HPA-singlecell-2nd-dual-head-pipeline

Solution overview at Kaggle

This pipeline achieve 0.55011 in both private and public leaderboard

I. HARDWARE:

Train:

• Ubuntu 20.04.1 LTS (~200GB free disk space)
• CPU: 8C
• RAM: 64GB
• 1 x RTX 3090

Predict

• Please refer to kaggle notebook sub1

II. SOFTWARE

• Please refer to Dockerfile
• Directly run container: docker pull steamedsheep/hpa_pipeline1:v1.15.

III. Data Setup

1. Make fold input and results at same dir of the repo.

2. Download pregenerated cells from kaggle notebook and decompress to input/train_cell_256, after decompression, there will be about 70GB png file in input/train_cell_256, since It's diffcult to use Kaggle API to download notebook output, Please download result.zip of each notebook below:

• train cell #1
• train_cell_#2
• Phil upload images #1
• Phil upload images #2
• Phil upload images #3
• ext not in Phil but rare

IV. Train and predict.

1. Start docker at solution dir sudo docker run -v $PWD:/workspace -it --gpus '"device=0"' steamedsheep/hpa_pipeline1:v1.2
2. /bin/bash train.sh

Full training

• run train.sh to get the result