Skip to content

Visualization of 1D CNNs filters space for Time Series Analysis

License

Notifications You must be signed in to change notification settings

MSD-IRIMAS/filter1D_visualization

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

38 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

2D Visualization of Convolution Filters on Time Series Data

This repository automatically generates an HTML (index.html) page showcasing a 2D space of 1D convolution-based filters, a common technique employed in Deep Learning for Time Series Classification. Inspired by the techniques outlined in Deep Learning for Time Series Classification, this tool leverages the power of the Bokeh library and Aeon toolkit to visualize these filters in an interactive and intuitive manner. Explore the spatial arrangement of filters and gain insights into their distribution, aiding in model understanding and interpretation.

Requirements

tensorflow
numpy
hydra-core
aeon
scikit-learn
bokeh
bs4
matplotlib

Usage

This code utilizes the hydra configuration setup, simply edit the parameters of the configuration file ('config/config_hydra.yaml`) and run the following:

python main.py

The different parameters include:

  • output_directory: The output directory where the results are stored
  • model_paths: A list of model paths to use, make sure to use the absolute path. PS: make sure you use .keras not .hdf5 files if you are tf >= 2.15
  • layer_indices: a list of layer indices in corresponding models
  • list_labels: a list of labels for each of the models
  • list_colors: a list of colors in HEX code, if None used colors are randomly generated
  • distance: a distance to compare filters among available distances in Aeon (default: 'dtw')
  • title: title of the html page (default: 'Filter 2D Visualization')

Example

We present below the generated figure of our code using the example models stored in example/.

This generates this html file, visualizing filters from the last conv layer of two different Lite classifiers trained on Car dataset.

example image

Acknowledgments

This work was supported by the ANR DELEGATION project (grant ANR-21-CE23-0014) of the French Agence Nationale de la Recherche.