An Exploration: A Clustering Algorithm Capable of Participating in Backpropagation.
This is a clustering algorithm capable of participating in backpropagation. In application, it can replace one-hot encoding and only requires knowledge of the similarity relationship between samples to train the model. There is also room for improvement in this structure.
After setting up the file paths and downloading the data sets, you can directly use the GPU or CPU to run the files named ~main.py in the gaussian_sample_clustering, image_recognition, and speaker_recognition folders.
"gaussian_sample_clustering" is used for basic validation of the clustering ability of PD-Selector. "image_recognition" involves loading PD-Selector onto a simple model to verify its participation in the model's training process. "speaker_recognition" is used to validate the application of PD-Selector in more complex scenarios.
"FashionMNIST", "ST-AEDS-20180100_1-OS"
numpy, torch, matplotlib, torchvision, sys, os