Is there some clean way to do K-Means clustering on Tensor data without converting it to numpy array.
I have a list of tensors and their corresponding labes and this is what I am doing.
def evaluateKMeansRaw(data, true_labels, n_clusters):
kmeans = KMeans(n_clusters=n_clusters,n_init=20)
kmeans.fit(data)
acc = cluster_acc(true_labels, kmeans.labels_)
nmi = metrics.normalized_mutual_info_score(true_labels, kmeans.labels_)
return acc, nmi
But this doesn’t work on the output of a Conv Layer
You cannot use scikit-learn on tensors as scikit-learn (and all of its methods) only work on numpy arrays.
However if you are not afraid to use custom implementations you could use something like this
Do we not have these algorithms as part of the framework as of now ?
check out this github repo. can be installed in a breeze using pip:
pip install kmeans-pytorch
find documentation here
I implemented NN, KNN and KMeans on a project I am working on only using PyTorch. You can find the implementation here with an example: Nearest Neighbor, K Nearest Neighbor and K Means (NN, KNN, KMeans) only using PyTorch · GitHub
>>> import torch as th
>>> from clustering import KNN
>>> data = th.Tensor([[1, 1], [0.88, 0.90], [-1, -1], [-1, -0.88]])
>>> labels = th.LongTensor([3, 3, 5, 5])
>>> test = th.Tensor([[-0.5, -0.5], [0.88, 0.88]])
>>> knn = KNN(data, labels)
>>> knn(test)
tensor([5, 3])