data_manager.py 10.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
# Copyright (c) 2020 VisualDL Authors. All Rights Reserve.
#
# 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 __future__ import absolute_import
import threading
import random
import collections

DEFAULT_PLUGIN_MAXSIZE = {
    "scalar": 300,
    "image": 10,
24 25
    "histogram": 100,
    "embeddings": 50000,
走神的阿圆's avatar
走神的阿圆 已提交
26 27
    "audio": 10,
    "pr_curve": 300
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
}


class Reservoir(object):
    """A map-to-arrays dict, with deterministic Reservoir Sampling.

    Store each reservoir bucket by key, and each bucket is a list sampling
    with reservoir algorithm.
    """

    def __init__(self, max_size, seed=0):
        """Creates a new reservoir.

        Args:
            max_size: The number of values to keep in the reservoir for each tag,
                if max_size is zero, all values will be kept in bucket.
            seed: The seed to initialize a random.Random().
            num_item_index: The index of data to add.

        Raises:
          ValueError: If max_size is not a nonnegative integer.
        """
        if max_size < 0 or max_size != round(max_size):
            raise ValueError("Max_size must be nonnegative integer.")
        self._max_size = max_size
        self._buckets = collections.defaultdict(
54 55
            lambda: _ReservoirBucket(max_size=self._max_size,
                                     random_instance=random.Random(seed))
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
        )
        self._mutex = threading.Lock()

    @property
    def keys(self):
        """Return all keys in self._buckets.

        Returns:
            All keys in reservoir buckets.
        :return:
        """
        with self._mutex:
            return list(self._buckets.keys())

    def _exist_in_keys(self, key):
        """Determine if key exists.

        Args:
            key: Key to determine if exists.

        Returns:
            True if key exists in buckets.keys, otherwise False.
        """
        return True if key in self._buckets.keys() else False

81 82
    def exist_in_keys(self, run, tag):
        """Determine if run_tag exists.
83 84 85 86

        For usage habits of VisualDL, actually call self._exist_in_keys()

        Args:
87
            run: Identity of one tablet.
88 89 90
            tag: Identity of one record in tablet.

        Returns:
91
            True if run_tag exists in buckets.keys, otherwise False.
92
        """
93
        key = run + "/" + tag
94 95 96 97 98 99 100 101
        return self._exist_in_keys(key)

    def _get_num_items_index(self, key):
        keys = self.keys
        if key not in keys:
            raise KeyError("Key %s not in buckets.keys()" % key)
        return self._buckets[key].num_items_index

102 103
    def get_num_items_index(self, run, tag):
        key = run + "/" + tag
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
        return self._get_num_items_index(key)

    def _get_items(self, key):
        """Get items with tag "key"

        Args:
            key: Key to finding bucket in reservoir buckets.

        Returns:
            One bucket in reservoir buckets by key.
        """
        keys = self.keys
        with self._mutex:
            if key not in keys:
                raise KeyError("Key %s not in buckets.keys()" % key)
            return self._buckets[key].items

121 122
    def get_items(self, run, tag):
        """Get items with tag 'run_tag'
123

124 125 126
        For usage habits of VisualDL, actually call self._get_items()

        Args:
127
            run: Identity of one tablet.
128 129 130
            tag: Identity of one record in tablet.

        Returns:
131
            One bucket in reservoir buckets by run and tag.
132
        """
133
        key = run + "/" + tag
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
        return self._get_items(key)

    def _add_item(self, key, item):
        """Add a new item to reservoir buckets with given tag as key.

        If bucket with key has not yet reached full size, each item will be
        added.

        If bucket with key is full, each item will be added with same
        probability.

        Add new item to buckets will always valid because self._buckets is a
        collection.defaultdict.

        Args:
            key: Tag of one bucket to add new item.
            item: New item to add to bucket.
        """
        with self._mutex:
            self._buckets[key].add_item(item)

155
    def add_item(self, run, tag, item):
156 157 158 159 160
        """Add a new item to reservoir buckets with given tag as key.

        For usage habits of VisualDL, actually call self._add_items()

        Args:
161
            run: Identity of one tablet.
162 163 164
            tag: Identity of one record in tablet.
            item: New item to add to bucket.
        """
165
        key = run + "/" + tag
166 167
        self._add_item(key, item)

168 169 170 171
    def _cut_tail(self, key):
        with self._mutex:
            self._buckets[key].cut_tail()

172
    def cut_tail(self, run, tag):
173
        """Pop the last item in reservoir buckets.
174

175 176 177 178
        Sometimes the tail of the retrieved data is abnormal 0. This
        method is used to handle this problem.

        Args:
179
            run: Identity of one tablet.
180 181
            tag: Identity of one record in tablet.
        """
182
        key = run + "/" + tag
183 184
        self._cut_tail(key)

185 186 187 188

class _ReservoirBucket(object):
    """Data manager for sampling data, use reservoir sampling.
    """
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
    def __init__(self, max_size, random_instance=None):
        """Create a _ReservoirBucket instance.

        Args:
            max_size: The maximum size of reservoir bucket. If max_size is
                zero, the bucket has unbounded size.
            random_instance: The random number generator. If not specified,
                default to random.Random(0)
            num_item_index: The index of data to add.

        Raises:
            ValueError: If args max_size is not a nonnegative integer.
        """
        if max_size < 0 or max_size != round(max_size):
            raise ValueError("Max_size must be nonnegative integer.")
        self._max_size = max_size
        self._random = random_instance if random_instance is not None else \
            random.Random(0)
        self._items = []
        self._mutex = threading.Lock()
        self._num_items_index = 0

    def add_item(self, item):
        """ Add an item to bucket, replacing an old item with probability.

        Use reservoir sampling to add a new item to sampling bucket,
        each item in a steam has same probability stay in the bucket.

        Args:
            item: The item to add to reservoir bucket.
        """
        with self._mutex:
            if len(self._items) < self._max_size or self._max_size == 0:
                self._items.append(item)
            else:
                r = self._random.randint(0, self._num_items_index)
                if r < self._max_size:
                    self._items.pop(r)
                    self._items.append(item)
            self._num_items_index += 1

    @property
    def items(self):
        """Get self._items

        Returns:
            All items.
        """
        with self._mutex:
            return self._items

    @property
    def num_items_index(self):
        with self._mutex:
            return self._num_items_index

246 247 248 249 250 251 252 253 254 255
    def cut_tail(self):
        """Pop the last item in reservoir buckets.

        Sometimes the tail of the retrieved data is abnormal 0. This
        method is used to handle this problem.
        """
        with self._mutex:
            self._items.pop()
            self._num_items_index -= 1

256 257 258 259

class DataManager(object):
    """Data manager for all plugin.
    """
260

261 262 263 264 265 266 267
    def __init__(self):
        """Create a data manager for all plugin.

        All kinds of plugin has own reservoir, stored in a dict with plugin
        name as key.

        """
268 269 270 271 272 273 274 275 276 277
        self._reservoirs = {
            "scalar":
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["scalar"]),
            "histogram":
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["histogram"]),
            "image":
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["image"]),
            "embeddings":
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["embeddings"]),
            "audio":
走神的阿圆's avatar
走神的阿圆 已提交
278 279 280
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["audio"]),
            "pr_curve":
            Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE["pr_curve"])
281
        }
282 283
        self._mutex = threading.Lock()

284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
    def add_reservoir(self, plugin):
        """Add reservoir to reservoirs.

        Every reservoir is attached to one plugin.

        Args:
            plugin: Key to get one reservoir bucket for one specified plugin.
        """
        with self._mutex:
            if plugin not in self._reservoirs.keys():
                self._reservoirs.update({
                    plugin:
                    Reservoir(max_size=DEFAULT_PLUGIN_MAXSIZE[plugin])
                })

299 300 301 302 303 304 305 306 307 308 309 310 311 312
    def get_reservoir(self, plugin):
        """Get reservoir by plugin as key.

        Args:
            plugin: Key to get one reservoir bucket for one specified plugin.

        Returns:
            Reservoir bucket for plugin.
        """
        with self._mutex:
            if plugin not in self._reservoirs.keys():
                raise KeyError("Key %s not in reservoirs." % plugin)
            return self._reservoirs[plugin]

313
    def add_item(self, plugin, run, tag, item):
314 315
        """Add item to one plugin reservoir bucket.

316
        Use 'run', 'tag' for usage habits of VisualDL.
317 318 319

        Args:
            plugin: Key to get one reservoir bucket.
320
            run: Each tablet has different 'run'.
321 322 323 324
            tag: Tag will be used to generate paths of tablets.
            item: The item to add to reservoir bucket.
        """
        with self._mutex:
325
            self._reservoirs[plugin].add_item(run, tag, item)
326 327 328 329 330 331 332 333 334 335 336 337

    def get_keys(self):
        """Get all plugin buckets name.

        Returns:
            All plugin keys.
        """
        with self._mutex:
            return self._reservoirs.keys()


default_data_manager = DataManager()