-
Notifications
You must be signed in to change notification settings - Fork 758
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request #445 from brj0/add-nndescent-algorithm
add nndescent algorithm
- Loading branch information
Showing
5 changed files
with
267 additions
and
2 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,5 @@ | ||
| FROM ann-benchmarks | ||
|
|
||
| RUN pip3 install --verbose numpy nndescent>=1.0.4 | ||
|
|
||
| RUN python3 -c 'import nndescent' |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,138 @@ | ||
| bit: | ||
| hamming: | ||
| - base_args: ['@metric'] | ||
| constructor: NNDescent | ||
| disabled: false | ||
| docker_tag: ann-benchmarks-nndescent | ||
| module: ann_benchmarks.algorithms.nndescent | ||
| name: nndescent | ||
| run_groups: | ||
| NN-120: | ||
| arg_groups: [{pruning_prob: [0.0, 1.0], leaf_size: 80, n_neighbors: [120], | ||
| pruning_degree_multiplier: [2.0, 2.5]}] | ||
| args: {} | ||
| query_args: [[0.08, 0.16, 0.2, 0.24, 0.28, 0.32, 0.36]] | ||
| NN-20: | ||
| arg_groups: [{pruning_prob: [0.75, 1.0], leaf_size: 32, n_neighbors: [20], | ||
| pruning_degree_multiplier: [1.0, 1.5]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.01, 0.02, 0.04, 0.08, 0.12, 0.16]] | ||
| NN-40: | ||
| arg_groups: [{pruning_prob: [0.5, 1.0], leaf_size: 48, n_neighbors: [40], | ||
| pruning_degree_multiplier: [1.5, 2.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.04, 0.08, 0.12, 0.16, 0.2, 0.24]] | ||
| NN-80: | ||
| arg_groups: [{pruning_prob: [0.25, 1.0], leaf_size: 64, n_neighbors: [80], | ||
| pruning_degree_multiplier: [1.75, 2.25]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.08, 0.12, 0.16, 0.2, 0.24, 0.28, 0.32]] | ||
| jaccard: | ||
| - base_args: ['@metric'] | ||
| constructor: NNDescent | ||
| disabled: false | ||
| docker_tag: ann-benchmarks-nndescent | ||
| module: ann_benchmarks.algorithms.nndescent | ||
| name: nndescent | ||
| run_groups: | ||
| NN-120: | ||
| arg_groups: [{pruning_prob: [1.0, 0.125], leaf_size: 80, n_neighbors: 120, | ||
| pruning_degree_multiplier: 1.0}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.12, 0.14, 0.16, 0.18, 0.2, 0.22]] | ||
| NN-20: | ||
| arg_groups: [{pruning_prob: [0.75, 1.0], leaf_size: 30, n_neighbors: 20, | ||
| pruning_degree_multiplier: 1.0}] | ||
| args: {} | ||
| query_args: [[0.0, 0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.12, 0.16, 0.2]] | ||
| NN-40: | ||
| arg_groups: [{pruning_prob: [0.5, 1.0], leaf_size: 30, n_neighbors: 40, | ||
| pruning_degree_multiplier: 1.0}] | ||
| args: {} | ||
| query_args: [[0.0, 0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.12, 0.16, 0.2]] | ||
| NN-80: | ||
| arg_groups: [{pruning_prob: [1.0, 0.25], leaf_size: 60, n_neighbors: 80, | ||
| pruning_degree_multiplier: 1.0}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.12, 0.14, 0.16, 0.18, 0.2, 0.22]] | ||
| float: | ||
| angular: | ||
| - base_args: ['@metric'] | ||
| constructor: NNDescent | ||
| disabled: false | ||
| docker_tag: ann-benchmarks-nndescent | ||
| module: ann_benchmarks.algorithms.nndescent | ||
| name: nndescent | ||
| run_groups: | ||
| NN-120-accurate: | ||
| arg_groups: [{pruning_prob: 0.125, leaf_size: 35, n_neighbors: 120, pruning_degree_multiplier: 2.5}] | ||
| args: {} | ||
| query_args: [[0.16, 0.2, 0.24, 0.28, 0.32, 0.36]] | ||
| NN-120-fast: | ||
| arg_groups: [{pruning_prob: 1.0, leaf_size: 20, n_neighbors: 120, pruning_degree_multiplier: 2.5}] | ||
| args: {} | ||
| query_args: [[0.0, 0.04, 0.08, 0.16, 0.2, 0.24, 0.28, 0.32]] | ||
| NN-20: | ||
| arg_groups: [{pruning_prob: [1.0], leaf_size: 20, n_neighbors: [20], pruning_degree_multiplier: [ | ||
| 0.5, 1.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12]] | ||
| NN-40: | ||
| arg_groups: [{pruning_prob: [0.5, 1.0], leaf_size: 25, n_neighbors: [40], | ||
| pruning_degree_multiplier: [1.5]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.14, 0.16]] | ||
| NN-80-accurate: | ||
| arg_groups: [{pruning_prob: 0.25, leaf_size: 30, n_neighbors: 80, pruning_degree_multiplier: 2.0}] | ||
| args: {} | ||
| query_args: [[0.08, 0.12, 0.16, 0.2, 0.24, 0.28, 0.32, 0.36]] | ||
| NN-80-fast: | ||
| arg_groups: [{pruning_prob: 1.0, leaf_size: 20, n_neighbors: 80, pruning_degree_multiplier: 2.0}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.08, 0.12, 0.16, 0.2, 0.24]] | ||
| any: | ||
| - base_args: ['@metric'] | ||
| constructor: NNDescent | ||
| disabled: false | ||
| docker_tag: ann-benchmarks-nndescent | ||
| module: ann_benchmarks.algorithms.nndescent | ||
| name: nndescent | ||
| run_groups: | ||
| NN-10-20: | ||
| arg_groups: [{pruning_prob: [1.0], leaf_size: 32, n_neighbors: [10, 20], | ||
| pruning_degree_multiplier: [1.5, 2.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.14, 0.16]] | ||
| NN-40-80: | ||
| arg_groups: [{pruning_prob: [0.0, 1.0], leaf_size: 64, n_neighbors: [40, | ||
| 80], pruning_degree_multiplier: [2.0, 2.5]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.04, 0.08, 0.12, 0.16, 0.2, 0.24, 0.28, 0.32]] | ||
| euclidean: | ||
| - base_args: ['@metric'] | ||
| constructor: NNDescent | ||
| disabled: false | ||
| docker_tag: ann-benchmarks-nndescent | ||
| module: ann_benchmarks.algorithms.nndescent | ||
| name: nndescent | ||
| run_groups: | ||
| NN-10: | ||
| arg_groups: [{pruning_prob: 1.0, leaf_size: 24, n_neighbors: 10, pruning_degree_multiplier: [ | ||
| 0.5, 1.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.1, 0.12]] | ||
| NN-20: | ||
| arg_groups: [{pruning_prob: 1.0, leaf_size: 24, n_neighbors: 20, pruning_degree_multiplier: [ | ||
| 0.75, 1.5]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.06, 0.08, 0.1, 0.12, 0.14, 0.16, 0.2]] | ||
| NN-40: | ||
| arg_groups: [{pruning_prob: [0.0, 1.0], leaf_size: 36, n_neighbors: 40, | ||
| pruning_degree_multiplier: [1.0, 2.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.02, 0.04, 0.08, 0.12, 0.16, 0.2, 0.24, 0.28, 0.32]] | ||
| NN-60: | ||
| arg_groups: [{pruning_prob: 0.0, leaf_size: 48, n_neighbors: 60, pruning_degree_multiplier: [ | ||
| 2.0, 3.0]}] | ||
| args: {} | ||
| query_args: [[0.0, 0.04, 0.08, 0.12, 0.16, 0.2, 0.24, 0.28, 0.32, 0.36]] |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,120 @@ | ||
| import numpy as np | ||
| import nndescent | ||
| import scipy.sparse | ||
|
|
||
| from ..base.module import BaseANN | ||
|
|
||
|
|
||
| class NNDescent(BaseANN): | ||
| def __init__(self, metric, index_param_dict): | ||
| if "n_neighbors" in index_param_dict: | ||
| self.n_neighbors = int(index_param_dict["n_neighbors"]) | ||
| else: | ||
| self.n_neighbors = 30 | ||
|
|
||
| if "pruning_degree_multiplier" in index_param_dict: | ||
| self.pruning_degree_multiplier = float( | ||
| index_param_dict["pruning_degree_multiplier"] | ||
| ) | ||
| else: | ||
| self.pruning_degree_multiplier = 1.5 | ||
|
|
||
| if "pruning_prob" in index_param_dict: | ||
| self.pruning_prob = float(index_param_dict["pruning_prob"]) | ||
| else: | ||
| self.pruning_prob = 1.0 | ||
|
|
||
| if "leaf_size" in index_param_dict: | ||
| self.leaf_size = int(index_param_dict["leaf_size"]) | ||
|
|
||
| self.is_sparse = metric in ["jaccard"] | ||
|
|
||
| self.nnd_metric = { | ||
| "angular": "dot", | ||
| "euclidean": "euclidean", | ||
| "hamming": "hamming", | ||
| "jaccard": "jaccard", | ||
| }[metric] | ||
|
|
||
| def fit(self, X): | ||
| if self.is_sparse: | ||
| # Convert to sparse matrix format | ||
| if type(X) == list: | ||
| sizes = [len(x) for x in X] | ||
| n_cols = max([max(x) for x in X]) + 1 | ||
| matrix = scipy.sparse.csr_matrix( | ||
| (len(X), n_cols), dtype=np.float32 | ||
| ) | ||
| matrix.indices = np.hstack(X).astype(np.int32) | ||
| matrix.indptr = np.concatenate([[0], np.cumsum(sizes)]).astype( | ||
| np.int32 | ||
| ) | ||
| matrix.data = np.ones( | ||
| matrix.indices.shape[0], dtype=np.float32 | ||
| ) | ||
| matrix.sort_indices() | ||
| X = matrix | ||
| else: | ||
| X = scipy.sparse.csr_matrix(X) | ||
|
|
||
| self.query_matrix = scipy.sparse.csr_matrix( | ||
| (1, X.shape[1]), dtype=np.float32 | ||
| ) | ||
| elif not isinstance(X, np.ndarray) or X.dtype != np.float32: | ||
| print("Convert data to float32") | ||
| X = np.asarray(X, dtype=np.float32) | ||
|
|
||
| # nndescent uses pointers to the data. Make shure X does not change | ||
| # outside of this scope. | ||
| self.X = X | ||
| self.index = nndescent.NNDescent( | ||
| self.X, | ||
| n_neighbors=self.n_neighbors, | ||
| metric=self.nnd_metric, | ||
| leaf_size=self.leaf_size, | ||
| pruning_degree_multiplier=self.pruning_degree_multiplier, | ||
| pruning_prob=self.pruning_prob, | ||
| verbose=True, | ||
| ) | ||
| # Make a dummy query to prepare the search graph. | ||
| if self.is_sparse: | ||
| empty_mtx = np.empty((0, X.shape[0]), dtype=np.float32) | ||
| empty_csr = scipy.sparse.csr_matrix(empty_mtx) | ||
| self.index.query(empty_csr, k=1, epsilon=0.1) | ||
| else: | ||
| empty_mtx = np.empty((0, X.shape[0]), dtype=np.float32) | ||
| self.index.query(empty_mtx, k=1, epsilon=0.1) | ||
|
|
||
| def set_query_arguments(self, epsilon=0.1): | ||
| self.epsilon = float(epsilon) | ||
|
|
||
| def query(self, v, n): | ||
| if self.is_sparse: | ||
| # Convert index array to sparse matrix format and query; the | ||
| # overhead of direct conversion is high for single queries | ||
| # (converting the entire test dataset and sending single rows is | ||
| # better), so we just populate the required structures. | ||
| if v.dtype == np.bool_: | ||
| self.query_matrix.indices = np.flatnonzero(v).astype(np.int32) | ||
| else: | ||
| self.query_matrix.indices = v.astype(np.int32) | ||
| size = self.query_matrix.indices.shape[0] | ||
| self.query_matrix.indptr = np.array([0, size], dtype=np.int32) | ||
| self.query_matrix.data = np.ones(size, dtype=np.float32) | ||
| ind, dist = self.index.query( | ||
| self.query_matrix, k=n, epsilon=self.epsilon | ||
| ) | ||
| else: | ||
| ind, dist = self.index.query( | ||
| v.reshape(1, -1).astype("float32"), k=n, epsilon=self.epsilon | ||
| ) | ||
| return ind[0] | ||
|
|
||
| def __str__(self): | ||
| return ( | ||
| f"NNDescent(n_neighbors={self.n_neighbors}, " | ||
| f"pruning_mult={self.pruning_degree_multiplier:.2f}, " | ||
| f"pruning_prob={self.pruning_prob:.3f}, " | ||
| f"epsilon={self.epsilon:.3f}, " | ||
| f"leaf_size={self.leaf_size:02d})" | ||
| ) |