b-sayah / omsignal-temie-series Goto Github PK

Here we propose a simple model that operates on ECG data using 1d convolutions on the time domain, and simultaneously analyzes the spectra of these timeseries using a some fully-connected layers. The model employs batch normalization throughout, with ReLU activations. This model is trained to jointly predict 3 real numbers (mean r-r interval, std r-r interval, mean t-r interval) and the participant ID (32 classes).

main.py can be executed to run all experiments. It will train a model using config.yml and output all results in logs/, models/, and img/.

evaluation/eval.py can be executed to run the best trained model against the test set.

All major training settings are handled in config.yml. Most options should be self-explainatory. The remaining ones are handled below:

config.yml:

• data: location of the input files.
• preprocessing:
  • noise_gain: scaling factor applied to additive noise for data augmentation.
• dataloader:
  • test_proportion: we combine and shuffle the training and validation sets to increase our training set size. This number [0 1] controls the proportion allocated to the validation set.
• models:
  • tspec:
    • ts_len: length of the input timeseries in X.
    • spec_len: length of the spectra of the input timeseries in X (typically ts_len/2).
    • hid_dim: size of all hidden layers in the network. Controls capacity.
    • num_layers: number of hidden layers of the spectra MLP component and the shared embedding components of the network. Controls capacity.
    • out_dims: a list of the predictor dimensions. If an element is 1, the task is treated as a regression. Otherwise it is treated as an n-class classification problem.
• training:
  • schedule_patience: number of epochs where the validation loss does not decrease to wait before decreasing the learning rate by an order of magnitude.
  • early_stopping_patience: number of epochs to wait where the validation loss does not decrease before deciding to stop training altogether.

Training Details:

This model is trained on all 4 tasks jointly. For evaluation, kendal's tau is used (a rank-order correlation measure) for all regression tasks. Therefore, MarginRankLoss is employed instead of MSELoss, as would be typically used for regression tasks. Classification is trained using CrossEntropyLoss. Optimization was done using stochastic gradient descent with momentum. A learning rate scheduler is employed such that the learning rate is reduced by an order of magnitude if the validation loss plateaus for more than 20 epochs.

Data:

The provided data is split into 3 binary files:

• MILA_TrainLabeledData.dat - labeled data for supervised training
• MILA_ValidationLabeledData.dat - labeled data for validation,
• MILA_UnlabeledData.dat - unlabeled data.

Each labeled dataset (MILA_TrainLabeledData.dat and MILA_ValidationLabeledData.dat) contains 5 windows of 30 second length ECG data (sampled at 125 Hz) for each of the 32 participants. For each participant, the samples in the MILA_TrainLabeledData.dat and MILA_ValidationLabeledData.dat datasets have were collected on independent days. The test data was collected on a 3rd day.

The labeled data looks like this:

• Shape = 160 x 3754 - where 160 = 5 x 32 corresponds to the number of windows.
• Column 0 to Column 3749 - Columns corresponding to the ECG data ( 30 seconds x 125 Hz = 3750 ). They contain float values.
• Column 3750 - Columns corresponding to the PR_Mean of the corresponding ECG sample. It contains float values.
• Column 3751 - Columns corresponding to the RT_Mean of the corresponding ECG sample. It contains float values.
• Column 3752 - Columns corresponding to the RR_StdDev of the corresponding ECG sample. It contains float values.
• Column 3753 - Columns corresponding to the ID of the participant. It contains int values.

The unlabeled data looks like this:

• Shape = 657233 x 3750 - where 657233 corresponds to the remaining number of unlabeled windows.
• Column 0 to Column 3749 - Columns corresponding to the ECG data ( 30 seconds x 125 Hz = 3750 ). They contain float values.

The ID column of the dataset can be mapped back and forth between the original range [0 43] and a machine-friendly range [0 32] using the ymap() method of the Data object: utils.Data.ymap(). This is used for reporting final values during evaluation.