8th place solution of kaggle czii competition

Quickstart

Install Dependencies

pip3 install -r requirements.txt

Training All Models

./get-data.sh # to get the data to train the models
./train.sh # to train the models
./soup.sh # to create model soup

Output Artifacts

• tiny unet model soup = checkpoints/tiny-unet/
• medium unet model soup = checkpoints/medium-unet/
• big unet model soup on denoised tomograms = sumo/checkpoint_finetune_denoised/checkpoint_finetune_denoised_soup.pth
• big unet model soup on all tomograms= sumo/checkpoint_finetune_all/checkpoint_finetune_all_soup.pth

Running this script does the following

• downloads data to data/
  • data/train for real experiments data
  • data/synthetic-data for real synthetic data
• train all models
  • pretrain and finetune tiny unet models, to checkpoints/tiny-unet
  • pretrain and finetune medium unet models, to checkpoints/medium-unet
  • pretrain big unet models, to sumo/checkpoint_pretraining
  • finetune big unet models on denoised tomograms, to sumo/checkpoint_finetune_denoised
  • finetune big unet models on all tomograms, to sumo/checkpoint_finetune_all
  • create model soup for each model trained

Key Assumptions

• we assume that user has the kaggle api setup as we use them to get the data from kaggle
• the data directory data/ is empty before ./get-data.sh is run
• all checkpoints directories are empty before any trainings are run

Inference

python ./05-inference.py # run the command for inference

• For inference all the tomograms should be placed under ./data/test/static/ folder
• The folder is not created by default and must be created by the user
• The result of inference is submission.csv

---

Dependencies

Hardware

User	OS	Memory	GPU
IAmParadox	Ubuntu	64Gb	RTX 4090
sirapoabchaikunsaeng	Ubuntu 22.04 + Ubuntu 24.04	64Gb + 500Gb Swap	RTX 3090 + RTX 4090
sersasj	Kaggle Kernel	-	2x T4
itsuki9180	TBD	TBD	TBD

Software

all software dependencies are listed inside requirements.txt, we did not do any additional software installations other than cuda and gpu drivers

---

Configuration Files

• config.yaml is the base config for tiny and medium unet models. As there's a lot of field in there, please refer to the comments inside the config file to see what each field does
• sumo/pretraining.yaml is the config for pretraining big-unet models
• sumo/finetune_denoised.yaml is the config for finetuning big-unet models on denoised tomograms
• sumo/finetune_all.yaml is the config for finetuning big-unet models on all tomograms