Development of signal decomposition algorithms based on PyTorch

[fanghaoming3]

Thank you for providing such valuable code. For someone like me who can only use Python but not Matlab, your work has helped me save time to a great extent. However, your project is developed using numpy and scipy in Python, and most of the signal decomposition algorithms run slowly. Therefore, I cannot embed it into my deep neural network for real-time signal decomposition operations.
Therefore, I hope that a signal decomposition library based on PyTorch can be developed for me to use in the future. The speed of signal decomposition can be increased through CUDA acceleration. @wwhenxuan

[wwhenxuan]

Thank you for providing such valuable code. For someone like me who can only use Python but not Matlab, your work has helped me save time to a great extent. However, your project is developed using numpy and scipy in Python, and most of the signal decomposition algorithms run slowly. Therefore, I cannot embed it into my deep neural network for real-time signal decomposition operations. Therefore, I hope that a signal decomposition library based on PyTorch can be developed for me to use in the future. The speed of signal decomposition can be increased through CUDA acceleration. @wwhenxuan

Thank you very much for your valuable @fanghaoming3 🥰🥰🥰.

The data science environment currently developed by PySDKit is Numpy and Scipy, which does have some problems in running speed, because this development environment can only perform operations through the CPU. In fact, I thought about this problem carefully before I was ready to develop PySDKit, but I finally decided to use only Numpy and Scipy for development, and other modules such as Matplotlib and tqdm as auxiliary.

On the one hand, because the deep learning environment of PyTorch is relatively large, it is not very friendly to download for those who want to quickly experience our algorithm. Numpy and Scipy are relatively easy to download, and there is no need to consider the CUDA version of the current host and whether CUDA acceleration can be used.
On the other hand, I did consider whether there would be users who would embed the code into a machine learning model or a deep neural network as part of it. However, it is worth noting that even the simplest signal decomposition algorithm, such as empirical mode decomposition, is more complex than Fourier and wavelet transforms, because it involves multiple iterations to solve extreme values ​​and upper and lower envelope spectra. Moreover, variational mode decomposition also involves variational optimization, so its complexity is relatively high. I personally test that you should first decompose the data when using it, process the input in advance, and then pass it into the deep neural network for feature extraction and learning.

However, in order to increase the speed of the algorithm through CUDA, we will further use PyTorch to rebuild our algorithm later. Thank you again for your valuable suggestions 🥳.

[wwhenxuan]

PtSDKit based on Pytorch might be our future work😋