Last active
October 26, 2022 14:30
-
-
Save betatim/ddf6a2fbff158703f051b6c3114698f0 to your computer and use it in GitHub Desktop.
This file contains hidden or 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
| import time | |
| import numpy as np | |
| import cupy as cp | |
| import sklearn | |
| from sklearn.cluster import KMeans | |
| from sklearn.datasets import make_blobs | |
| from sklearn.model_selection import train_test_split | |
| import cuml | |
| # XXX Attempt to speed up first run | |
| from sklearn_gpu.kmeans import KMeansEngine | |
| n_features = 16# * 2 | |
| n_centers = 12 | |
| X, y_true = make_blobs( | |
| n_samples=3_000_000, centers=n_centers, n_features=n_features, cluster_std=0.60, random_state=10 | |
| ) | |
| X, X_, y_true, y_true_ = train_test_split(X, y_true, random_state=42, test_size=3000) | |
| #X, y_true = make_blobs( | |
| # n_samples=300_000, centers=8, n_features=n_features, cluster_std=1.20, random_state=0 | |
| #) | |
| kmeans_args = dict(n_clusters=n_centers, random_state=42, n_init=1) | |
| #@profile | |
| def default(): | |
| # Default implementation | |
| km = KMeans(**kmeans_args) | |
| tic = time.perf_counter_ns() | |
| for _ in range(1): | |
| km.fit(X) | |
| toc = time.perf_counter_ns() | |
| y_pred_ = km.predict(X_) | |
| print("CPU run took:", (toc-tic)/1_000_000) | |
| #@profile | |
| def cuml_(): | |
| km3 = cuml.KMeans(**kmeans_args) | |
| tic = time.perf_counter_ns() | |
| for _ in range(10): | |
| km3.fit(X) | |
| toc = time.perf_counter_ns() | |
| y_pred3_ = km3.predict(X_) | |
| print("CUML run took:", (toc-tic)/1_000_000) | |
| #@profile | |
| def sklearn_gpu(): | |
| # Using the accelerated version | |
| with sklearn.config_context(engine_provider="sklearn_gpu"): | |
| km2 = KMeans(**kmeans_args) | |
| tic = time.perf_counter_ns() | |
| for _ in range(10): | |
| km2.fit(X) | |
| toc = time.perf_counter_ns() | |
| y_pred2_ = km2.predict(X_) | |
| # XXX the first run is always way way slow. Maybe because things need to be imported? | |
| # XXX mitigated (a bit) by running the cuml KMeans first | |
| print("GPU run took:", (toc-tic)/1_000_000) | |
| if __name__ == "__main__": | |
| default() | |
| #cuml_() | |
| #cuml_() | |
| #sklearn_gpu() | |
| #sklearn_gpu() |
This file contains hidden or 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
| Timer unit: 1e-06 s | |
| Total time: 0.59432 s | |
| File: /home/nfs/thead/git/scikit-learn/sklearn/cluster/_kmeans.py | |
| Function: fit at line 1525 | |
| Line # Hits Time Per Hit % Time Line Contents | |
| ============================================================== | |
| 1525 @profile | |
| 1526 def fit(self, X, y=None, sample_weight=None): | |
| 1527 """Compute k-means clustering. | |
| 1528 | |
| 1529 Parameters | |
| 1530 ---------- | |
| 1531 X : {array-like, sparse matrix} of shape (n_samples, n_features) | |
| 1532 Training instances to cluster. It must be noted that the data | |
| 1533 will be converted to C ordering, which will cause a memory | |
| 1534 copy if the given data is not C-contiguous. | |
| 1535 If a sparse matrix is passed, a copy will be made if it's not in | |
| 1536 CSR format. | |
| 1537 | |
| 1538 y : Ignored | |
| 1539 Not used, present here for API consistency by convention. | |
| 1540 | |
| 1541 sample_weight : array-like of shape (n_samples,), default=None | |
| 1542 The weights for each observation in X. If None, all observations | |
| 1543 are assigned equal weight. | |
| 1544 | |
| 1545 .. versionadded:: 0.20 | |
| 1546 | |
| 1547 Returns | |
| 1548 ------- | |
| 1549 self : object | |
| 1550 Fitted estimator. | |
| 1551 """ | |
| 1552 1 931.0 931.0 0.2 self._validate_params() | |
| 1553 1 31.0 31.0 0.0 engine = self._get_engine() | |
| 1554 | |
| 1555 2 72906.0 36453.0 12.3 X, y, sample_weight = engine.prepare_fit( | |
| 1556 1 0.0 0.0 0.0 X, | |
| 1557 1 0.0 0.0 0.0 y=y, | |
| 1558 1 1.0 1.0 0.0 sample_weight=sample_weight, | |
| 1559 ) | |
| 1560 1 27.0 27.0 0.0 self._check_params_vs_input(X) | |
| 1561 | |
| 1562 1 1.0 1.0 0.0 best_inertia, best_labels = None, None | |
| 1563 | |
| 1564 2 3.0 1.5 0.0 for i in range(self._n_init): | |
| 1565 # Initialize centers | |
| 1566 1 142986.0 142986.0 24.1 centers_init = engine.init_centroids(X) | |
| 1567 1 6.0 6.0 0.0 if self.verbose: | |
| 1568 print("Initialization complete") | |
| 1569 | |
| 1570 # run a k-means once | |
| 1571 2 372880.0 186440.0 62.7 labels, inertia, centers, n_iter_ = engine.kmeans_single( | |
| 1572 1 1.0 1.0 0.0 X, | |
| 1573 1 2.0 2.0 0.0 sample_weight, | |
| 1574 1 2.0 2.0 0.0 centers_init, | |
| 1575 ) | |
| 1576 | |
| 1577 # determine if these results are the best so far | |
| 1578 # we chose a new run if it has a better inertia and the clustering is | |
| 1579 # different from the best so far (it's possible that the inertia is | |
| 1580 # slightly better even if the clustering is the same with potentially | |
| 1581 # permuted labels, due to rounding errors) | |
| 1582 1 3.0 3.0 0.0 if best_inertia is None or ( | |
| 1583 inertia < best_inertia | |
| 1584 and not engine.is_same_clustering(labels, best_labels, self.n_clusters) | |
| 1585 ): | |
| 1586 1 1.0 1.0 0.0 best_labels = labels | |
| 1587 1 1.0 1.0 0.0 best_centers = centers | |
| 1588 1 1.0 1.0 0.0 best_inertia = inertia | |
| 1589 1 1.0 1.0 0.0 best_n_iter = n_iter_ | |
| 1590 | |
| 1591 1 32.0 32.0 0.0 engine.unshift_centers(X, best_centers) | |
| 1592 | |
| 1593 1 4473.0 4473.0 0.8 distinct_clusters = np.unique(best_labels).shape[0] | |
| 1594 1 2.0 2.0 0.0 if distinct_clusters < self.n_clusters: | |
| 1595 warnings.warn( | |
| 1596 "Number of distinct clusters ({}) found smaller than " | |
| 1597 "n_clusters ({}). Possibly due to duplicate points " | |
| 1598 "in X.".format(distinct_clusters, self.n_clusters), | |
| 1599 ConvergenceWarning, | |
| 1600 stacklevel=2, | |
| 1601 ) | |
| 1602 | |
| 1603 1 12.0 12.0 0.0 self.cluster_centers_ = best_centers | |
| 1604 1 4.0 4.0 0.0 self._n_features_out = self.cluster_centers_.shape[0] | |
| 1605 1 4.0 4.0 0.0 self.labels_ = best_labels | |
| 1606 1 2.0 2.0 0.0 self.inertia_ = best_inertia | |
| 1607 1 6.0 6.0 0.0 self.n_iter_ = best_n_iter | |
| 1608 1 1.0 1.0 0.0 return self | |
| Total time: 0.074355 s | |
| File: /home/nfs/thead/git/scikit-learn/sklearn/cluster/_kmeans.py | |
| Function: predict at line 1610 | |
| Line # Hits Time Per Hit % Time Line Contents | |
| ============================================================== | |
| 1610 @profile | |
| 1611 def predict(self, X, sample_weight=None): | |
| 1612 """Predict the closest cluster each sample in X belongs to. | |
| 1613 | |
| 1614 In the vector quantization literature, `cluster_centers_` is called | |
| 1615 the code book and each value returned by `predict` is the index of | |
| 1616 the closest code in the code book. | |
| 1617 | |
| 1618 Parameters | |
| 1619 ---------- | |
| 1620 X : {array-like, sparse matrix} of shape (n_samples, n_features) | |
| 1621 New data to predict. | |
| 1622 | |
| 1623 sample_weight : array-like of shape (n_samples,), default=None | |
| 1624 The weights for each observation in X. If None, all observations | |
| 1625 are assigned equal weight. | |
| 1626 | |
| 1627 Returns | |
| 1628 ------- | |
| 1629 labels : ndarray of shape (n_samples,) | |
| 1630 Index of the cluster each sample belongs to. | |
| 1631 """ | |
| 1632 1 78.0 78.0 0.1 check_is_fitted(self) | |
| 1633 1 42.0 42.0 0.1 engine = self._get_engine() | |
| 1634 1 627.0 627.0 0.8 X, sample_weight = engine.prepare_prediction(X, sample_weight) | |
| 1635 1 73608.0 73608.0 99.0 return engine.get_labels(X, sample_weight) |
This file contains hidden or 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
| Timer unit: 1e-06 s | |
| Total time: 2.99799 s | |
| File: /home/nfs/thead/git/scikit-learn/sklearn/cluster/_kmeans.py | |
| Function: fit at line 1525 | |
| Line # Hits Time Per Hit % Time Line Contents | |
| ============================================================== | |
| 1525 @profile | |
| 1526 def fit(self, X, y=None, sample_weight=None): | |
| 1527 """Compute k-means clustering. | |
| 1528 | |
| 1529 Parameters | |
| 1530 ---------- | |
| 1531 X : {array-like, sparse matrix} of shape (n_samples, n_features) | |
| 1532 Training instances to cluster. It must be noted that the data | |
| 1533 will be converted to C ordering, which will cause a memory | |
| 1534 copy if the given data is not C-contiguous. | |
| 1535 If a sparse matrix is passed, a copy will be made if it's not in | |
| 1536 CSR format. | |
| 1537 | |
| 1538 y : Ignored | |
| 1539 Not used, present here for API consistency by convention. | |
| 1540 | |
| 1541 sample_weight : array-like of shape (n_samples,), default=None | |
| 1542 The weights for each observation in X. If None, all observations | |
| 1543 are assigned equal weight. | |
| 1544 | |
| 1545 .. versionadded:: 0.20 | |
| 1546 | |
| 1547 Returns | |
| 1548 ------- | |
| 1549 self : object | |
| 1550 Fitted estimator. | |
| 1551 """ | |
| 1552 1 1042.0 1042.0 0.0 self._validate_params() | |
| 1553 1 37.0 37.0 0.0 engine = self._get_engine() | |
| 1554 | |
| 1555 2 772429.0 386214.5 25.8 X, y, sample_weight = engine.prepare_fit( | |
| 1556 1 1.0 1.0 0.0 X, | |
| 1557 1 0.0 0.0 0.0 y=y, | |
| 1558 1 0.0 0.0 0.0 sample_weight=sample_weight, | |
| 1559 ) | |
| 1560 1 23.0 23.0 0.0 self._check_params_vs_input(X) | |
| 1561 | |
| 1562 1 1.0 1.0 0.0 best_inertia, best_labels = None, None | |
| 1563 | |
| 1564 2 4.0 2.0 0.0 for i in range(self._n_init): | |
| 1565 # Initialize centers | |
| 1566 1 1735836.0 1735836.0 57.9 centers_init = engine.init_centroids(X) | |
| 1567 1 4.0 4.0 0.0 if self.verbose: | |
| 1568 print("Initialization complete") | |
| 1569 | |
| 1570 # run a k-means once | |
| 1571 2 434253.0 217126.5 14.5 labels, inertia, centers, n_iter_ = engine.kmeans_single( | |
| 1572 1 0.0 0.0 0.0 X, | |
| 1573 1 1.0 1.0 0.0 sample_weight, | |
| 1574 1 1.0 1.0 0.0 centers_init, | |
| 1575 ) | |
| 1576 | |
| 1577 # determine if these results are the best so far | |
| 1578 # we chose a new run if it has a better inertia and the clustering is | |
| 1579 # different from the best so far (it's possible that the inertia is | |
| 1580 # slightly better even if the clustering is the same with potentially | |
| 1581 # permuted labels, due to rounding errors) | |
| 1582 1 4.0 4.0 0.0 if best_inertia is None or ( | |
| 1583 inertia < best_inertia | |
| 1584 and not engine.is_same_clustering(labels, best_labels, self.n_clusters) | |
| 1585 ): | |
| 1586 1 1.0 1.0 0.0 best_labels = labels | |
| 1587 1 1.0 1.0 0.0 best_centers = centers | |
| 1588 1 1.0 1.0 0.0 best_inertia = inertia | |
| 1589 1 1.0 1.0 0.0 best_n_iter = n_iter_ | |
| 1590 | |
| 1591 1 47.0 47.0 0.0 engine.unshift_centers(X, best_centers) | |
| 1592 | |
| 1593 1 54274.0 54274.0 1.8 distinct_clusters = np.unique(best_labels).shape[0] | |
| 1594 1 2.0 2.0 0.0 if distinct_clusters < self.n_clusters: | |
| 1595 warnings.warn( | |
| 1596 "Number of distinct clusters ({}) found smaller than " | |
| 1597 "n_clusters ({}). Possibly due to duplicate points " | |
| 1598 "in X.".format(distinct_clusters, self.n_clusters), | |
| 1599 ConvergenceWarning, | |
| 1600 stacklevel=2, | |
| 1601 ) | |
| 1602 | |
| 1603 1 14.0 14.0 0.0 self.cluster_centers_ = best_centers | |
| 1604 1 8.0 8.0 0.0 self._n_features_out = self.cluster_centers_.shape[0] | |
| 1605 1 3.0 3.0 0.0 self.labels_ = best_labels | |
| 1606 1 2.0 2.0 0.0 self.inertia_ = best_inertia | |
| 1607 1 3.0 3.0 0.0 self.n_iter_ = best_n_iter | |
| 1608 1 0.0 0.0 0.0 return self | |
| Total time: 0.065059 s | |
| File: /home/nfs/thead/git/scikit-learn/sklearn/cluster/_kmeans.py | |
| Function: predict at line 1610 | |
| Line # Hits Time Per Hit % Time Line Contents | |
| ============================================================== | |
| 1610 @profile | |
| 1611 def predict(self, X, sample_weight=None): | |
| 1612 """Predict the closest cluster each sample in X belongs to. | |
| 1613 | |
| 1614 In the vector quantization literature, `cluster_centers_` is called | |
| 1615 the code book and each value returned by `predict` is the index of | |
| 1616 the closest code in the code book. | |
| 1617 | |
| 1618 Parameters | |
| 1619 ---------- | |
| 1620 X : {array-like, sparse matrix} of shape (n_samples, n_features) | |
| 1621 New data to predict. | |
| 1622 | |
| 1623 sample_weight : array-like of shape (n_samples,), default=None | |
| 1624 The weights for each observation in X. If None, all observations | |
| 1625 are assigned equal weight. | |
| 1626 | |
| 1627 Returns | |
| 1628 ------- | |
| 1629 labels : ndarray of shape (n_samples,) | |
| 1630 Index of the cluster each sample belongs to. | |
| 1631 """ | |
| 1632 1 73.0 73.0 0.1 check_is_fitted(self) | |
| 1633 1 33.0 33.0 0.1 engine = self._get_engine() | |
| 1634 1 490.0 490.0 0.8 X, sample_weight = engine.prepare_prediction(X, sample_weight) | |
| 1635 1 64463.0 64463.0 99.1 return engine.get_labels(X, sample_weight) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment