Memory-Gated Recurrent Networks (AAAI 2021)
arXiv: https://arxiv.org/abs/2012.13121
Python == 3.6. Pytorch == 1.5. By default, the codes and the pre-trained models run on GPUs.
Simulated data is stored in simulation_data/data_generation_matlab/simulation_data in the CSV format. The first two columns in each csv file contain the observations of
To train and evaluate new models, please refer to NN_main_simulation_regression.py. To validate pre-trained models, please refer to validation_regression.py. To reproduce table 1 and figure 2 in the paper, please refer to simulation_results_summary.ipynb.
Please download the MIMIC-III data set, and follow the instructions in the mimic-3 repository to preprocess the data. You should finish all the steps in Section Building a benchmark and Train / validation split. After these steps, you should be able to run main.py of the four tasks in the mimic-3 repository.
To accelerate training, we save the processed data into the NPZ format. We are not able to directly provide these files due to the constrained usage of the MIMIC-III data set. Please run {task}_data_preparation.py in folder mimic3_utils for each task. A folder named by the task will be created under mimic3_utils. For example, for the in-hospital mortality task, please run mimic3_utils/ihm_data_preparation.py. Note that these files make use of multi-processors by default and are time-consuming.
To train new models, please refer to NN_main_mimic3_{task}.py of each task. For example, for the in-hospital mortality task, please run NN_main_mimic3_ihm.py. The hyper-parameters used to obtain results in the paper are saved in the respective NN_main_mimic3_{task}.py of each task. We repeat each experiment three times and the result with the best validation performance is reported. To validate pre-trained models, please run validation_mimic3_{task}.py of each task.
The UEA data sets can be downloaded here. You may run the uea_utils/uea_dataset_preparation.py to automatically download and preprocess the data sets. Folders named by data sets will be created under uea_utils.
The hyper-parameters used to obtain results in the paper are saved in uea_utils/param_dict.py. To train and evaluate new models, please refer to NN_main_uea.py. We repeat each experiment three times and the best result is reported. To validate pre-trained models, please refer to validation_uea.py.