1# Copyright 2017 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"""Lookup operations.""" 16 17from tensorflow.python.data.experimental.ops.cardinality import assert_cardinality 18from tensorflow.python.framework import dtypes 19from tensorflow.python.framework import ops 20from tensorflow.python.framework import tensor_spec 21from tensorflow.python.ops import gen_experimental_dataset_ops as ged_ops 22from tensorflow.python.ops import lookup_ops 23from tensorflow.python.ops import math_ops 24from tensorflow.python.util.tf_export import tf_export 25 26 27def _check_table_initializer_element_spec(element_spec): 28 """Raises an error if the given table initializer element spec is invalid.""" 29 base_error = ("Datasets used to initialize lookup tables must " 30 "produce elements in the form (key, value), where " 31 "the keys and values are scalar tensors. ") 32 specific_error = None 33 if len(element_spec) != 2: 34 raise ValueError(base_error + "However, the given dataset produces " 35 f"{len(element_spec)} components instead of two " 36 "(key, value) components. Full dataset element spec: " 37 f"{element_spec}.") 38 if not isinstance(element_spec[0], tensor_spec.TensorSpec): 39 raise ValueError(base_error + "However, the given dataset produces " 40 f"non-Tensor keys of type {type(element_spec[0])}.") 41 if not isinstance(element_spec[1], tensor_spec.TensorSpec): 42 raise ValueError(base_error + "However, the given dataset produces " 43 f"non-Tensor values of type {type(element_spec[1])}.") 44 if element_spec[0].shape.rank not in (None, 0): 45 raise ValueError( 46 base_error + "However, the given dataset produces " 47 f"non-scalar key Tensors of rank {element_spec[0].shape.rank}.") 48 if element_spec[1].shape.rank not in (None, 0): 49 raise ValueError( 50 base_error + "However, the given dataset produces " 51 f"non-scalar value Tensors of rank {element_spec[1].shape.rank}.") 52 53 54@tf_export("data.experimental.DatasetInitializer") 55class DatasetInitializer(lookup_ops.TableInitializerBase): 56 """Creates a table initializer from a `tf.data.Dataset`. 57 58 Sample usage: 59 60 >>> keys = tf.data.Dataset.range(100) 61 >>> values = tf.data.Dataset.range(100).map( 62 ... lambda x: tf.strings.as_string(x * 2)) 63 >>> ds = tf.data.Dataset.zip((keys, values)) 64 >>> init = tf.data.experimental.DatasetInitializer(ds) 65 >>> table = tf.lookup.StaticHashTable(init, "") 66 >>> table.lookup(tf.constant([0, 1, 2], dtype=tf.int64)).numpy() 67 array([b'0', b'2', b'4'], dtype=object) 68 69 Attributes: 70 dataset: A `tf.data.Dataset` object that produces tuples of scalars. The 71 first scalar is treated as a key and the second as value. 72 Raises: ValueError if `dataset` doesn't conform to specifications. 73 """ 74 75 def __init__(self, dataset): 76 """Creates a table initializer from a `tf.data.Dataset`. 77 78 Args: 79 dataset: A `tf.data.Dataset` object that produces tuples of scalars. The 80 first scalar is treated as a key and the second as value. 81 Raises: ValueError if `dataset` doesn't conform to specifications. 82 Returns: A `DatasetInitializer` object 83 """ 84 # Assert that the dataset element spec is a tuple of TensorSpecs where 85 # each tensor is a scalar. 86 self.dataset = dataset 87 elem_spec = self.dataset.element_spec 88 _check_table_initializer_element_spec(elem_spec) 89 90 key_type = elem_spec[0].dtype 91 value_type = elem_spec[1].dtype 92 super(DatasetInitializer, self).__init__(key_type, value_type) 93 94 def initialize(self, table): 95 lookup_ops.check_table_dtypes(table, self._key_dtype, self._value_dtype) 96 init_op = ged_ops.initialize_table_from_dataset( 97 table.resource_handle, self.dataset._variant_tensor) # pylint: disable=protected-access 98 ops.add_to_collection(ops.GraphKeys.TABLE_INITIALIZERS, init_op) 99 return init_op 100 101 102@tf_export("data.experimental.table_from_dataset") 103def table_from_dataset(dataset=None, 104 num_oov_buckets=0, 105 vocab_size=None, 106 default_value=None, 107 hasher_spec=lookup_ops.FastHashSpec, 108 key_dtype=dtypes.string, 109 name=None): 110 """Returns a lookup table based on the given dataset. 111 112 This operation constructs a lookup table based on the given dataset of pairs 113 of (key, value). 114 115 Any lookup of an out-of-vocabulary token will return a bucket ID based on its 116 hash if `num_oov_buckets` is greater than zero. Otherwise it is assigned the 117 `default_value`. 118 The bucket ID range is 119 `[vocabulary size, vocabulary size + num_oov_buckets - 1]`. 120 121 Sample Usages: 122 123 >>> keys = tf.data.Dataset.range(100) 124 >>> values = tf.data.Dataset.range(100).map( 125 ... lambda x: tf.strings.as_string(x * 2)) 126 >>> ds = tf.data.Dataset.zip((keys, values)) 127 >>> table = tf.data.experimental.table_from_dataset( 128 ... ds, default_value='n/a', key_dtype=tf.int64) 129 >>> table.lookup(tf.constant([0, 1, 2], dtype=tf.int64)).numpy() 130 array([b'0', b'2', b'4'], dtype=object) 131 132 Args: 133 dataset: A dataset containing (key, value) pairs. 134 num_oov_buckets: The number of out-of-vocabulary buckets. 135 vocab_size: Number of the elements in the vocabulary, if known. 136 default_value: The value to use for out-of-vocabulary feature values. 137 Defaults to -1. 138 hasher_spec: A `HasherSpec` to specify the hash function to use for 139 assignation of out-of-vocabulary buckets. 140 key_dtype: The `key` data type. 141 name: A name for this op (optional). 142 143 Returns: 144 The lookup table based on the given dataset. 145 146 Raises: 147 ValueError: If 148 * `dataset` does not contain pairs 149 * The 2nd item in the `dataset` pairs has a dtype which is incompatible 150 with `default_value` 151 * `num_oov_buckets` is negative 152 * `vocab_size` is not greater than zero 153 * The `key_dtype` is not integer or string 154 """ 155 elem_spec = dataset.element_spec 156 _check_table_initializer_element_spec(elem_spec) 157 if default_value is None: 158 default_value = -1 159 if not (elem_spec[1].dtype.is_integer or elem_spec[1].dtype.is_floating): 160 raise ValueError("`default_value` must be specified when creating a " 161 "table from a dataset that produces values of type " 162 f"{elem_spec[1].dtype}.") 163 if num_oov_buckets < 0: 164 raise ValueError("`num_oov_buckets` must be greater than or equal to 0, " 165 f"got {num_oov_buckets}.") 166 if (not isinstance(vocab_size, ops.Tensor) and vocab_size is not None and 167 vocab_size < 1): 168 raise ValueError(f"`vocab_size` must be greater than 0, got {vocab_size}.") 169 if (not key_dtype.is_integer) and (dtypes.string != key_dtype.base_dtype): 170 raise TypeError("`key_dtype` must be either an integer or string type, " 171 f"but got {key_dtype}") 172 if vocab_size is not None: 173 if isinstance(vocab_size, ops.Tensor): 174 vocab_size = math_ops.cast(vocab_size, dtypes.int64) 175 dataset = dataset.take(vocab_size) 176 dataset = dataset.apply(assert_cardinality(vocab_size)) 177 with ops.name_scope(name, "string_to_index"): 178 initializer = DatasetInitializer(dataset) 179 with ops.name_scope(None, "hash_table"): 180 table = lookup_ops.StaticHashTableV1(initializer, default_value) 181 if num_oov_buckets: 182 table = lookup_ops.IdTableWithHashBuckets( 183 table, 184 num_oov_buckets=num_oov_buckets, 185 hasher_spec=hasher_spec, 186 key_dtype=key_dtype) 187 return table 188 189 190@tf_export("data.experimental.index_table_from_dataset") 191def index_table_from_dataset(dataset=None, 192 num_oov_buckets=0, 193 vocab_size=None, 194 default_value=-1, 195 hasher_spec=lookup_ops.FastHashSpec, 196 key_dtype=dtypes.string, 197 name=None): 198 """Returns an index lookup table based on the given dataset. 199 200 This operation constructs a lookup table based on the given dataset of keys. 201 202 Any lookup of an out-of-vocabulary token will return a bucket ID based on its 203 hash if `num_oov_buckets` is greater than zero. Otherwise it is assigned the 204 `default_value`. 205 The bucket ID range is 206 `[vocabulary size, vocabulary size + num_oov_buckets - 1]`. 207 208 Sample Usages: 209 210 >>> ds = tf.data.Dataset.range(100).map(lambda x: tf.strings.as_string(x * 2)) 211 >>> table = tf.data.experimental.index_table_from_dataset( 212 ... ds, key_dtype=dtypes.int64) 213 >>> table.lookup(tf.constant(['0', '2', '4'], dtype=tf.string)).numpy() 214 array([0, 1, 2]) 215 216 Args: 217 dataset: A dataset of keys. 218 num_oov_buckets: The number of out-of-vocabulary buckets. 219 vocab_size: Number of the elements in the vocabulary, if known. 220 default_value: The value to use for out-of-vocabulary feature values. 221 Defaults to -1. 222 hasher_spec: A `HasherSpec` to specify the hash function to use for 223 assignation of out-of-vocabulary buckets. 224 key_dtype: The `key` data type. 225 name: A name for this op (optional). 226 227 Returns: 228 The lookup table based on the given dataset. 229 230 Raises: 231 ValueError: If 232 * `num_oov_buckets` is negative 233 * `vocab_size` is not greater than zero 234 * The `key_dtype` is not integer or string 235 """ 236 return table_from_dataset(dataset.enumerate().map(lambda v, k: (k, v)), 237 num_oov_buckets, vocab_size, default_value, 238 hasher_spec, key_dtype, name) 239