Environments and Requirements

• Windows/Ubuntu both works
• gpu memory>= 8GB required
• CUDA version testted on 11.1
• python=3.8

To install requirements:

pip install -r requirements.txt

Dataset

• Download data from https://neurips22-cellseg.grand-challenge.org/dataset/
• folder structures will be describe in preprocessing section

Preprocessing

Running the data preprocessing code:

python create_annotation_for_yolo.py --input_imgs_path <path_to_input_data> --input_masks_path <path_to_input_data> --output_path <path_to_output_data>

Then organize your data dir as follows to fit the training yaml:

data
--images
----train
----val
--labels
----train
----val

Training

1. To train the model(s) in the paper, run this command:

cell.yaml is provided

download yolov5s6.pt from https://github.com/ultralytics/yolov5/releases/download/v7.0/yolov5s.pt

Then run

python train.py --img 1280 --batch 8 --epochs 50 --data ./cell.yaml --weights yolov5s6.pt --multi-scale --seed 42 --name fold_1

You can download trained models here:
https://drive.google.com/file/d/1ty6jXk-ZnuA4f23ZvlLTvRdWhVY8-JM3/view?usp=drive_link

https://drive.google.com/file/d/1p42Wm1zEHprFeDW8KxgYJa7PYsGnL4GZ/view?usp=drive_link

https://drive.google.com/file/d/1NkbHfEZWdMAZBR8toWk4Nvf_c3srd-Cn/view?usp=drive_link

https://drive.google.com/file/d/1n7o7dvIygyI-JjECKRwW2hkl4puitWvp/view?usp=drive_link

https://drive.google.com/file/d/1D2Ynk9loeJZHAKjS351na7f-6wrIetjh/view?usp=drive_link

2. To fine-tune the model on a customized dataset, run this command: first preprocessing data,

python create_annotation_for_yolo.py --input_imgs_path <path_to_input_data> --input_masks_path <path_to_input_data> --output_path <path_to_output_data>

then train the model

python train.py --img 1280 --batch 8 --epochs 50 --data ./youdata.yaml --weights yolov5s6.pt --multi-scale --seed 42 --name your_run_name

Inference

1. To infer the testing cases, run this command:

python custom_det.py --img 1280 --source patched_cache --weights runs/fold_4.pt runs/fold_3.pt runs/fold_2.pt runs/fold_1.pt runs/fold_0.pt --name testa --max-det 20000 --half --iou-thres 0.5 --conf-thres=0.4 --save-txt --save-conf --line-thickness 1 --hide-labels --project patched_cache/detect --nosave

docker container run --gpus "device=0" -m 28G --name algorithm --rm -v $PWD/CellSeg_Test/:/workspace/inputs/ -v $PWD/algorithm_results/:/workspace/outputs/ algorithm:latest /bin/bash -c "sh predict.sh"

2. Colab jupyter notebook https://colab.research.google.com/drive/1hrsotfDiyyWpdlWIY2JpkzNLjURE5joG?usp=sharing
3. Docker containers on DockerHub https://hub.docker.com/r/baibizhe/overoverfitting