-
Notifications
You must be signed in to change notification settings - Fork 527
/
kneighbors_classifier.py
452 lines (396 loc) · 15 KB
/
kneighbors_classifier.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
#
# Copyright (c) 2020-2023, NVIDIA CORPORATION.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from cuml.internals.safe_imports import gpu_only_import
from cuml.dask.common.input_utils import DistributedDataHandler
from cuml.dask.common.input_utils import to_output
from cuml.dask.common import parts_to_ranks
from cuml.dask.common import flatten_grouped_results
from cuml.dask.common.utils import raise_mg_import_exception
from cuml.dask.common.utils import wait_and_raise_from_futures
from raft_dask.common.comms import get_raft_comm_state
from cuml.dask.neighbors import NearestNeighbors
from dask.dataframe import Series as DaskSeries
from dask.distributed import get_worker
import dask.array as da
from uuid import uuid1
from cuml.internals.safe_imports import cpu_only_import
np = cpu_only_import("numpy")
pd = cpu_only_import("pandas")
cudf = gpu_only_import("cudf")
class KNeighborsClassifier(NearestNeighbors):
"""
Multi-node Multi-GPU K-Nearest Neighbors Classifier Model.
K-Nearest Neighbors Classifier is an instance-based learning technique,
that keeps training samples around for prediction, rather than trying
to learn a generalizable set of model parameters.
Parameters
----------
n_neighbors : int (default=5)
Default number of neighbors to query
batch_size: int (optional, default 2000000)
Maximum number of query rows processed at once. This parameter can
greatly affect the throughput of the algorithm. The optimal setting
of this value will vary for different layouts and index to query
ratios, but it will require `batch_size * n_features * 4` bytes of
additional memory on each worker hosting index partitions.
handle : cuml.Handle
Specifies the cuml.handle that holds internal CUDA state for
computations in this model. Most importantly, this specifies the CUDA
stream that will be used for the model's computations, so users can
run different models concurrently in different streams by creating
handles in several streams.
If it is None, a new one is created.
verbose : int or boolean, default=False
Sets logging level. It must be one of `cuml.common.logger.level_*`.
See :ref:`verbosity-levels` for more info.
"""
def __init__(
self, *, client=None, streams_per_handle=0, verbose=False, **kwargs
):
super().__init__(client=client, verbose=verbose, **kwargs)
self.streams_per_handle = streams_per_handle
def fit(self, X, y):
"""
Fit a multi-node multi-GPU K-Nearest Neighbors Classifier index
Parameters
----------
X : array-like (device or host) shape = (n_samples, n_features)
Index data.
Acceptable formats: dask CuPy/NumPy/Numba Array
y : array-like (device or host) shape = (n_samples, n_features)
Index labels data.
Acceptable formats: dask CuPy/NumPy/Numba Array
Returns
-------
self : KNeighborsClassifier model
"""
if not isinstance(X._meta, (np.ndarray, pd.DataFrame, cudf.DataFrame)):
raise ValueError("This chunk type is not supported")
self.data_handler = DistributedDataHandler.create(
data=[X, y], client=self.client
)
# uniq_labels: set of possible labels for each labels column
# n_unique: number of possible labels for each labels column
uniq_labels = []
if self.data_handler.datatype == "cupy":
if y.ndim == 1:
uniq_labels.append(da.unique(y))
else:
n_targets = y.shape[1]
for i in range(n_targets):
uniq_labels.append(da.unique(y[:, i]))
else:
if isinstance(y, DaskSeries):
uniq_labels.append(y.unique())
else:
n_targets = len(y.columns)
for i in range(n_targets):
uniq_labels.append(y.iloc[:, i].unique())
uniq_labels = da.compute(uniq_labels)[0]
if hasattr(uniq_labels[0], "values_host"): # for cuDF Series
uniq_labels = list(map(lambda x: x.values_host, uniq_labels))
elif hasattr(uniq_labels[0], "values"): # for pandas Series
uniq_labels = list(map(lambda x: x.values, uniq_labels))
self.uniq_labels = np.sort(np.array(uniq_labels))
self.n_unique = list(map(lambda x: len(x), self.uniq_labels))
return self
@staticmethod
def _func_create_model(sessionId, **kwargs):
try:
from cuml.neighbors.kneighbors_classifier_mg import (
KNeighborsClassifierMG as cumlKNN,
)
except ImportError:
raise_mg_import_exception()
handle = get_raft_comm_state(sessionId, get_worker())["handle"]
return cumlKNN(handle=handle, **kwargs)
@staticmethod
def _func_predict(
model,
index,
index_parts_to_ranks,
index_nrows,
query,
query_parts_to_ranks,
query_nrows,
uniq_labels,
n_unique,
ncols,
rank,
convert_dtype,
probas_only,
):
if probas_only:
return model.predict_proba(
index,
index_parts_to_ranks,
index_nrows,
query,
query_parts_to_ranks,
query_nrows,
uniq_labels,
n_unique,
ncols,
rank,
convert_dtype,
)
else:
return model.predict(
index,
index_parts_to_ranks,
index_nrows,
query,
query_parts_to_ranks,
query_nrows,
uniq_labels,
n_unique,
ncols,
rank,
convert_dtype,
)
def predict(self, X, convert_dtype=True):
"""
Predict labels for a query from previously stored index
and index labels.
The process is done in a multi-node multi-GPU fashion.
Parameters
----------
X : array-like (device or host) shape = (n_samples, n_features)
Query data.
Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
convert_dtype : bool, optional (default = True)
When set to True, the predict method will automatically
convert the data to the right formats.
Returns
-------
predictions : Dask futures or Dask CuPy Arrays
"""
query_handler = DistributedDataHandler.create(
data=X, client=self.client
)
self.datatype = query_handler.datatype
comms = KNeighborsClassifier._build_comms(
self.data_handler, query_handler, self.streams_per_handle
)
worker_info = comms.worker_info(comms.worker_addresses)
"""
Build inputs and outputs
"""
self.data_handler.calculate_parts_to_sizes(comms=comms)
query_handler.calculate_parts_to_sizes(comms=comms)
data_parts_to_ranks, data_nrows = parts_to_ranks(
self.client, worker_info, self.data_handler.gpu_futures
)
query_parts_to_ranks, query_nrows = parts_to_ranks(
self.client, worker_info, query_handler.gpu_futures
)
"""
Each Dask worker creates a single model
"""
key = uuid1()
models = dict(
[
(
worker,
self.client.submit(
self._func_create_model,
comms.sessionId,
**self.kwargs,
workers=[worker],
key="%s-%s" % (key, idx),
),
)
for idx, worker in enumerate(comms.worker_addresses)
]
)
"""
Invoke knn_classify on Dask workers to perform distributed query
"""
key = uuid1()
knn_clf_res = dict(
[
(
worker_info[worker]["rank"],
self.client.submit(
self._func_predict,
models[worker],
self.data_handler.worker_to_parts[worker]
if worker in self.data_handler.workers
else [],
data_parts_to_ranks,
data_nrows,
query_handler.worker_to_parts[worker]
if worker in query_handler.workers
else [],
query_parts_to_ranks,
query_nrows,
self.uniq_labels,
self.n_unique,
X.shape[1],
worker_info[worker]["rank"],
convert_dtype,
False,
key="%s-%s" % (key, idx),
workers=[worker],
),
)
for idx, worker in enumerate(comms.worker_addresses)
]
)
wait_and_raise_from_futures(list(knn_clf_res.values()))
"""
Gather resulting partitions and return result
"""
out_futures = flatten_grouped_results(
self.client, query_parts_to_ranks, knn_clf_res
)
comms.destroy()
return to_output(out_futures, self.datatype).squeeze()
def score(self, X, y, convert_dtype=True):
"""
Predict labels for a query from previously stored index
and index labels.
The process is done in a multi-node multi-GPU fashion.
Parameters
----------
X : array-like (device or host) shape = (n_samples, n_features)
Query test data.
Acceptable formats: dask CuPy/NumPy/Numba Array
y : array-like (device or host) shape = (n_samples, n_features)
Labels test data.
Acceptable formats: dask CuPy/NumPy/Numba Array
Returns
-------
score
"""
y_pred = self.predict(X, convert_dtype=convert_dtype)
if not isinstance(y_pred, da.Array):
y_pred = y_pred.to_dask_array(lengths=True)
if not isinstance(y, da.Array):
y = y.to_dask_array(lengths=True)
y_true = y.squeeze()
matched = y_pred == y_true
mean_match = matched.mean()
return float(mean_match.compute())
def predict_proba(self, X, convert_dtype=True):
"""
Provide score by comparing predictions and ground truth.
Parameters
----------
X : array-like (device or host) shape = (n_samples, n_features)
Query data.
Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
convert_dtype : bool, optional (default = True)
When set to True, the predict method will automatically
convert the data to the right formats.
Returns
-------
probabilities : Dask futures or Dask CuPy Arrays
"""
query_handler = DistributedDataHandler.create(
data=X, client=self.client
)
self.datatype = query_handler.datatype
comms = KNeighborsClassifier._build_comms(
self.data_handler, query_handler, self.streams_per_handle
)
worker_info = comms.worker_info(comms.worker_addresses)
"""
Build inputs and outputs
"""
self.data_handler.calculate_parts_to_sizes(comms=comms)
query_handler.calculate_parts_to_sizes(comms=comms)
data_parts_to_ranks, data_nrows = parts_to_ranks(
self.client, worker_info, self.data_handler.gpu_futures
)
query_parts_to_ranks, query_nrows = parts_to_ranks(
self.client, worker_info, query_handler.gpu_futures
)
"""
Each Dask worker creates a single model
"""
key = uuid1()
models = dict(
[
(
worker,
self.client.submit(
self._func_create_model,
comms.sessionId,
**self.kwargs,
workers=[worker],
key="%s-%s" % (key, idx),
),
)
for idx, worker in enumerate(comms.worker_addresses)
]
)
"""
Invoke knn_classify on Dask workers to perform distributed query
"""
key = uuid1()
knn_prob_res = dict(
[
(
worker_info[worker]["rank"],
self.client.submit(
self._func_predict,
models[worker],
self.data_handler.worker_to_parts[worker]
if worker in self.data_handler.workers
else [],
data_parts_to_ranks,
data_nrows,
query_handler.worker_to_parts[worker]
if worker in query_handler.workers
else [],
query_parts_to_ranks,
query_nrows,
self.uniq_labels,
self.n_unique,
X.shape[1],
worker_info[worker]["rank"],
convert_dtype,
True,
key="%s-%s" % (key, idx),
workers=[worker],
),
)
for idx, worker in enumerate(comms.worker_addresses)
]
)
wait_and_raise_from_futures(list(knn_prob_res.values()))
n_outputs = len(self.n_unique)
def _custom_getter(o):
def func_get(f, idx):
return f[o][idx]
return func_get
"""
Gather resulting partitions and return result
"""
outputs = []
for o in range(n_outputs):
futures = flatten_grouped_results(
self.client,
query_parts_to_ranks,
knn_prob_res,
getter_func=_custom_getter(o),
)
outputs.append(to_output(futures, self.datatype))
comms.destroy()
if n_outputs == 1:
return da.concatenate(outputs, axis=0)
return tuple(outputs)