Sartorius - Cell Instance Segmentation

https://www.kaggle.com/c/sartorius-cell-instance-segmentation

Environment setup

Build docker image

bash .dev_scripts/build.sh

Set env variables

export DATA_DIR="/path/to/data"
export CODE_DIR="/path/to/this/repo"

Start a docker container

bash .dev_scripts/start.sh all

Data preparation

Download competition data from Kaggle
Download LIVECell dataset from https://github.com/sartorius-research/LIVECell (we didn't use the data provided by Kaggle)
Unzip the files as follows

├── LIVECell_dataset_2021
│   ├── images
│   ├── livecell_coco_train.json
│   ├── livecell_coco_val.json
│   └── livecell_coco_test.json
├── train
├── train_semi_supervised
└── train.csv

Start a docker container and run the following commands

mkdir /data/checkpoints/
python tools/prepare_livecell.py
python tools/prepare_kaggle.py

The results should look like the

├── LIVECell_dataset_2021
│   ├── images
│   ├── train_8class.json
│   ├── val_8class.json
│   ├── test_8class.json
│   ├── livecell_coco_train.json
│   ├── livecell_coco_val.json
│   └── livecell_coco_test.json
├── train
├── train_semi_supervised
├── checkpoints
├── train.csv
├── dtrainval.json
├── dtrain_g0.json
└── dval_g0.json

Training

Download COCO pretrained YOLOX-x weights from https://github.com/Megvii-BaseDetection/YOLOX

Convert the weights

python tools/convert_official_yolox.py /path/to/yolox_x.pth /path/to/data/checkpoints/yolox_x_coco.pth

Start a docker container and run the following commands for training

# train detector using the LIVECell dataset
python tools/det/train.py configs/det/yolox_x_livecell.py

# predict bboxes of LIVECell validataion data
python tools/det/test.py configs/det/yolox_x_livecell.py work_dirs/yolox_x_livecell/epoch_30.pth --out work_dirs/yolox_x_livecell/val_preds.pkl --eval bbox

# finetune the detector on competition data(train split)
python tools/det/train.py configs/det/yolox_x_kaggle.py --load-from work_dirs/yolox_x_livecell/epoch_15.pth

# predict bboxes of competition data(val split)
python tools/det/test.py configs/det/yolox_x_kaggle.py work_dirs/yolox_x_kaggle/epoch_30.pth --out work_dirs/yolox_x_kaggle/val_preds.pkl --eval bbox

# train segmentor using LIVECell dataset
python tools/seg/train.py configs/seg/upernet_swin-t_livecell.py

# finetune the segmentor on competition data(train split)
python tools/seg/train.py configs/seg/upernet_swin-t_kaggle.py --load-from work_dirs/upernet_swin-t_livecell/epoch_1.pth

# predict instance masks of competition data(val split)
python tools/seg/test.py configs/seg/upernet_swin-t_kaggle.py work_dirs/upernet_swin-t_kaggle/epoch_10.pth --out work_dirs/upernet_swin-t_kaggle/val_results.pkl --eval dummy

Solution of Kaggle competition: Sartorius - Cell Instance Segmentation

Related tags

Overview

Sartorius - Cell Instance Segmentation

Environment setup

Data preparation

Training

Owner

Neural style transfer as a class in PyTorch

WORD: Revisiting Organs Segmentation in the Whole Abdominal Region

[ICML 2020] DrRepair: Learning to Repair Programs from Error Messages

Self-Supervised Learning

MapReader: A computer vision pipeline for the semantic exploration of maps at scale

Official implementation of "Learning to Discover Cross-Domain Relations with Generative Adversarial Networks"

Code for project: "Learning to Minimize Remainder in Supervised Learning".

PyTorch implementation of NeurIPS 2021 paper: "CoFiNet: Reliable Coarse-to-fine Correspondences for Robust Point Cloud Registration"

Implementation of Nalbach et al. 2017 paper.

Python with OpenCV - MediaPip Framework Hand Detection

Constructing interpretable quadratic accuracy predictors to serve as an objective function for an IQCQP problem that represents NAS under latency constraints and solve it with efficient algorithms.

A pytorch reproduction of { Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation }.

Free like Freedom

Multiview Neural Surface Reconstruction by Disentangling Geometry and Appearance

PPO Lagrangian in JAX

Unleashing Transformers: Parallel Token Prediction with Discrete Absorbing Diffusion for Fast High-Resolution Image Generation from Vector-Quantized Codes

CHERRY is a python library for predicting the interactions between viral and prokaryotic genomes

Serve TensorFlow ML models with TF-Serving and then create a Streamlit UI to use them

An implementation of the "Attention is all you need" paper without extra bells and whistles, or difficult syntax

A lightweight face-recognition toolbox and pipeline based on tensorflow-lite