This repository contains code for the 58th place solution in the Google Research - Identify Contrails to Reduce Global Warming competition on Kaggle. My solution write-up can be found here.
Use a different seed for each model regardless of the validation fold
- Seed changes data loading and initial model weights (which will result in more different models for the ensemble)
- Ex. 1st place solution for Feedback Prize - Predicting Effective Arguments Competition
Analyze model predictions
- Missed the 0.5 pixel shift that made TTA not work
- 5th Place Solution, 0.5 Pixel Shift Explanation
More Creativity, Trust your findings
- Could have trained on each participant, or on soft labels Discussionby @Ryches
- I found that bicubic worked best in CV, but did not stick with it..
- The best tips are not always on the discussion forums
Use checkpointing during training
- Enables the usage of pausing and resuming training runs without wasting computation
Team up with other Kagglers.
- This is the best way to learn!
- Efficientnetv2, DPT
- Losses: Tversky, LogCoshDice
- Downsampling Interpolation Methods
- Removing Islands
- Openmmlab (Upernet, Swin)
- Deep supervision
- Created a diverse ensemble using different backbones
- Attempted novel ideas (deep supervision, custom model architectures, etc)
- Iterated quickly with small-scale versions of images/models, and scaled up at the end
- First segmentation competition!