https://github.com/doxakis/cosinesimilaritydistancesongpu

Compute cosine similarity distances for all combinations of the dataset on the gpu with CUDA
https://github.com/doxakis/cosinesimilaritydistancesongpu

cuda

Last synced: 2 months ago
JSON representation

Compute cosine similarity distances for all combinations of the dataset on the gpu with CUDA

• Host: GitHub
• URL: https://github.com/doxakis/cosinesimilaritydistancesongpu
• Owner: doxakis
• License: mit
• Created: 2018-06-14T04:00:58.000Z (about 8 years ago)
• Default Branch: master
• Last Pushed: 2018-06-14T04:22:01.000Z (about 8 years ago)
• Last Synced: 2025-07-10T04:34:36.538Z (11 months ago)
• Topics: cuda
• Language: C#
• Size: 10.7 KB
• Stars: 0
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Cosine similarity distances on GPU

Compute cosine similarity distances for all combinations of the dataset on the gpu with CUDA

This was coded in c# with the library Alea GPU. A similar logic could be reuse in any language (Python, c++, etc.)

I plan to use it as a preprocessing step before running HDBSCAN (text clustering in a unsupervised way).

Calculating distances could make the algorithm faster and can be a way to scale out. (No need to use PCA to reduce the complexity)

This is more like a proof of concept.

Please note that the first time the kernel function run, a JIT compilation occur. It takes about 1 sec.

I would recommend to run it when starting your application if possible to minimize the impact on perceived performance.

# Future works

- Batch processing (if the array is too large, it does not work. We got : System.Exception: '[CUDAError] CUDA_ERROR_OUT_OF_MEMORY')

- Find optimal parameter (determine if it's better to use CPU only)

# Copyright and license

Code released under the MIT license.