FastAutoAugment Time Series

Author: Matthieu OLEKHNOVITCH

This is an unofficial implementation of the paper Fast AutoAugment for time series data. The original paper is for image data. The code is inspired from the official implementation of the paper Fast AutoAugment.

⚀ Dataset and Model

The dataset used is the UCR Time Series Classification Archive. The model used is a ResNet model.

⚁ Architecture

The architecture of the augmented model is as follows:

⚂ Transformations

The transformations used are as follows:

• Identity
• Additive Noise
• Random Crop
• Drift
• Reverse

⚃ Requirements

• Python 3.12

⛓ Usage

Prepare the environment

I personally recommend using uv as package manager. You can install it using the following command:

pip install uv

You can then create the environment and install the dependencies using the following commands:

uv venv

. venv/bin/activate

uv pip install -r requirements.txt

Configure your cuda device and WnB link in the config file 'FastAA/config.py'

Set the device to 'cuda' or 'cpu' if you want to override automatic device detection. Set is_wandb to True if you want to log the results to WnB. You may have to log in to WnB using the following command:

wandb login

Basic run of augmentation on single dataset

You can run FastAA augmentation comparison using the following command:

python FastAA/main.py --dataset=ECG5000 --compare --runs=5

Other parameters are detailed in the help:

python FastAA/main.py --help

Run augmentation on all datasets

You can run FastAA augmentation on all datasets using the following command:

python FastAA/run_full_datasets_exploration.py

Note: This will take a long time to run. You can run it in the background using the following command:

nohup python FastAA/run_full_datasets_exploration.py &

The results will be saved in the data/logs folder as soon as they are computed.

Plot the metrics comparison between FastAA and baseline

Plot the metrics comparison with the command:

python results/plot_result_table.py 

And access the last computed results in results/metrics_comparison.png :

We can see that significant improvements are made on some datasets, while others are not improved, or even worsened. Therefore, we propose to study the impact of the transformations on the datasets and try to understand why some datasets are improved and others are not.

📈 Results