How to apply k-means clustering to 4D tensor?

General Discussion
#1

I received tensor of 4 dimensions (805, 8, 8, 128) from latent space of autoencoder (as input it was tensor of (1108, 32, 32, 3) - 1108 arrays of 32x32 pictures of 3 colors)

now I want to apply k-means to my feature tensor of 4 dimensions (805, 8, 8, 128), but every kmeans function I try (from sklearn, scipy etc) it through an error that number of dimensions different than 1 or 2 are not supported

how could i find clusters above all my arrays of feature tensor?

#3

As e.g. scikit-learn max 2d arrays inputs. But you can reshape the input.

See python 3.x - Applying Kmeans on 3D data with Scikit-learn - Stack Overflow

#4

E.g. for a basic example with an Autoencoder and MNIST see:

#6

thank you for the links! i think i just don’t understand how reshaping array would not lead to loosing information (it was also in a comments to stackoverflow question but still not getting the clear understanding)

like when i flatten the array from 3d to 1d it turns out to be completely different data, isn’t it?

#7

If don’t think that the Autoencoder output will care about the structure of the data you can try to have an UMAP pass like in:

1 Like