wangyiqiu/dbscan-python

Theoretically Efficient and Practical Parallel DBSCAN

dbscan-python's Issues

Can the memory footprint be reduced?

With a 14178107*8 vector, a 108GB memory machine is quickly used up. Is there a way to reduce the memory footprint?

The train output:

Input: 14178107 points, dimension 8
scheduler = Parlay-HomeGrown
num-threads = 16
num-cell = 12333095
compute-grid = 5.06638

Question: Setting the number of threads

I would like to perform some tests with varying number of threads. What is the most convenient way to run your method for a desired number of threads? Thank you.

DBSCAN do not free memory

Thank for your fast DBSCAN realization. I have a problem. Calling dbscan.DBSCAN(x) consums additional memory. If I call dbscan.DBSCAN(x) n time consums n*V memory, where V is memory for one dbscan.DBSCAN(x) calling.

ST-DBSCAN extension possible?

I was wondering if this implementation could be extended to deal with spatio-temporal data (allowing one to specify a threshold for both space and time eps)?

pypi module not working

hello, the compiled module works like a charm but the module from pypi fails to load the DBSCAN method with error:
cannot import name 'DBSCAN' from 'dbscan' (unknown location)

awesome work btw

Parallel Approximate DBSCAN Clustering

Does this repository support parallel approximate DBSCAN clustering? If so, please instruct me on how to use it. Thank you.

Python Binding dbscan.DBSCAN Error

Hello,

I am trying to implement the library on python. Did the installation using pip3 install dbscan and I get this error:
ModuleNotFoundError: No module named 'dbscan.DBSCAN'
I am using python 3.9.12.

Thanks for the help.

Abnormal calculation result

When I set eps=0.01, the result is abnormal, but when eps=0.02, the result is normal.
I have provided data 1.zip.
Uploading 1.zip…

dbscan result is wrong with the latest two commits

I used src/make.sh to compile the code and generated *.so file successfully, but got a wrong clustering result.

The commit 646cc6 and 08dd4a has the same wrong results.

But I found 83c5696 which produced example.png can generate right result, but it seems to have memory leak issues.

Edit: The latest commit which work normally is 8c6afc

Followed all instructions and post CMAKE, make -j does nothing, no error

Hi,
finally compiled binary dbscan and can run it.
The sample file given in the tutorial is 40 lines long. When I run with it and eps 0.1, minpts 10 dbscan runs and outputs all points as outliers. Is this OK?
thanks!

Large N Overflows

Hi!

I'm working with a very large dataset of over 100M points and this implementation is working beautifully except for the documented warning.

"Large n, the program behavior might be undefined due to overflow"

This causes clusters to appear that are very distant from each other. Is there any way to improve this behavior? I can clean this up by thresholding dust, but there are some manual operations I can't do that would make my segmentation much nicer due to this problem.

Thank you so much for this wonderful DBSCAN implementation. I wouldn't be complaining if it wasn't capable of handling such a huge number of points already!

Cannot import in Conda environment

In my conda environment:
>>> import dbscan Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/zpyang/anaconda3/envs/visnet/lib/python3.9/site-packages/dbscan/__init__.py", line 22, in <module> from dbscan.DBSCAN import DBSCAN ModuleNotFoundError: No module named 'dbscan.DBSCAN'
but it is successfully imported while outside the conda environment. Not sure why that's the case.

[Question] Maximum dimensionality

Why maximum dimensionality is restricted to 20? Are You planning to add this feature in the future?

I would like to use parallel version of DBSCAN algorithm to cluster high dimensional data (Transformer models embeddings which have 512 dimensions).