This is a python code to train CNN model, and run evaluation or prediction on ECG (Electrocardiography) data challenge to detect invertions in ECG data.
NOTE: Sadly, I'm not the owner of the data, try to ask if dataset is available at git repository Détection d'inversions ECG
CNN model defined with Keras framework and used Tensorflow backend.
Model architecture motivated by the current state-of-the-art in image processing - Convolutional Neural Networks.
Contrary to images, ECG signal is 1D signal. So convolutions are applied as 1D filters.
This model has a simple architecture:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) (None, 750, 12) 0
_________________________________________________________________
conv1d_1 (Conv1D) (None, 748, 64) 2368
_________________________________________________________________
max_pooling1d_1 (MaxPooling (None, 374, 64) 0
_________________________________________________________________
conv1d_2 (Conv1D) (None, 372, 128) 24704
_________________________________________________________________
max_pooling1d_2 (MaxPooling (None, 186, 128) 0
_________________________________________________________________
conv1d_3 (Conv1D) (None, 184, 256) 98560
_________________________________________________________________
max_pooling1d_3 (MaxPooling (None, 92, 256) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 23552) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 23552) 0
_________________________________________________________________
output (Dense) (None, 1) 23553
=================================================================
Total params: 149,185.0
Trainable params: 149,185.0
Model trained with cross entropy loss function, max of 20 epochs and early stopping option when there are no improvement in loss function for more that 3 epochs.
On MacBook Pro 13" (2015) CPU training takes around 3 min.
Notes on training: training dataset contains around 1K samples and to improve generalisation capabilities of the network there is a simple data augmentation method in model.py script. For every original sample there are 5 more samples generated with random horizontal shift (we can say 'in timeline'). So the total volume of training data is higher that the original input data.
This script tested in the following environment:
- Python 2.7
- TensorFlow 1.0 (CPU-only)
- Keras 2.0.2
- Numpy 1.12.0
- h5py 2.6.0
- MacOS Sierra
Install python requirements:
pip install --requirement requirements.txt
Usage:
python model.py [-h] --run-mode {TRAIN,EVAL,PRED} --data-csv DATA_CSV
--labels-csv LABELS_CSV [--model-name MODEL_NAME]
where arguments are:
-h, --help show this help message and exit
--run-mode {TRAIN,EVAL,PRED}
Perform one of the following operations on model use
these commands: TRAIN : train model, EVAL : evaluate
model PRED : make prediction with model
--data-csv DATA_CSV Raw data CSV file
--labels-csv LABELS_CSV
Labels CSV file. Labels are ignored in PRED mode
--model-name MODEL_NAME
Optional model name to be added as suffix to output
files
For example, if you have your train data in the file input_training.csv and target labels in output_training.csv, then this command will train model and save it with suffix my_cnn:
python model.py --run-mode TRAIN \
--data-csv input_training.csv \
--labels-csv output_training.csv \
--model-name my_cnn
Alexander Usoltsev 2017