Use autoencoder can achieve the goal of dimension reduction. After dimension reduction, it'll become easier to train the data. Here, we use kmeans as an example.
Pleace go to this google drive link to download the image data. The data has been transformed into npy file, which cam be read by numpy library. https://drive.google.com/open?id=1Yoja5z8whnc8o0eUpvmJ5zRKKjmXgGc7