Unsupervised Interpretable Pattern Discovery in Time Series Using Autoencoders

Link to pdf version of the article

Launch the method

usage: python3 AETest.py

Mandatory arguments:

• dataset_path
  • Path of the json dataset file.
• doc_length
  • Length of a temporal document.
• doc_height
  • Height of a temporal document.
• nb_filters
  • Number of filters given.
• filters_length
  • Length of the given filters.
• weights_path
  • Path of the network weigths file.
• iterations
  • Number of training iterations.
• batches_size
  • Number of examples in each batches
• gradient_algorithm
  • Algorithm used for gradient descent (SGD, momentum, ADAM).
• learning_rate
  • Learning rate used in training.
• momentum
  • Momentum used in training.
• lambdaGL
  • Group lasso coefficient.
• lambdaL
  • lasso coefficient
• lambdaKL
  • Kullback on latent coefficient.
• expe_file
  • Used as prefix of the output files path (format: /).
• gpu
  • GPU-to-be-used index.

optional arguments:

• -h, --help
  • show help
• --train
  • Launch on training mode.
• --scratch
  • Enable learning from scratch.

The temporal documents are under the form of simple text files, the json dataset file contain a dictionary with the keys "train" and "test" corresponding to lists of path of document (see given example).

When there is several GPUs on the machine, launch "CUDA_VISIBLE_DEVICES= python3 AETest.py " to have a correct memory allocation.

Generate synthetic data

usage: python3 GenerateTemporalDocument.py

Mandatory arguments:

• nb_doc
  • Number of temporal document to be generated.
• nb_test_tdoc
  • Number of test examples in dataset
• doc_length
  • Length of a temporal document
• doc_height
  • Height of a temporal document
• font_name
  • Path of the font file.
• font_size
  • Size of the font used.
• motifs
  • Words used as motifs (seperated with '_').
• motifs_length
  • Maximum length of a motif.
• nb_motifs_doc
  • Number of motifs in each document.
• min_nb_occ_motifs
  • Minimum number of observations in a motif.
• max_nb_occ_motifs
  • Maximum number of observations in a motif.
• noise
  • Proportion of the mean number of obervations in the generated documents added as salt-and-pepper noise.
• repository
  • Repository where generate the dataset.

optional arguments:

• -h, --help
  • show help
• --gene_img
  • Also generate png versions of the generated temporal documents.

The example of dataset was created by the command

• python3 GenerateTemporalDocument.py --gene_img 5 2 500 25 JennaSue.ttf 20 Egg_Eggplant_Plant 45 10 3500 4500 0.33 Example_Dataset