TensorFlow implementation (old master branch) is available at tf_legacy branch.
Repository for our NeurIPS 2019 paper, Diffeomorphic Temporal Alignment Nets co-authored by: Ron Shapira Weber, Matan Eyal, Nicki Skafte Detlefsen, Oren Shriki and Oren Freifeld.
Ron Shapira Weber (email: ronsha@post.bgu.ac.il)
- Standard Python(>=3.6) packages: numpy, matplotlib, tqdm, seaborn
- PyTorch >= 1.4
- tslean == 0.1.19 (requires for DTW based methods - SoftDTW, DBA and NCC)
- libcpab == 2.0
- For Nvidia GPU iimplementation: CUDA==11.0 + appropriate cuDNN as well. You can follow the instructions here.
For the native PyTorch implementation (slower), we support all operating system. For the fast CUDA implementation of libcpab, we only support Linux.
We recommend installing a virtual environment via Anaconda. For instance:
conda create -n dtan python=3.7 numpy matplotlib seaborn tqdm
licpab [2] is a python package supporting the CPAB transformations [1] in Numpy, Tensorflow and Pytorch.
Install libcpab
Note 1: you might have to recompile the dynamic libraries under /libcpab/tensorflow/
git clone https://github.com/SkafteNicki/libcpab
Add libcpab to your python path:
export PYTHONPATH=$PYTHONPATH:$YOUR_FOLDER_PATH/libcpab
Make sure libcpab was installed properly (Run one of the demos).
Clone the repository:
git clone https://github.com/BGU-CS-VIL/dtan.git
# move to pytorch branch
git checkout pytorch
Add DTAN to your python path:
export PYTHONPATH=$PYTHONPATH:$YOUR_FOLDER_PATH/dtan
Try the example code under dtan/exmaples (see also our Usage section below):
python UCR_alignment.py
- To initialize the model:
from DTAN.DTAN_layer import DTAN as dtan_model
# Init model to GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = dtan_model(signal_len=int, channels=int, tess=[int,], n_recurrence=int,
zero_boundary=bool, device='gpu').to(device)
Under the 'examples' dir you can find example scripts for training and running DTAN time-series joint alignment.
- UCR time-series classification archive [3] alignment example.
To run simply enter:
python UCR_alignment.py
Note that here we only provide one dataset from the UCR archive - ECGFiveDays. For the entire archive, please visit: https://www.cs.ucr.edu/~eamonn/time_series_data/
The script supports the following flags:
optional arguments:
-h, --help show this help message and exit
--tess_size TESS_SIZE
CPA velocity field partition
--smoothness_prior smoothness prior flag
--no_smoothness_prior
no smoothness prior flag
--lambda_smooth LAMBDA_SMOOTH
lambda_smooth, larger values -> smoother warps
--lambda_var LAMBDA_VAR
lambda_var, larger values -> larger warps
--n_recurrences N_RECURRENCES
number of recurrences of R-DTAN
--zero_boundary ZERO_BOUNDARY
zero boundary constrain
--n_epochs N_EPOCHS number of epochs
--batch_size BATCH_SIZE
batch size
--lr LR learning rate
-
Usage Example - Running with and without smoothness prior:
See the jupyter notebook, illustrating the importance of the smoothness prior. -
UCR Nearest Centroid Classification (NCC): Coming soon to PyTorch version
Here we provide an end-to-end pipeline for the NCC experiment described in our paper. The script uses the same hyper-parameters (i.e., lambda_var, lambda_smooth, n_recurrences) used in the paper, depending on the UCR dataset. To run the pipeline on the 'ECGFiveDays' dataset go to the 'examples' dir and simply run:
python UCR_NCC.py
You can change the dataset inside the script.
[1] @article{freifeld2017transformations,
title={Transformations Based on Continuous Piecewise-Affine Velocity Fields},
author={Freifeld, Oren and Hauberg, Soren and Batmanghelich, Kayhan and Fisher, John W},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2017},
publisher={IEEE}
}
[2] @misc{detlefsen2018,
author = {Detlefsen, Nicki S.},
title = {libcpab},
year = {2018},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/SkafteNicki/libcpab}},
}
[3] @misc{UCRArchive,
title={The UCR Time Series Classification Archive},
author={ Chen, Yanping and Keogh, Eamonn and Hu, Bing and Begum, Nurjahan and Bagnall, Anthony and Mueen, Abdullah and Batista, Gustavo},
year={2015},
month={July},
note = {\url{www.cs.ucr.edu/~eamonn/time_series_data/}}
}
This software is released under the MIT License (included with the software). Note, however, that if you are using this code (and/or the results of running it) to support any form of publication (e.g., a book, a journal paper, a conference paper, a patent application, etc.) then we request you will cite our paper:
@inproceedings{weber2019diffeomorphic,
title={Diffeomorphic Temporal Alignment Nets},
author={Weber, Ron A Shapira and Eyal, Matan and Skafte, Nicki and Shriki, Oren and Freifeld, Oren},
booktitle={Advances in Neural Information Processing Systems},
pages={6570--6581},
year={2019}
}