Note
Go to the end to download the full example code. or to run this example in your browser via Binder
TSNE中的近似最近邻#
本示例展示了如何在管道中链接KNeighborsTransformer和TSNE。
它还展示了如何包装 nmslib
和 pynndescent
包以替换KNeighborsTransformer并执行近似最近邻。这些包可以通过 pip install nmslib pynndescent
进行安装。
注意:在KNeighborsTransformer中,我们使用的定义包括每个训练点作为其自身的邻居在 n_neighbors
的计数中,并且为了兼容性原因,当 mode == 'distance'
时会计算一个额外的邻居。
请注意,我们在提议的 nmslib
包装中也做了同样的处理。
# 作者:scikit-learn 开发者
# SPDX-License-Identifier: BSD-3-Clause
首先我们尝试导入这些包,并在缺失时警告用户。
我们定义了一个包装类,用于实现 nmslib
的 scikit-learn API,以及一个加载函数。
import joblib
import numpy as np
from scipy.sparse import csr_matrix
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.datasets import fetch_openml
from sklearn.utils import shuffle
class NMSlibTransformer(TransformerMixin, BaseEstimator):
"""用于将nmslib作为sklearn的KNeighborsTransformer的包装器"""
def __init__(self, n_neighbors=5, metric="euclidean", method="sw-graph", n_jobs=-1):
self.n_neighbors = n_neighbors
self.method = method
self.metric = metric
self.n_jobs = n_jobs
def fit(self, X):
self.n_samples_fit_ = X.shape[0]
# 请参阅手册中的更多指标
# https://github.com/nmslib/nmslib/tree/master/manual
space = {
"euclidean": "l2",
"cosine": "cosinesimil",
"l1": "l1",
"l2": "l2",
}[self.metric]
self.nmslib_ = nmslib.init(method=self.method, space=space)
self.nmslib_.addDataPointBatch(X.copy())
self.nmslib_.createIndex()
return self
def transform(self, X):
n_samples_transform = X.shape[0]
# 出于兼容性原因,由于每个样本都被视为其自身的邻居,因此将多计算一个额外的邻居。
n_neighbors = self.n_neighbors + 1
if self.n_jobs < 0:
# 与 joblib 对负 n_jobs 值的处理相同:
# 特别地, `n_jobs == -1` 表示“线程数与 CPU 数相同”。
num_threads = joblib.cpu_count() + self.n_jobs + 1
else:
num_threads = self.n_jobs
results = self.nmslib_.knnQueryBatch(
X.copy(), k=n_neighbors, num_threads=num_threads
)
indices, distances = zip(*results)
indices, distances = np.vstack(indices), np.vstack(distances)
indptr = np.arange(0, n_samples_transform * n_neighbors + 1, n_neighbors)
kneighbors_graph = csr_matrix(
(distances.ravel(), indices.ravel(), indptr),
shape=(n_samples_transform, self.n_samples_fit_),
)
return kneighbors_graph
def load_mnist(n_samples):
"""加载MNIST,打乱数据,只返回n_samples。"""
mnist = fetch_openml("mnist_784", as_frame=False)
X, y = shuffle(mnist.data, mnist.target, random_state=2)
return X[:n_samples] / 255, y[:n_samples]
我们对不同的精确/近似最近邻变换器进行基准测试。
import time
from sklearn.manifold import TSNE
from sklearn.neighbors import KNeighborsTransformer
from sklearn.pipeline import make_pipeline
datasets = [
("MNIST_10000", load_mnist(n_samples=10_000)),
("MNIST_20000", load_mnist(n_samples=20_000)),
]
n_iter = 500
perplexity = 30
metric = "euclidean"
# TSNE 需要一定数量的邻居,这取决于困惑度参数。
# 由于我们将每个样本都包含为其自身的邻居,因此需要增加一个。
n_neighbors = int(3.0 * perplexity + 1) + 1
tsne_params = dict(
init="random", # pca not supported for sparse matrices
perplexity=perplexity,
method="barnes_hut",
random_state=42,
n_iter=n_iter,
learning_rate="auto",
)
transformers = [
(
"KNeighborsTransformer",
KNeighborsTransformer(n_neighbors=n_neighbors, mode="distance", metric=metric),
),
(
"NMSlibTransformer",
NMSlibTransformer(n_neighbors=n_neighbors, metric=metric),
),
(
"PyNNDescentTransformer",
PyNNDescentTransformer(
n_neighbors=n_neighbors, metric=metric, parallel_batch_queries=True
),
),
]
for dataset_name, (X, y) in datasets:
msg = f"Benchmarking on {dataset_name}:"
print(f"\n{msg}\n" + str("-" * len(msg)))
for transformer_name, transformer in transformers:
longest = np.max([len(name) for name, model in transformers])
start = time.time()
transformer.fit(X)
fit_duration = time.time() - start
print(f"{transformer_name:<{longest}} {fit_duration:.3f} sec (fit)")
start = time.time()
Xt = transformer.transform(X)
transform_duration = time.time() - start
print(f"{transformer_name:<{longest}} {transform_duration:.3f} sec (transform)")
if transformer_name == "PyNNDescentTransformer":
start = time.time()
Xt = transformer.transform(X)
transform_duration = time.time() - start
print(
f"{transformer_name:<{longest}} {transform_duration:.3f} sec"
" (transform)"
)
示例输出:
在MNIST_10000上的基准测试:#
KNeighborsTransformer 0.007 秒(拟合) KNeighborsTransformer 1.139 秒(转换) NMSlibTransformer 0.208 秒(拟合) NMSlibTransformer 0.315 秒(转换) PyNNDescentTransformer 4.823 秒(拟合) PyNNDescentTransformer 4.884 秒(转换) PyNNDescentTransformer 0.744 秒(转换)
在MNIST_20000上的基准测试:#
KNeighborsTransformer 0.011 秒(拟合) KNeighborsTransformer 5.769 秒(转换) NMSlibTransformer 0.733 秒(拟合) NMSlibTransformer 1.077 秒(转换) PyNNDescentTransformer 14.448 秒(拟合) PyNNDescentTransformer 7.103 秒(转换) PyNNDescentTransformer 1.759 秒(转换)
请注意, PyNNDescentTransformer
在第一次 fit
和第一次 transform
时会花费更多时间,这是由于 numba 即时编译器的开销。但在第一次调用之后,编译后的 Python 代码会被 numba 缓存在缓存中,后续调用不会再有这种初始开销。KNeighborsTransformer
和 NMSlibTransformer
在这里只运行一次,因为它们的 fit
和 transform
时间更稳定(它们没有 PyNNDescentTransformer 的冷启动问题)。
import matplotlib.pyplot as plt
from matplotlib.ticker import NullFormatter
transformers = [
("TSNE with internal NearestNeighbors", TSNE(metric=metric, **tsne_params)),
(
"TSNE with KNeighborsTransformer",
make_pipeline(
KNeighborsTransformer(
n_neighbors=n_neighbors, mode="distance", metric=metric
),
TSNE(metric="precomputed", **tsne_params),
),
),
(
"TSNE with NMSlibTransformer",
make_pipeline(
NMSlibTransformer(n_neighbors=n_neighbors, metric=metric),
TSNE(metric="precomputed", **tsne_params),
),
),
]
# 初始化图表
nrows = len(datasets)
ncols = np.sum([1 for name, model in transformers if "TSNE" in name])
fig, axes = plt.subplots(
nrows=nrows, ncols=ncols, squeeze=False, figsize=(5 * ncols, 4 * nrows)
)
axes = axes.ravel()
i_ax = 0
for dataset_name, (X, y) in datasets:
msg = f"Benchmarking on {dataset_name}:"
print(f"\n{msg}\n" + str("-" * len(msg)))
for transformer_name, transformer in transformers:
longest = np.max([len(name) for name, model in transformers])
start = time.time()
Xt = transformer.fit_transform(X)
transform_duration = time.time() - start
print(
f"{transformer_name:<{longest}} {transform_duration:.3f} sec"
" (fit_transform)"
)
# 绘制 TSNE 嵌入图,该图在不同方法之间应该非常相似
axes[i_ax].set_title(transformer_name + "\non " + dataset_name)
axes[i_ax].scatter(
Xt[:, 0],
Xt[:, 1],
c=y.astype(np.int32),
alpha=0.2,
cmap=plt.cm.viridis,
)
axes[i_ax].xaxis.set_major_formatter(NullFormatter())
axes[i_ax].yaxis.set_major_formatter(NullFormatter())
axes[i_ax].axis("tight")
i_ax += 1
fig.tight_layout()
plt.show()
示例输出:
在MNIST_10000上的基准测试:#
使用内部NearestNeighbors的TSNE 24.828秒(fit_transform) 使用KNeighborsTransformer的TSNE 20.111秒(fit_transform) 使用NMSlibTransformer的TSNE 21.757秒(fit_transform)
在MNIST_20000上的基准测试:#
使用内部NearestNeighbors的TSNE 51.955秒(fit_transform) 使用KNeighborsTransformer的TSNE 50.994秒(fit_transform) 使用NMSlibTransformer的TSNE 43.536秒(fit_transform)
我们可以观察到,默认的 TSNE
估计器及其内部的 NearestNeighbors
实现,在性能方面大致等同于包含 TSNE
和 KNeighborsTransformer
的管道。这是预期的,因为这两个管道在内部都依赖于相同的 NearestNeighbors
实现来执行精确的邻居搜索。近似的 NMSlibTransformer
在最小的数据集上已经比精确搜索稍快,但这种速度差异预计在样本数量较大的数据集上会更加显著。
请注意,并非所有的近似搜索方法都能保证比默认的精确搜索方法提高速度:实际上,自从scikit-learn 1.1以来,精确搜索的实现已经显著改进。此外,蛮力精确搜索方法在 fit
时不需要构建索引。因此,要在:class:~sklearn.manifold.TSNE
管道的上下文中获得整体性能提升,近似搜索在 transform
时的收益需要大于在 fit
时构建近似搜索索引所花费的额外时间。
最终,TSNE算法本身也是计算密集型的,与最近邻搜索无关。因此,将最近邻搜索步骤加速5倍,并不会使整个流程加速5倍。
Related examples