• Download anaconda for your system

  wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh

• Run the installer bash Miniconda3-latest-Linux-x86_64.sh

• Create an environment conda env create -n sl python=3.6.4 -f environment.yml

• Activate the created environment source activate sl

IMPORTANT:

• Download the sample data first: cd data && bash download-sample-data.sh
• When training/prediciting the processed data is by default stored in <data_dir>/tmp/. To store the data at different location you must set the TMPDIR environment variable accordingly. This variable is set automatically on leonahrd when requesting scratch space. It is recommended to store the data in /cluster/scratch/ and request enough scratch space on the executing compute nodes.
• Be aware the required space for the processed data scales with the number of neighbors. Additionally for the Physionet18 dataset there is an additional factor 4 (!) from the conversion from int8 (raw data) to float32.

All preprocessing steps, model architecture and parameters are saved in the model file. Just specify a saved model and subject file/directory.

First activate the conda environment and change to the bin directory:

conda activate sl && cd bin/

Then specify a model and subject to predict. The prediction is saved to .csv.

E.g. for Sleep-EDF
python classify.py --model ../models/cv_sleepedf_Fz_2D_singlechanexp2_6464962FC_MP/fold0/checkpoint.pth.tar --subject ../data/sleepedf/SC4001E0-PSG

E.g. for Physionet18

python classify.py --model ../models/3067_LateFusion.pth.tar --subject ../data/physionet-challenge-train/tr11-0064/

First activate the conda environment and change to the bin directory: conda activate sl && cd bin/

Training the single channel expert model local on the F3M2 channel:

python train.py with F3M2_2DP singlechanexp save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

On leonard we request 5 computing nodes with scratch space and a GPU:

module load python_gpu/3.6.4

bsub -n 5 -W 03:59 -R "rusage[mem=6000,scratch=25000,ngpus_excl_p=1]" ~/miniconda3/envs/sl/bin/python train.py with F3M2_2DP singlechanexp save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

Training the AMOE model local on all channels (6 EEG, 1 EOG, 3 EMG) with 10% channel dropout:

python train.py with PHYSIONET_EEG_EOG_EMG_2D AMOE_RS160 ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

On leonard we request 5 computing nodes with scratch space and a GPU:

module load python_gpu/3.6.4

bsub -n 5 -W 03:59 -R "rusage[mem=6000,scratch=65000,ngpus_excl_p=1]" ~/miniconda3/envs/sl/bin/python train.py with PHYSIONET_EEG_EOG_EMG_2D AMOE_RS160 ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

Training the Attention model local on all channels (6 EEG, 1 EOG, 3 EMG) with 10% channel dropout:

python train.py with PHYSIONET_EEG_EOG_EMG_2D AttentionNet_RS160 ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

On leonard we request 5 computing nodes with scratch space and a GPU:

module load python_gpu/3.6.4

bsub -n 5 -W 03:59 -R "rusage[mem=6000,scratch=65000,ngpus_excl_p=1]" ~/miniconda3/envs/sl/bin/python train.py with PHYSIONET_EEG_EOG_EMG_2D AttentionNet_RS160 ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

Training the LateFusion model local on all channels (6 EEG, 1 EOG, 3 EMG) with 10% channel dropout:

python train.py with PHYSIONET_EEG_EOG_EMG_2D LateFusion ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

On leonard we request 5 computing nodes with scratch space and a GPU:

module load python_gpu/3.6.4

bsub -n 5 -W 03:59 -R "rusage[mem=6000,scratch=65000,ngpus_excl_p=1]" ~/miniconda3/envs/sl/bin/python train.py with PHYSIONET_EEG_EOG_EMG_2D LateFusion ChannelDropout10 save_model=True ds.data_dir=../data/physionet-challenge-train/ ds.train_csv=../cfg/physionet18/physio18_debug.csv ds.val_csv=../cfg/physionet18/physio18_debug_val.csv

To reproduce the 20-fold CV results on the Sleep-EDF dataset (Fpz channel) run the following bash script in the bin directory: bash train-sleepedf-singlechanexp.sh --data_dir=<path to data>

To run the cross validation on leonhard adjust bash train-sleepedf-singlechanexp.sh --data_dir=<path to data> --leonhard

!!! EVALUATE CROSS VALIDATION !!!

To evaluate a trained model on a test set refer to the command line documentation of bin/validate.py:

python validate.py --h usage: validate.py [-h] [--model MODEL] [--data_dir DATA_DIR] [--subject_csv SUBJECT_CSV] [--output_dir OUTPUT_DIR] [--channel_drop CHANNEL_DROP]

This will create a .npz file for each subject in the subject_csv and save in the output_dir. It is recommended to save the results in

reports/results/<dataset>/<experiment>/<model arch>/<channel>

A jupyter notebook can then be used analyze overall accuracy, the confusion matrix, etc. An example can be found here.

All the steps to load a trained model, view its training parameters and how to score on a single subject is illustrated in this jupyter notebook.