angus924/hydra

HYDRA: Competing convolutional kernels for fast and accurate time series classification

ROCKET · MINIROCKET · HYDRA

HYDRA

HYDRA: Competing convolutional kernels for fast and accurate time series classification

arXiv:2203.13652 (preprint)

We demonstrate a simple connection between dictionary methods for time series classification, which involve extracting and counting symbolic patterns in time series, and methods based on transforming input time series using convolutional kernels, namely ROCKET and its variants. We show that by adjusting a single hyperparameter it is possible to move by degrees between models resembling dictionary methods and models resembling ROCKET. We present HYDRA, a simple, fast, and accurate dictionary method for time series classification using competing convolutional kernels, combining key aspects of both ROCKET and conventional dictionary methods. HYDRA is faster and more accurate than the most accurate existing dictionary methods, and can be combined with ROCKET and its variants to further improve the accuracy of these methods.

Please cite as:

@article{dempster_etal_2022,
  author  = {Dempster, Angus and Schmidt, Daniel F and Webb, Geoffrey I},
  title   = {{HYDRA}: Competing convolutional kernels for fast and accurate time series classification},
  year    = {2022},
  journal = {arXiv:2203.13652}
}

Results

UCR Archive (112 Datasets, 30 Resamples)

• Hydra
• Hydra+MiniRocket, Hydra+MultiRocket, Hydra+Rocket

Requirements

• Python
• PyTorch
• NumPy
• scikit-learn (or similar)

Code

hydra.py

hydra_multivariate.py†

softmax.py*

† experimental
* Hydra + SGD for larger datasets (i.e., more than approx. 10,000 training examples)

Examples

from hydra import Hydra, SparseScaler
from sklearn.linear_model import RidgeClassifierCV

[...] # load data (torch.FloatTensor, shape = (num_examples, 1, length))

transform = Hydra(X_training.shape[-1])

X_training_transform = transform(X_training)
X_test_transform = transform(X_test)

scaler = SparseScaler()

X_training_transform = scaler.fit_transform(X_training_transform)
X_test_transform = scaler.transform(X_test_transform)

classifier = RidgeClassifierCV(alphas = np.logspace(-3, 3, 10))
classifier.fit(X_training_transform, Y_training)

predictions = classifier.predict(X_test_transform)

Acknowledgements

We thank Professor Eamonn Keogh and all the people who have contributed to the UCR time series classification archive. Figures in our paper showing mean ranks were produced using code from Ismail Fawaz et al. (2019).

🐲

hydra's Issues

Randomness causes the model to be unstable and impractical!

Due to the presence of random numbers, the performance of the model is very unstable, and the quality of the results depends entirely on luck. Can self. W be pre trained to make the model stable and robust? thanks

Add an example

Could you please upload an example using a multi-channel database?I'm a beginner. I want to repeat the program.

what does it mean hydra_multivariate.py is experimental

what does it mean hydra_multivariate.py
is experimental

is it possible to implement hydra with numpy?

pytorch is too heavy for me.

Can I use it in a very large dataset, eg . million

how to use minirocket_multi and HydraMultivariate together?

so (sample, feature, length) gets transformed into (sample, features) by both mini rocket multi and hydra multi, so should I just transform the data separately and combine them (sample, rocket feature + hydra feature)? or is any alternative way present?

Why use simple linear classifiers?

Thank you for your great works (i.e., rocket, minirocket, multirocket and hydra), which are really insightful and valuable!

I am curious why you use simple linear classifiers in all your works. I guess there may be two reasons. First, the feature dimension is large, so it is better to select simple classifiers with fewer parameters. Second, simple linear classifiers really work fine and achieve the SOTA on the UCR archive. Are there any other reasons?

Since the random convolution kernel transform is highly integrated and independent of subsequent classifications, I wonder if it is feasible to make some efforts on the classifiers.

Thank you for your time!

hope multivariate version

GOOD JOB, Thank U.

Why use .clamp(0).sqrt() on transformed input?

Hi,

I've been reading your paper and looking through the code, but I have difficulty understanding why you use .clamp(0).sqrt() on the transformed input data. See, e.g., line 94 in softmax.py:

94 X_training_transform = transform.batch(X_training).clamp(0).sqrt()

Have I missed something in the paper or somewhere else?

Thank you in advance for any help you can provide!

Best regards, Kristian