15)] in PyTorch. In unet.py, while UNetVannila follows the paper's architecture(i.e. convolution layers don't have padding.), UNet has convolutions which use padding. In this repo, we use UNet mainly.

TODO

• Tackle with "Shaded White removing problem"
  • Can't identify shaded white parts of items.
  • Higher Brightness degree?
• Dilated Conv?
• test.py

Done

• get_dataset
• get_dataloader
• get_model
• get_optimizer
• get_scheduler
• Trainer
• MSELoss
• BCE+Dice Loss(from kaggle)
  • Messed up with Loss values.
  • BCE-only works well. So, dice loss degrades the performance.
• Lessen the number of parameters(3rd place solution uses 8M UNet)
  • Works well.(not better.)
• Double Check duplciation between train and test data.
  • Using diffrent ids b/w train and test data.
• BCELoss
  • In the literature, every paper uses bce rather than mse. We stick to this.
• Add More DAs
  • Rotation(45)
  • Color Jitter(Brightness/Contreast/Saturation/Hue)
  • Gaussian Blur?
• From ConvBlock to ResBlock (in bottleneck.)
  • Ref:DeepResUNet
    • Use pre-act resblock for all blocks.(not only bottleneck)
• Add Random Grayscale to catch the shape of the objects
  • Not good...?
• Quantitative Evaluation.(Dice)