I’m working to do color clustering (aka color palette, common colors, dominant colors) of images. One very useful technique is k-means clustering which is implemented in Scholar.

After k-means clustering I’d like to be able to get the silhouette_score which is implemented at Scholar.Metrics.Clustering.silhouette_score/3. That would seem to be a very natural step to validate or refuge the selection of the cluster size that was used (num_clusters: 16 below).

However the output generated by fit/2 doesn’t seem to play well with with silhouette_score/2 despite them being in a similar domain.

iex> kmeans = Image.Nx.kmeans(i, num_clusters: 16)
%Scholar.Cluster.KMeans{
  num_iterations: #Nx.Tensor<
    s64
    EXLA.Backend<host:0, 0.2302565275.2944008210.1119>
    24
  >,
  clusters: #Nx.Tensor<
    f32[16][3]
    EXLA.Backend<host:0, 0.2302565275.2944008210.1120>
    [
      [29.555273056030273, 22.40241813659668, 18.988496780395508],
      [221.6188507080078, 186.17681884765625, 115.16548919677734],
      [158.05177307128906, 123.88009643554688, 53.33770751953125],
      [127.1727294921875, 95.57713317871094, 41.13895034790039],
      [62.95863342285156, 47.129539489746094, 33.71294021606445],
      [97.46968841552734, 79.16098022460938, 61.60565185546875],
      [204.48558044433594, 216.12306213378906, 231.07981872558594],
      [161.0527801513672, 182.42922973632812, 208.91531372070312],
      [131.0211944580078, 107.85687255859375, 78.05144500732422],
      [190.57359313964844, 153.97003173828125, 62.70090103149414],
      [193.35650634765625, 161.45425415039062, 100.00911712646484],
      [162.95822143554688, 135.37939453125, 89.99324035644531],
      [222.13687133789062, 185.55056762695312, 75.35253143310547],
      [95.11575317382812, 67.52027893066406, 28.772855758666992],
      [250.12782287597656, 223.23638916015625, 149.09202575683594],
      [244.928466796875, 210.7081298828125, 104.9043960571289]
    ]
  >,
  inertia: #Nx.Tensor<
    f32
    EXLA.Backend<host:0, 0.2302565275.2944008210.1121>
    33850068.0
  >,
  labels: #Nx.Tensor<
    s64[91767]
    EXLA.Backend<host:0, 0.2302565275.2944008210.1122>
    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...]
  >
}

The function Image.Nx.kmeans/2 is just:

    def kmeans(%Vimage{} = image, options \\ []) do
      {_count, colors} = unique_colors(image)
      Scholar.Cluster.KMeans.fit(colors, options)
    end

It feels like it would be reasonable to be able to call Scholar.Metrics.Clustering.silhouette_score(colors, kmeans.clusters) but that’s very definitely not the signature.

Any advice or guidance on how to apply the results from fit/2 to silhouette_score/2 would be greatly appreciated.