This is a CPU and GPU implementation of the KDD 2017 paper Towards an Optimal Subspace for K-Means^[1]. The CPU implementation (mode=cpu) is written in Numpy. There are 2 GPU implementations:

1. Using only PyCUDA and scikit-cuda API (mode=gpu)
2. Using PyCUDA with custom kernels optimized for this algorithm (mode=gpu_custom)

• Numpy
• PyCUDA: pip install pycuda
• scikit-cuda: Install from source as described here. We tested using commit #249538c.
• Matplotlib (for plots)
• scikit-learn (for computing NMI score)

Note: We tested this implementation only on Python 2. There are some issues with the GPU version on Python 3.

Go to src/

python main.py -d=<dataset_name> -k=<number_of_clusters> -mode=<mode>

For help: python main.py -h

3 available modes: cpu, gpu, gpu_custom

python main.py -d=wine -k=3 -mode=cpu

[i] Itr 1: 24 points changed
[i] Itr 2: 7 points changed
[i] Itr 3: 7 points changed
[i] Itr 4: 2 points changed
[i] Itr 5: 1 points changed
[i] Itr 6: 0 points changed

[i] Results
[*] m: 2
[*] NMI: 0.87590

[1] Mautz et. al. Towards an Optimal Subspace for K-Means