4th place Solution, Google Research - Identify Contrails to Reduce Global Warming

1st stage, pseudolabeling

The training parameters assume that an instance with 4 gpus with 48GB are used. Due to usage of bfloat16 only Ampere+ generation is supported.

Bootstrap pseudolabels

At this stage 4 folds of UNets were trained for 3 different encoders

• 512x512 resolution
• encoders: EfficientNet V2L, MaxVit Base, EfficientNet L2NS
• loss: 50BCE + 0.2Dice
• Optimizer: AdamW, wd=1e-2, cosine LR 1e-4, 70 epochs
• Augmentations: HFlip, 90, 270
• each fold on a single GPU, mixed precision with bfloat16

Predict OOF ensemble with 4TTA

second round pseudolabels

Same params as for bootstrap step though with 50% during training pseudo labels were used (only for 0-3, 5-7 frames)

Final stage

Training

• using train/val split
• 768x768 resolution
• encoders: EfficientNet V2L, MaxVit Base, EfficientNet L2NS, EfficientNet V2XL,
• loss: 50BCE + 0.2Dice
• Optimizer: AdamW, wd=1e-2, cosine LR 1e-4, 70 epochs
• Augmentations: HFlip, 90, 270
• DDP, mixed precision with bfloat16

SWA

The last step is SWA to average 5 best checkpoints for each model

How to reproduce

Inference is implemented in Kaggle kernel Kernel inference

All the steps for training are defined in a single script train_4folds_and_final which requires only a single parameter with path to the root dataset directory.

train_4folds_and_final.sh <you dataset dir>