-
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:
- 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.
- Before submitting any any task to the scheduling system, load the appropriate python module:
module load python_gpu/3.6.4
Training the single channel expert model local on the EEG channel:
bsub -n 5 -W 4:00 -R "rusage[mem=6000,scratch=25000,ngpus_excl_p=1]" python train.py with caro_EEG_2D singlechanexp save_model=True ds.fold=0 ds.data_dir=[DATA DIRECTORY] ds.train_csv=../cfg/caro/debug_train.csv ds.val_csv=../cfg/caro/debug_val.csv
The attention model requires trained experts models. Training the attention model local on all channels:
bsub -n 5 -W 5:00 -R "rusage[mem=6000,scratch=25000,ngpus_excl_p=1]" python train.py with caro_all_2D_no_sweat AttentionNet_RS160_caro save_model=True ds.fold=0 ds.data_dir=[DATA DIRECTORY] ds.train_csv=../cfg/caro/debug_train.csv ds.val_csv=../cfg/caro/debug_val.csv
Training an unsupervized DSSM requires trained expert models. By changing the "unsupervized" flag, you can switch between the supervized and unsupervized DSSM:
bsub -n 5 -W 10:00 -R "rusage[mem=18000,scratch=60000,ngpus_excl_p=1]" python train.py with caro_all_2D_no_sweat DSSM_caro save_model=True unsupervized=[True/False] ds.fold=0 ds.data_dir=[DATA DIRECTORY] ds.train_csv=../cfg/caro/debug_train.csv ds.val_csv=../cfg/caro/debug_val.csv
This model is responsible for translating unsupervized DSSM states into sleep stage predictions. It requires a trained unsupervized DSSM.
bsub -n 5 -W 5:00 -R "rusage[mem=6000,scratch=25000,ngpus_excl_p=1]" python train.py with caro_all_2D_no_sweat Sleep_Classifier_caro save_model=True ds.fold=0 ds.data_dir=[DATA DIRECTORY] ds.train_csv=../cfg/caro/debug_train.csv ds.val_csv=../cfg/caro/debug_val.csv
The available run-options can be found in config.py. Some of the more important ones: save_model: whether to save the model or not unsupervized: should be False (default) except for the unsupervized DSSM ds.data_dir: where the data lies ds.train_csv/ds.val_csv: where train/validation splits are saved log_dir: where to save the logs. this includes checkpoints and spectograms ds.nbrs: how many neighbouring samples to consider ds.osnbrs: whether the neighbouring samples should be onesided (default=False), meaning coming before the sample to predict instead of symetrically around it. ds.fold: what fold to train/validate on. this is used for cross-validation ms.label_nbrs: whether to calculate the loss from neighbouring samples as well ms.dropout: dropout in the convolutional layers ms.hidden size: dimensionality of nn layers ms.hidden size: dimensionality of cnn layers
To train and evaluate a model via cross-validation, you can run the script train_caro.sh. Make sure to set the correct paths in the script before, and to specify the dataset (e.g. caro or caro_new) at the top of the script. With the variabes providied at the beginning of the script you can also skip training or cross-validation.
Cross-validation 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.