ssswokao / extramae

Official Implementation of Time Series Generation with Masked Autoencoder (ExtraMAE)

Instructions for Reimplementation

We provide the official implementation of Time Series Generation with Masked Autoencoder under the folder MAI.

You may implement experiments on Stock, Sine, and Energy according to the README.md under MAI.

We also provide three jupyter notes for you to preliminarily check the results.

• ExtraMAE.ipynb shows experiments on visualization, discrimination scores, prediction scores, and ablation study.
• Imputaition.ipynb shows results for imputation
• MaskRatio.ipynb shows results in the study of mask ratios.

The jupyter notes rely on large experiment results. Therefore, the readers cannot run them directly. You may find the links for all results of experiments and runnable jupyter notes in SupplementaryMaterial.pdf.

Masked Autoencoder with Extrapolator (ExtraMAE)

Intro

The folder MAI provides the official implementation of Time Series Generation with Masked Autoencoder.

We summarize the prediction, discrimination, and visualization results in summary.ipynb. You may refer to it directly or reimplement our experiments by the steps below.

Repository Structure

Model

Pytorch implementation of ExtraMAE can be found in ./MAI/models.py. All supporting files are stored in ./MAI/modules/*.

modules include:

• utils.py stores all utility functions
• generation.py generates masks and synthetic data
• visualization.py visualize the original data and synthetic data by PCA plots, t-SNE plots, and dot plots.

Datasets

Datasets are stored in ./MAI/data/*

Metrics

Metrics for assessing the quality of synthetic data are in ./MAI/metrics/*:

• Predictive Scores: We train a predictor on the synthetic data and do testing on the original data. Mean absolute value (MAE) on the original data is reported as the predictive score. (The lower, the better. )
• Discriminative Scores: We train a discriminator to tell synthetic data from the original data. Accuracy on the testing set is reported as the discriminative score. (The lower, the better. )

Configuration

config_stock.json configs for the stock dataset.

config_energy.json configs for energy dataset.

config_sine.json configs for sine dataset.

requirements.txt stores all packages needed.

Results

A experiment on Stock named AE0_EM0_RE50000 , for example, has a private folder ./MAI/storage/AE0_EM0_RE50000 for its own and there are three subfolders under it.

• model stores model related files
• pics stores visualization for the evaluation of ExtraMAE by generation mode random_once
• synthesis stores visualization for three other generation modes:
  • cross_average
  • cross_concate
  • random_average

summary.ipynb summarizes the results.

Implementation

1. Create a conda environment named MAI.

   conda create -n MAI python=3.6
   conda activate MAI

2. Enter the folder of the repository MAI in the terminal.

3. Get your environment ready.

   pip install -r requirements.txt

4. Run main.py with the config you need.

   python main.py --config_dir=stock_config.json

   python main.py --config_dir=energy_config.json

   python main.py --config_dir=sine_config.json

5. In case you wanna change the configuration for different experiments and instances, just modify the config.json and save it.

6. To see the summarized results of predictive scores and discriminative scores, send the experiment_name to method find_results in MAI/summary.ipynb.