solution.md

Summary

• 3D encoder and 2D/1D encoder

  • 1/2 or 1/4 resolution prediction

  • very simple decoder

Data preprocessing

• split the 2nd fragment into two image vertically: 4 folds

Model

• 3D CNN encoder and 2D Encoder
  • encoder: based on @samfc10 's Notebook: Vesuvius Challenge - 3D ResNet Training
    • remove maxpooling after the 1st CNN
    • use attention before reduce D-dim
    • use resnet18 and resnet34
  • decoder
    • a single 2D CNN layer
    • upsample with a nearest interpolation
  • output resolution is downsampled to 1/2. then upsample with a bilinear interpolation
• 3D transformer encoder and linear decoder
  • encoder: use MViTv2-s of the PyTorch official implementation and a pre-trained model
    • modify forward function to get each scale output
    • replace MaxPool3d into MaxPool2d for the reproducibility
  • decoder: a single linear and patch expanding to upscale low resolutions 3D images
    • patch expanding is from Swin-Unet
  • output resolution is downsampled to 1/2 or 1/4, then upsample with a bilinear interpolation

Training

• amp
• torch.compile
• label smoothing
• cutout
• cutmix
• mixup
• data augmentation
  • referred @ts 's notebook: 2.5d segmentaion baseline [training]
  • referred @tattaka and @mipypf 's
• patch_size=224 and stride=112
  • stride=75 or stride=56 didn't work

Inference

• fp16 inference
• stride=75
  • better than stride=112
• ignore edge of output prediction
  • use only the red area of prediction (figure below)