DeepAnt Implementation

This is an Implementation of the DeepAnt Architecture for Anomaly Detection in Univariate TimeSeries using Pytorch Lightning.

The Folder Structure goes as follows:

• checkpoints: Checkpoint of the best model. It can be loaded to check inference.
• data: Data used in the training Process.
• DeepAnt.ipynb: The Notebook showing the training process plus evaluation. All the code is modularized into utils.py and deepant.py.

In this case the Algorithm was applied to one of the Numenta Benchmarks mentioned in the paper. In order to reproduce the algorithm the same Hyperparameters were used when available:

• SEQ_LEN: 10

• 2 1D-Convolutional Blocks:

  • Number of Filters: 32
  • Conv Kernel 3x3
  • Relu Activation Function
  • MaxPool Kernel: 2x2

• 1 Fully Connected Block

  • Fully Connected Hidden Layer: 40
  • ReLU Activation Function
  • Fully Connected Dropout Rate: 0.25
  • p_w: 1

• Training Epochs: 30

• Adam Optimizer:

  • Learning Rate: 1e-5
  • MAE Loss

• Error Calculation: L2 Norm