bhendi-boi/ann-snn

ANN

• Download ann.ipynb and ann_train.csv to replicate my work
• You need these libraries:
  1. matplotlib
  2. pytorch
  3. numpy
  4. pandas
• In addition to them you need inbuilt modules like copy and random

Results

• got around 98% accuracy for Arrhythmia_classifier

  

• ANN and SNN networks can be calculated now. Download spikes.ipynb file and run it with the biases and weights you got after running the Arrhythmia_classifier model from ann.ipynb .

SNN

• Download the snn.ipynb and 4vectors.csv file to replicate my work.

• You need these libraries:

  1. snntorch
  2. pytorch
  3. sklearn
  4. imbalencelearn
  5. matplotlib
  6. pandas
  7. numpy

• Additionally, you can run the same model on google colab.

• If you have an Nvidia GPU, install CUDA to take advantage of your GPU which highly reduces the time taken to train the model.

Network Structure

• num of inputs - 4
• num of hidden layers - 8
• num of output - 5

Input feature vectors include rr_interval, qrs_interval, st_interval, qt_interval.

A quick overview of preprocessing.

If you want to update the flowchart, open flowchart.excalidraw on excalidraw

How to play with the parameters

• parameters you can change include beta for the LIF neurons, the sampling strategy, learning_rate specified in the ADAM optimizer.
• beta is the decay rate of membrane potential, clipped between 0 and 1 during the forward-pass.
• To know more about LIF neurons, visit this website.
• sampling strategy is used to over sample classes whose count is very less. Whatever the number you assign to a particular key, that will be the number of elements in the final dataset before splitting for training and testing.
• The over sampling method used is SMOTE.

Results