Re: [Scikit-learn-general] Making approximate nearest neighbor search more efficient
Hi Joel,
I was on vacation during past 3 days. I''ll look into this asap and let you
all know.
I also did some profiling, but only with the usage of `pairwise_distance`
method. Brute force technique directly uses that for the entire query
array, but LSH uses that in a loop and I noticed there is a huge lag. I'll
first confirm your claims. I can start working on this but I think I'll
need your or some other contributers' reviewing as well . I'll do this if
it's possible.
On Sun, Aug 2, 2015 at 3:50 AM, Joel Nothman joel.noth...@gmail.com wrote:
@Maheshakya, will you be able to do work in the near future on speeding up
the ascending phase instead? Or should another contributor take that baton?
Not only does it seem to be a major contributor to runtime, but it is
independent of metric and hashing mechanism (within binary hashes), and
hence the most fundamental component of LSHForest.T
On 30 July 2015 at 22:28, Joel Nothman joel.noth...@gmail.com wrote:
(sorry, I should have said the first b layers, not 2**b layers, producing
a memoization of 2**b offsets
On 30 July 2015 at 22:22, Joel Nothman joel.noth...@gmail.com wrote:
One approach to fixing the ascending phase would ensure that
_find_matching_indices is only searching over parts of the tree that have
not yet been explored, while currently it searches over the entire index at
each depth.
My preferred, but more experimental, solution is to memoize where the
first 2**b layers of the tree begin and end in the index, for small b. So
if our index stored:
[[0, 0, 0, 1, 1, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 1, 1, 0, 0, 0],
[0, 1, 1, 0, 0, 1, 1, 0],
[1, 0, 0, 0, 0, 1, 0, 1],
[1, 0, 0, 1, 0, 1, 0, 1],
[1, 1, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 0, 0, 0]]
and b=2, we'd memoize offsets for prefixes of size 2:
[0, # 00
3, # 01
6, # 10
8, # 11
]
Given a query like [0, 1, 1, 0, 0, 0, 0, 0], it's easy to shift down to
leave the first b bits [0, 1] remaining, and look them up in the array just
defined to identify a much narrower search space [3, 6) matching that
prefix in the overall index.
Indeed, given the min_hash_match constraint, not having this sort of
thing for b = min_hash_match seems wasteful.
This provides us O(1) access to the top layers of the tree when
ascending, and makes the searchsorted calls run in log(n / (2 ** b)) time
rather than log(n). It is also much more like traditional LSH. However, it
complexifies the code, as we now have to consider two strategies for
descent/ascent.
On 30 July 2015 at 21:46, Joel Nothman joel.noth...@gmail.com wrote:
What makes you think this is the main bottleneck? While it is not an
insignificant consumer of time, I really doubt this is what's making
scikit-learn's LSH implementation severely underperform with respect to
other implementations.
We need to profile. In order to do that, we need some sensible
parameters that users might actually want, e.g. number of features for
{dense, sparse} cases, index size, target 10NN precision and recall
(selecting corresponding n_estimators and n_candidates). Ideally we'd
consider real-world datasets. And of course, these should be sensible for
whichever metric we're operating over, and whether we're doing KNN or
Radius searches.
I don't know if it's realistic, but I've profiled the following
bench_plot_approximate_neighbors settings:
Building NearestNeighbors for 10 samples in 100 dimensions
LSHF parameters: n_estimators = 15, n_candidates = 100
Building LSHForest for 10 samples in 100 dimensions
Done in 1.492s
Average time for lshf neighbor queries: 0.005s
Average time for exact neighbor queries: 0.002s
Average Accuracy : 0.88
Speed up: 0.5x
Of 4.77s total time spent in LSHForest.kneighbors for a 1000-query
matrix, we have:
- 0.03 spent in _query (hashing and descending)
- 0.91 spent in _compute_distances (exact distance calculation)
- 3.80 remaining in _get_candidates (ascending phase), almost all
of which is spent in _find_matching_indices
Cutting exact distance calculation to 0s will still not get this faster
than the exact approach. Of course, your mileage may vary, but this
suggests to me you're barking up the wrong tree (no pun intended).
On 30 July 2015 at 19:43, Maheshakya Wijewardena
pmaheshak...@gmail.com wrote:
Hi,
I've started to look into the matter of improving performance of
LSHForest. As we have discussed sometime before(in fact, quite a long
time), main concern is to Cythonize distance calculations. Currently, this
done by iteratively moving over all the query vectors when `kneighbors`
method is called for a set of query vectors. It has been identified that
iterating over each query with Python loops is a huge overhead. I have
implemented a few Cython hacks to demonstrate the distance calculation in
LSHForest and I was able to get an approximate speedup 10x compared
Re: [Scikit-learn-general] About contributing code
Hi Andreas, Thank you for your reply! The NFL belongs in the k-NN family of algorithms for classification / regression. In general, for a query point, NFL works like k-NN but instead of using the k feature points to determine the results, it generalized any two feature points of the same class by a feature line and then computes the distance of the query point from its projection on feature lines. It is supposed to improve the results of NN in some cases and especially on patter recognition tasks. You can have a look here http://www.scholarpedia.org/article/Nearest_feature_line http://www.dsp.toronto.edu/juwei/Publication/IEEE_NN.pdf I recently learnt about it while studying for a feature extraction course. I haven't implemented / tested it yet and I am new to sci-kit learn but I thought it would be fun (from my side) to implement it and have the opportunity to contribute too. I know that this is not how things work for you. To add a feature to the library it must be something that is going to be useful for many people so that there is a reason for maintaining it. Having a look at the issue page on github I think I can help with some minor contributions like adding more related projects in documentation (I will see If I can do that asap). Besides that my idea is to implement something from scratch so that I will be able to get familiar with the project step by step and if I end up with something that you will want to add to the library I will be very happy to contribute. Thank you for your time. I will continue watching the issue page and maybe help with something. Best Regards, Prokopis On Tue, Jul 28, 2015 at 8:43 PM, Andreas Mueller t3k...@gmail.com wrote: Hi Gryllos. Before contributing a new feature (which is usually a major undertaking) it us usually a good idea to get started working on known issues, have a look at the issue tracker. I'm not familiar with the feature line approach. Can you elaborate and provide a reference? Please see the FAQ for our policy on algorithms we like to include: http://scikit-learn.org/dev/faq.html#can-i-add-this-new-algorithm-that-i-or-someone-else-just-published http://scikit-learn.org/dev/faq.html#can-i-add-this-classical-algorithm-from-the-80s Cheers, Andy On 07/23/2015 06:54 PM, Prokopis Gryllos wrote: Hi everyone, I would like to contribute code to the project and I was thinking of implementing a nearest feature line approach to the nearest neighbor class. As it is suggested in the instruction set about contributing I thought it would be best to ask you first before I start working on it. Thank you in advance, I am waiting for your reply! Best Regards, Gryllos Prokopis -- ___ Scikit-learn-general mailing listScikit-learn-general@lists.sourceforge.nethttps://lists.sourceforge.net/lists/listinfo/scikit-learn-general -- ___ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general -- ___ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
Re: [Scikit-learn-general] Making approximate nearest neighbor search more efficient
@Maheshakya, will you be able to do work in the near future on speeding up
the ascending phase instead? Or should another contributor take that baton?
Not only does it seem to be a major contributor to runtime, but it is
independent of metric and hashing mechanism (within binary hashes), and
hence the most fundamental component of LSHForest.
On 30 July 2015 at 22:28, Joel Nothman joel.noth...@gmail.com wrote:
(sorry, I should have said the first b layers, not 2**b layers, producing
a memoization of 2**b offsets)
On 30 July 2015 at 22:22, Joel Nothman joel.noth...@gmail.com wrote:
One approach to fixing the ascending phase would ensure that
_find_matching_indices is only searching over parts of the tree that have
not yet been explored, while currently it searches over the entire index at
each depth.
My preferred, but more experimental, solution is to memoize where the
first 2**b layers of the tree begin and end in the index, for small b. So
if our index stored:
[[0, 0, 0, 1, 1, 0, 0, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 0, 1, 0, 1, 0, 1, 0],
[0, 1, 0, 0, 0, 0, 0, 0],
[0, 1, 0, 1, 1, 0, 0, 0],
[0, 1, 1, 0, 0, 1, 1, 0],
[1, 0, 0, 0, 0, 1, 0, 1],
[1, 0, 0, 1, 0, 1, 0, 1],
[1, 1, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 0, 0, 0]]
and b=2, we'd memoize offsets for prefixes of size 2:
[0, # 00
3, # 01
6, # 10
8, # 11
]
Given a query like [0, 1, 1, 0, 0, 0, 0, 0], it's easy to shift down to
leave the first b bits [0, 1] remaining, and look them up in the array just
defined to identify a much narrower search space [3, 6) matching that
prefix in the overall index.
Indeed, given the min_hash_match constraint, not having this sort of
thing for b = min_hash_match seems wasteful.
This provides us O(1) access to the top layers of the tree when
ascending, and makes the searchsorted calls run in log(n / (2 ** b)) time
rather than log(n). It is also much more like traditional LSH. However, it
complexifies the code, as we now have to consider two strategies for
descent/ascent.
On 30 July 2015 at 21:46, Joel Nothman joel.noth...@gmail.com wrote:
What makes you think this is the main bottleneck? While it is not an
insignificant consumer of time, I really doubt this is what's making
scikit-learn's LSH implementation severely underperform with respect to
other implementations.
We need to profile. In order to do that, we need some sensible
parameters that users might actually want, e.g. number of features for
{dense, sparse} cases, index size, target 10NN precision and recall
(selecting corresponding n_estimators and n_candidates). Ideally we'd
consider real-world datasets. And of course, these should be sensible for
whichever metric we're operating over, and whether we're doing KNN or
Radius searches.
I don't know if it's realistic, but I've profiled the following
bench_plot_approximate_neighbors settings:
Building NearestNeighbors for 10 samples in 100 dimensions
LSHF parameters: n_estimators = 15, n_candidates = 100
Building LSHForest for 10 samples in 100 dimensions
Done in 1.492s
Average time for lshf neighbor queries: 0.005s
Average time for exact neighbor queries: 0.002s
Average Accuracy : 0.88
Speed up: 0.5x
Of 4.77s total time spent in LSHForest.kneighbors for a 1000-query
matrix, we have:
- 0.03 spent in _query (hashing and descending)
- 0.91 spent in _compute_distances (exact distance calculation)
- 3.80 remaining in _get_candidates (ascending phase), almost all of
which is spent in _find_matching_indices
Cutting exact distance calculation to 0s will still not get this faster
than the exact approach. Of course, your mileage may vary, but this
suggests to me you're barking up the wrong tree (no pun intended).
On 30 July 2015 at 19:43, Maheshakya Wijewardena pmaheshak...@gmail.com
wrote:
Hi,
I've started to look into the matter of improving performance of
LSHForest. As we have discussed sometime before(in fact, quite a long
time), main concern is to Cythonize distance calculations. Currently, this
done by iteratively moving over all the query vectors when `kneighbors`
method is called for a set of query vectors. It has been identified that
iterating over each query with Python loops is a huge overhead. I have
implemented a few Cython hacks to demonstrate the distance calculation in
LSHForest and I was able to get an approximate speedup 10x compared to
current distance calculation with a Python loop. However, I came across
some blockers while trying to do this and need some clarifications.
What I need to know is, do we use a mechanism to release GIL when we
want to parallelize. One of my observations is `pairwise_distance` uses all
the cores even when I don't specify `n_jobs` parameter which is 1 in
default. Is this an expected behavior?
If I want to release GIL, can I use OpenMP module in Cython? Or is that
a task of Joblib?
Any input on this is highly appreciated.
Best regards,
--
