pingchang818/tdstf

TDSTF

This is the github repository for the paper "A Transformer-based Diffusion Probabilistic Model for Heart Rate and Blood Pressure Forecasting in Intensive Care Unit" (https://doi.org/10.1016/j.cmpb.2024.108060)

MIMIC-III data

Download the dataset at

https://physionet.org/content/mimiciii/1.4/

Environment

Anaconda3-2023.03-0-Windows-x86_64

Pytorch=2.1.1 + cuda=11.8

Data preprocessing

Create empty folders: "/save", "/preprocess/data", "/preprocess/data/MIMICIII", and "/preprocess/data/first"

Download the MIMIC-III data to "/preprocess/data/MIMICIII"

Run the files step_1.py through step_4.py in order in the folder "/preprocess"

Experiments

Run the file main.py

To test a pretrained model, please assign the model folder name to the parameter "modelfolder"

Acknowledgements

A part of the codes is based on CSDI and STraTS

tdstf's Issues

Memory Error

I tried to run the code. In data preprocessing step2 a 4.23G sets.pkl file is generated,when reading this file 【sets = pickle.load()】in step3， a memory error is reported. 【My computer: GeForce RTX 3090, 3TB.】how can I fix this? ~👀

gt=samples_y[:,2] ?【the calculation of MSE】

I'm sorry to bother you due to a question I don't understand, the value of "samples_y" in the "forecast" function has been changed and is no longer equal to the original ground truth【pic 1】, why is gt=samples_y[:,2]【pic 2】. In that case is the MSE calculation no longer correct?

Unable to get results reported in the paper

Hello @PingChang818 , thank you so much for providing the source code for the interesting paper. I've been trying to reproduce the results from the paper, but haven't been able to do so.

So I followed the readme file, downloaded the data, ran through preprocessing, and did training and evaluation. There wasn't any problem aside from getting the paths right for the dataset.

I trained the model twice with python main.py, and evaluated each checkpoint twice. That gave us 4 results:

model 1 run 1
CRPS: 0.7520603882639032
MSE: 622.348388671875

model 1 run 2
CRPS: 0.5939008311221474
MSE: 1333.6546630859375

model 2 run 1
CRPS: 0.830156125520405
MSE: 232.1817626953125

model 2 run 2
CRPS: 0.7796238849037572
MSE: 774.9886474609375

I understand that there is some inevitable randomness in diffusion models, but these results seem far away from Table 2. Is it correct to use hyperparameters in config/base.yaml? Is there something else that I'm missing? Thanks!