add k-d tree #447
add k-d tree #447
Conversation
Depends on: 4336440 from PR mratsim#444.
`nonzero` is now a well behaved proc, which returns a Tensor instead of a nested seq. :)
We don't want to export `<` for tensors that does a element wise comparison until something is smaller or larger in either argument. As I couldn't make the compiler happy with a `bind `<`` in the code, let's work around it by creating a `distinct` type that we only use in the heap queue! 🥳
|
Only have to turn a few scipy comparisons into a proper test now. Seems to work fine now. An example of a 2D tree based on 1000 random points as a non balanced tree. The color corresponds to the node ID in which the data is. The shape corresponds to the result of a ball query around the red point with a radius of 0.1. All found points in the search correspond are drawn as points. The rest as crosses.
And the same input data to create a balanced tree (cutting at the median instead of the middle):
Both are using the default leaf size of 16. edit: The plots and stuffAs is evident from these plots up there, I wrote some code to plot such trees. Of course the plotting part is only really possible in the 2D case. But the hyperrectangle + extracting the bounds of each node could come in handy for the k-d case too. I wasn't sure whether to put it into this module directly. While looking around I saw that lonely |
|
Ok, aside from the question about whether I should add the plotting / hyperrectangle stuff, this is done from my side. Since this is a bit bigger, I'd appreciate if someone could take a look. :) |
|
Ugh, I forgot about the annoying |
|
Let's hope it works now. |
This is almost a straight up port from scipy's KD tree (the Python implementation, not
cKDTree, which is a C implementation).In principle the implementation works and gives the same results as the scipy tree.
Things left to do:
seq[tuple]to something more reasonable, maybe tuple ofTensor[T], Tensor[int].<comparison of nodesmaybe unify the metricsused here with a general concept of metrics in arraymancer (given that e.g. dbscan allows for different metrics, maybe kmeans does too, haven't checked). NOTE: this will be done after the DBSCAN PR is merged (I'll try to get that done over the weekend). Then we can have a generalmetricsinspatialthat can be used in clustering and here and by itself.