johnelvis / ecg-tianchi Goto Github PK

中文版readme

source: “合肥高新杯”心电人机智能大赛

Round 1 F1-score: 0.8491 Round 2 F1-score: 0.9228 In round 1, I rank (5/2353). But my final result failed to be submitted in round 2 due to some committee reason, so I do not know the final online score.

I treat this as 34-class classification (in round 1, it's 55-class). I use ResNet 50 / 101 and ResNext 50.
This are the scores of all models on offline validation set, which rely on to give the weights in model ensembling.

The age and gender feature have been added to above models. I also test those models which do not have the two features. Although the validation set score of models without these two features will be lower, the online score for all models after ensemble can increase.

ensemble score:

Yet the highest score on validation set is 0.9226.

(*basic means no age and gender features)

Hardware: NVIDIA V100 (32 GB) GPU * 1 Operating System: CentOS Linux 7 Software: Python 3.7, PyTorch 1.1

a. Create a conda virtual environment and activate it.

conda create -n tianchi python=3.7 -y
conda activate tianchi

b. Install PyTorch stable or nightly and torchvision following the official instructions, e.g.,

pip install requirements.txt

c. data download

cd data
./data_download.sh

d. data prepare

cd code
python data_preparing.py

In the data enhancement section I referenced the following code: https://github.com/JavisPeng/ecg_pytorch?spm=5176.12282029.0.0.3d952737ec5tuc

I used weighted binary_cross_entropy. The weight depends on the amount of every class.

weight = 1 / log( sum(classes)  + 1e-5)

cd code
./train.sh configs/dev_testA/ResNet50.yaml
./train.sh configs/dev_testA/ResNet50_noweight.yaml
./train.sh configs/dev_testA/ResNet101.yaml
./train.sh configs/dev_testA/ResNeXt50.yaml
python configs/ensemble.py

To train models without age and gender featrue, modify model name in yaml file to 'modelname+_Basic'

python main.py --configs=configs/ensemble.yaml -v

python main.py --configs=configs/ensemble.yaml -e

If you have any questions, feel free to connact: Yuzhe Zhou HongKong University Email: yuzhe36 DOT connect DOT hku DOT hk