Coverage for /pythoncovmergedfiles/medio/medio/usr/local/lib/python3.8/site-packages/keras/src/layers/preprocessing/integer_lookup.py: 31%
36 statements
« prev ^ index » next coverage.py v7.4.0, created at 2024-01-03 07:57 +0000
1# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Keras string lookup preprocessing layer."""
18import numpy as np
19import tensorflow.compat.v2 as tf
21from keras.src.engine import base_preprocessing_layer
22from keras.src.layers.preprocessing import index_lookup
24# isort: off
25from tensorflow.python.platform import tf_logging as logging
26from tensorflow.python.util.tf_export import keras_export
29@keras_export(
30 "keras.layers.IntegerLookup",
31 "keras.layers.experimental.preprocessing.IntegerLookup",
32 v1=[],
33)
34class IntegerLookup(index_lookup.IndexLookup):
35 """A preprocessing layer which maps integer features to contiguous ranges.
37 This layer maps a set of arbitrary integer input tokens into indexed integer
38 output via a table-based vocabulary lookup. The layer's output indices will
39 be contiguously arranged up to the maximum vocab size, even if the input
40 tokens are non-continguous or unbounded. The layer supports multiple options
41 for encoding the output via `output_mode`, and has optional support for
42 out-of-vocabulary (OOV) tokens and masking.
44 The vocabulary for the layer must be either supplied on construction or
45 learned via `adapt()`. During `adapt()`, the layer will analyze a data set,
46 determine the frequency of individual integer tokens, and create a
47 vocabulary from them. If the vocabulary is capped in size, the most frequent
48 tokens will be used to create the vocabulary and all others will be treated
49 as OOV.
51 There are two possible output modes for the layer. When `output_mode` is
52 `"int"`, input integers are converted to their index in the vocabulary (an
53 integer). When `output_mode` is `"multi_hot"`, `"count"`, or `"tf_idf"`,
54 input integers are encoded into an array where each dimension corresponds to
55 an element in the vocabulary.
57 The vocabulary can optionally contain a mask token as well as an OOV token
58 (which can optionally occupy multiple indices in the vocabulary, as set
59 by `num_oov_indices`).
60 The position of these tokens in the vocabulary is fixed. When `output_mode`
61 is `"int"`, the vocabulary will begin with the mask token at index 0,
62 followed by OOV indices, followed by the rest of the vocabulary. When
63 `output_mode` is `"multi_hot"`, `"count"`, or `"tf_idf"` the vocabulary will
64 begin with OOV indices and instances of the mask token will be dropped.
66 For an overview and full list of preprocessing layers, see the preprocessing
67 [guide](https://www.tensorflow.org/guide/keras/preprocessing_layers).
69 Args:
70 max_tokens: Maximum size of the vocabulary for this layer. This should
71 only be specified when adapting the vocabulary or when setting
72 `pad_to_max_tokens=True`. If None, there is no cap on the size of the
73 vocabulary. Note that this size includes the OOV and mask tokens.
74 Defaults to `None`.
75 num_oov_indices: The number of out-of-vocabulary tokens to use. If this
76 value is more than 1, OOV inputs are modulated to determine their OOV
77 value. If this value is 0, OOV inputs will cause an error when calling
78 the layer. Defaults to `1`.
79 mask_token: An integer token that represents masked inputs. When
80 `output_mode` is `"int"`, the token is included in vocabulary and mapped
81 to index 0. In other output modes, the token will not appear in the
82 vocabulary and instances of the mask token in the input will be dropped.
83 If set to None, no mask term will be added. Defaults to `None`.
84 oov_token: Only used when `invert` is True. The token to return for OOV
85 indices. Defaults to `-1`.
86 vocabulary: Optional. Either an array of integers or a string path to a
87 text file. If passing an array, can pass a tuple, list, 1D numpy array,
88 or 1D tensor containing the integer vocbulary terms. If passing a file
89 path, the file should contain one line per term in the vocabulary. If
90 this argument is set, there is no need to `adapt()` the layer.
91 vocabulary_dtype: The dtype of the vocabulary terms, for example
92 `"int64"` or `"int32"`. Defaults to `"int64"`.
93 idf_weights: Only valid when `output_mode` is `"tf_idf"`. A tuple, list,
94 1D numpy array, or 1D tensor or the same length as the vocabulary,
95 containing the floating point inverse document frequency weights, which
96 will be multiplied by per sample term counts for the final `tf_idf`
97 weight. If the `vocabulary` argument is set, and `output_mode` is
98 `"tf_idf"`, this argument must be supplied.
99 invert: Only valid when `output_mode` is `"int"`. If True, this layer will
100 map indices to vocabulary items instead of mapping vocabulary items to
101 indices. Defaults to `False`.
102 output_mode: Specification for the output of the layer. Values can be
103 `"int"`, `"one_hot"`, `"multi_hot"`, `"count"`, or `"tf_idf"`
104 configuring the layer as follows:
105 - `"int"`: Return the vocabulary indices of the input tokens.
106 - `"one_hot"`: Encodes each individual element in the input into an
107 array the same size as the vocabulary, containing a 1 at the element
108 index. If the last dimension is size 1, will encode on that
109 dimension. If the last dimension is not size 1, will append a new
110 dimension for the encoded output.
111 - `"multi_hot"`: Encodes each sample in the input into a single array
112 the same size as the vocabulary, containing a 1 for each vocabulary
113 term present in the sample. Treats the last dimension as the sample
114 dimension, if input shape is (..., sample_length), output shape will
115 be (..., num_tokens).
116 - `"count"`: As `"multi_hot"`, but the int array contains a count of
117 the number of times the token at that index appeared in the sample.
118 - `"tf_idf"`: As `"multi_hot"`, but the TF-IDF algorithm is applied to
119 find the value in each token slot.
120 For `"int"` output, any shape of input and output is supported. For all
121 other output modes, currently only output up to rank 2 is supported.
122 Defaults to `"int"`.
123 pad_to_max_tokens: Only applicable when `output_mode` is `"multi_hot"`,
124 `"count"`, or `"tf_idf"`. If True, the output will have its feature axis
125 padded to `max_tokens` even if the number of unique tokens in the
126 vocabulary is less than max_tokens, resulting in a tensor of shape
127 [batch_size, max_tokens] regardless of vocabulary size. Defaults to
128 False.
129 sparse: Boolean. Only applicable when `output_mode` is `"multi_hot"`,
130 `"count"`, or `"tf_idf"`. If True, returns a `SparseTensor` instead of a
131 dense `Tensor`. Defaults to `False`.
133 Examples:
135 **Creating a lookup layer with a known vocabulary**
137 This example creates a lookup layer with a pre-existing vocabulary.
139 >>> vocab = [12, 36, 1138, 42]
140 >>> data = tf.constant([[12, 1138, 42], [42, 1000, 36]]) # Note OOV tokens
141 >>> layer = tf.keras.layers.IntegerLookup(vocabulary=vocab)
142 >>> layer(data)
143 <tf.Tensor: shape=(2, 3), dtype=int64, numpy=
144 array([[1, 3, 4],
145 [4, 0, 2]])>
147 **Creating a lookup layer with an adapted vocabulary**
149 This example creates a lookup layer and generates the vocabulary by
150 analyzing the dataset.
152 >>> data = tf.constant([[12, 1138, 42], [42, 1000, 36]])
153 >>> layer = tf.keras.layers.IntegerLookup()
154 >>> layer.adapt(data)
155 >>> layer.get_vocabulary()
156 [-1, 42, 1138, 1000, 36, 12]
158 Note that the OOV token -1 have been added to the vocabulary. The remaining
159 tokens are sorted by frequency (42, which has 2 occurrences, is first) then
160 by inverse sort order.
162 >>> data = tf.constant([[12, 1138, 42], [42, 1000, 36]])
163 >>> layer = tf.keras.layers.IntegerLookup()
164 >>> layer.adapt(data)
165 >>> layer(data)
166 <tf.Tensor: shape=(2, 3), dtype=int64, numpy=
167 array([[5, 2, 1],
168 [1, 3, 4]])>
171 **Lookups with multiple OOV indices**
173 This example demonstrates how to use a lookup layer with multiple OOV
174 indices. When a layer is created with more than one OOV index, any OOV
175 tokens are hashed into the number of OOV buckets, distributing OOV tokens in
176 a deterministic fashion across the set.
178 >>> vocab = [12, 36, 1138, 42]
179 >>> data = tf.constant([[12, 1138, 42], [37, 1000, 36]])
180 >>> layer = tf.keras.layers.IntegerLookup(
181 ... vocabulary=vocab, num_oov_indices=2)
182 >>> layer(data)
183 <tf.Tensor: shape=(2, 3), dtype=int64, numpy=
184 array([[2, 4, 5],
185 [1, 0, 3]])>
187 Note that the output for OOV token 37 is 1, while the output for OOV token
188 1000 is 0. The in-vocab terms have their output index increased by 1 from
189 earlier examples (12 maps to 2, etc) in order to make space for the extra
190 OOV token.
192 **One-hot output**
194 Configure the layer with `output_mode='one_hot'`. Note that the first
195 `num_oov_indices` dimensions in the ont_hot encoding represent OOV values.
197 >>> vocab = [12, 36, 1138, 42]
198 >>> data = tf.constant([12, 36, 1138, 42, 7]) # Note OOV tokens
199 >>> layer = tf.keras.layers.IntegerLookup(
200 ... vocabulary=vocab, output_mode='one_hot')
201 >>> layer(data)
202 <tf.Tensor: shape=(5, 5), dtype=float32, numpy=
203 array([[0., 1., 0., 0., 0.],
204 [0., 0., 1., 0., 0.],
205 [0., 0., 0., 1., 0.],
206 [0., 0., 0., 0., 1.],
207 [1., 0., 0., 0., 0.]], dtype=float32)>
209 **Multi-hot output**
211 Configure the layer with `output_mode='multi_hot'`. Note that the first
212 `num_oov_indices` dimensions in the multi_hot encoding represent OOV tokens
214 >>> vocab = [12, 36, 1138, 42]
215 >>> data = tf.constant([[12, 1138, 42, 42],
216 ... [42, 7, 36, 7]]) # Note OOV tokens
217 >>> layer = tf.keras.layers.IntegerLookup(
218 ... vocabulary=vocab, output_mode='multi_hot')
219 >>> layer(data)
220 <tf.Tensor: shape=(2, 5), dtype=float32, numpy=
221 array([[0., 1., 0., 1., 1.],
222 [1., 0., 1., 0., 1.]], dtype=float32)>
224 **Token count output**
226 Configure the layer with `output_mode='count'`. As with multi_hot output,
227 the first `num_oov_indices` dimensions in the output represent OOV tokens.
229 >>> vocab = [12, 36, 1138, 42]
230 >>> data = tf.constant([[12, 1138, 42, 42],
231 ... [42, 7, 36, 7]]) # Note OOV tokens
232 >>> layer = tf.keras.layers.IntegerLookup(
233 ... vocabulary=vocab, output_mode='count')
234 >>> layer(data)
235 <tf.Tensor: shape=(2, 5), dtype=float32, numpy=
236 array([[0., 1., 0., 1., 2.],
237 [2., 0., 1., 0., 1.]], dtype=float32)>
239 **TF-IDF output**
241 Configure the layer with `output_mode='tf_idf'`. As with multi_hot output,
242 the first `num_oov_indices` dimensions in the output represent OOV tokens.
244 Each token bin will output `token_count * idf_weight`, where the idf weights
245 are the inverse document frequency weights per token. These should be
246 provided along with the vocabulary. Note that the `idf_weight` for OOV
247 tokens will default to the average of all idf weights passed in.
249 >>> vocab = [12, 36, 1138, 42]
250 >>> idf_weights = [0.25, 0.75, 0.6, 0.4]
251 >>> data = tf.constant([[12, 1138, 42, 42],
252 ... [42, 7, 36, 7]]) # Note OOV tokens
253 >>> layer = tf.keras.layers.IntegerLookup(
254 ... output_mode='tf_idf', vocabulary=vocab, idf_weights=idf_weights)
255 >>> layer(data)
256 <tf.Tensor: shape=(2, 5), dtype=float32, numpy=
257 array([[0. , 0.25, 0. , 0.6 , 0.8 ],
258 [1.0 , 0. , 0.75, 0. , 0.4 ]], dtype=float32)>
260 To specify the idf weights for oov tokens, you will need to pass the entire
261 vocabularly including the leading oov token.
263 >>> vocab = [-1, 12, 36, 1138, 42]
264 >>> idf_weights = [0.9, 0.25, 0.75, 0.6, 0.4]
265 >>> data = tf.constant([[12, 1138, 42, 42],
266 ... [42, 7, 36, 7]]) # Note OOV tokens
267 >>> layer = tf.keras.layers.IntegerLookup(
268 ... output_mode='tf_idf', vocabulary=vocab, idf_weights=idf_weights)
269 >>> layer(data)
270 <tf.Tensor: shape=(2, 5), dtype=float32, numpy=
271 array([[0. , 0.25, 0. , 0.6 , 0.8 ],
272 [1.8 , 0. , 0.75, 0. , 0.4 ]], dtype=float32)>
274 When adapting the layer in tf_idf mode, each input sample will be considered
275 a document, and idf weight per token will be calculated as
276 `log(1 + num_documents / (1 + token_document_count))`.
278 **Inverse lookup**
280 This example demonstrates how to map indices to tokens using this layer.
281 (You can also use `adapt()` with `inverse=True`, but for simplicity we'll
282 pass the vocab in this example.)
284 >>> vocab = [12, 36, 1138, 42]
285 >>> data = tf.constant([[1, 3, 4], [4, 0, 2]])
286 >>> layer = tf.keras.layers.IntegerLookup(vocabulary=vocab, invert=True)
287 >>> layer(data)
288 <tf.Tensor: shape=(2, 3), dtype=int64, numpy=
289 array([[ 12, 1138, 42],
290 [ 42, -1, 36]])>
292 Note that the first index correspond to the oov token by default.
295 **Forward and inverse lookup pairs**
297 This example demonstrates how to use the vocabulary of a standard lookup
298 layer to create an inverse lookup layer.
300 >>> vocab = [12, 36, 1138, 42]
301 >>> data = tf.constant([[12, 1138, 42], [42, 1000, 36]])
302 >>> layer = tf.keras.layers.IntegerLookup(vocabulary=vocab)
303 >>> i_layer = tf.keras.layers.IntegerLookup(
304 ... vocabulary=layer.get_vocabulary(), invert=True)
305 >>> int_data = layer(data)
306 >>> i_layer(int_data)
307 <tf.Tensor: shape=(2, 3), dtype=int64, numpy=
308 array([[ 12, 1138, 42],
309 [ 42, -1, 36]])>
311 In this example, the input token 1000 resulted in an output of -1, since
312 1000 was not in the vocabulary - it got represented as an OOV, and all OOV
313 tokens are returned as -1 in the inverse layer. Also, note that for the
314 inverse to work, you must have already set the forward layer vocabulary
315 either directly or via `adapt()` before calling `get_vocabulary()`.
316 """
318 def __init__(
319 self,
320 max_tokens=None,
321 num_oov_indices=1,
322 mask_token=None,
323 oov_token=-1,
324 vocabulary=None,
325 vocabulary_dtype="int64",
326 idf_weights=None,
327 invert=False,
328 output_mode="int",
329 sparse=False,
330 pad_to_max_tokens=False,
331 **kwargs,
332 ):
333 if not tf.dtypes.as_dtype(vocabulary_dtype).is_integer:
334 raise ValueError(
335 "`vocabulary_dtype` must be an integer dtype. "
336 f"Received: {vocabulary_dtype}"
337 )
339 # Legacy versions of the IntegerLookup layer set layer dtype to int64,
340 # instead of the output type. If we see this and output mode is not
341 # "int", clear the setting so we don't switch types for old SavedModels.
342 if (
343 output_mode != "int"
344 and "dtype" in kwargs
345 and (kwargs["dtype"] == tf.int64 or kwargs["dtype"] == "int64")
346 ):
347 del kwargs["dtype"]
349 # Support deprecated args for this layer.
350 if "max_values" in kwargs:
351 logging.log_first_n(
352 logging.WARN,
353 "max_values is deprecated, use max_tokens instead.",
354 1,
355 )
356 max_tokens = kwargs["max_values"]
357 del kwargs["max_values"]
358 if "mask_value" in kwargs:
359 logging.log_first_n(
360 logging.WARN,
361 "mask_value is deprecated, use mask_token instead.",
362 1,
363 )
364 mask_token = kwargs["mask_value"]
365 del kwargs["mask_value"]
366 if "oov_value" in kwargs:
367 logging.log_first_n(
368 logging.WARN,
369 "oov_value is deprecated, use oov_token instead.",
370 1,
371 )
372 oov_token = kwargs["oov_value"]
373 del kwargs["oov_value"]
375 # If max_tokens is set, the token must be greater than 1 - otherwise we
376 # are creating a 0-element vocab, which doesn't make sense.
377 if max_tokens is not None and max_tokens <= 1:
378 raise ValueError(
379 "If `max_tokens` is set for `IntegerLookup`, it must be "
380 f"greater than 1. Received: max_tokens={max_tokens}."
381 )
383 if num_oov_indices < 0:
384 raise ValueError(
385 "The value of `num_oov_indices` argument for `IntegerLookup` "
386 "must >= 0. Received num_oov_indices="
387 f"{num_oov_indices}."
388 )
390 # Make sure mask and oov are of the dtype we want.
391 mask_token = None if mask_token is None else np.int64(mask_token)
392 oov_token = None if oov_token is None else np.int64(oov_token)
394 super().__init__(
395 max_tokens=max_tokens,
396 num_oov_indices=num_oov_indices,
397 mask_token=mask_token,
398 oov_token=oov_token,
399 vocabulary=vocabulary,
400 vocabulary_dtype=vocabulary_dtype,
401 idf_weights=idf_weights,
402 invert=invert,
403 output_mode=output_mode,
404 sparse=sparse,
405 pad_to_max_tokens=pad_to_max_tokens,
406 **kwargs,
407 )
408 base_preprocessing_layer.keras_kpl_gauge.get_cell("IntegerLookup").set(
409 True
410 )
412 # We override this method solely to generate a docstring.
413 def adapt(self, data, batch_size=None, steps=None):
414 """Computes a vocabulary of interger terms from tokens in a dataset.
416 Calling `adapt()` on an `IntegerLookup` layer is an alternative to
417 passing in a precomputed vocabulary on construction via the
418 `vocabulary` argument. An `IntegerLookup` layer should always be either
419 adapted over a dataset or supplied with a vocabulary.
421 During `adapt()`, the layer will build a vocabulary of all integer
422 tokens seen in the dataset, sorted by occurrence count, with ties broken
423 by sort order of the tokens (high to low). At the end of `adapt()`, if
424 `max_tokens` is set, the vocabulary wil be truncated to `max_tokens`
425 size. For example, adapting a layer with `max_tokens=1000` will compute
426 the 1000 most frequent tokens occurring in the input dataset. If
427 `output_mode='tf-idf'`, `adapt()` will also learn the document
428 frequencies of each token in the input dataset.
430 In order to make `StringLookup` efficient in any distribution context,
431 the vocabulary is kept static with respect to any compiled `tf.Graph`s
432 that call the layer. As a consequence, if the layer is adapted a second
433 time, any models using the layer should be re-compiled. For more
434 information see
435 `tf.keras.layers.experimental.preprocessing.PreprocessingLayer.adapt`.
437 `adapt()` is meant only as a single machine utility to compute layer
438 state. To analyze a dataset that cannot fit on a single machine, see
439 [Tensorflow Transform](
440 https://www.tensorflow.org/tfx/transform/get_started) for a
441 multi-machine, map-reduce solution.
443 Arguments:
444 data: The data to train on. It can be passed either as a
445 `tf.data.Dataset`, or as a numpy array.
446 batch_size: Integer or `None`.
447 Number of samples per state update.
448 If unspecified, `batch_size` will default to 32.
449 Do not specify the `batch_size` if your data is in the
450 form of datasets, generators, or `keras.utils.Sequence` instances
451 (since they generate batches).
452 steps: Integer or `None`.
453 Total number of steps (batches of samples)
454 When training with input tensors such as
455 TensorFlow data tensors, the default `None` is equal to
456 the number of samples in your dataset divided by
457 the batch size, or 1 if that cannot be determined. If x is a
458 `tf.data` dataset, and 'steps' is None, the epoch will run until
459 the input dataset is exhausted. When passing an infinitely
460 repeating dataset, you must specify the `steps` argument. This
461 argument is not supported with array inputs.
462 """
463 super().adapt(data, batch_size=batch_size, steps=steps)