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 text dataset generation utilities.""" 16from __future__ import absolute_import 17from __future__ import division 18from __future__ import print_function 19 20import numpy as np 21 22from tensorflow.python.data.ops import dataset_ops 23from tensorflow.python.keras.preprocessing import dataset_utils 24from tensorflow.python.ops import io_ops 25from tensorflow.python.ops import string_ops 26from tensorflow.python.util.tf_export import keras_export 27 28 29@keras_export('keras.preprocessing.text_dataset_from_directory', v1=[]) 30def text_dataset_from_directory(directory, 31 labels='inferred', 32 label_mode='int', 33 class_names=None, 34 batch_size=32, 35 max_length=None, 36 shuffle=True, 37 seed=None, 38 validation_split=None, 39 subset=None, 40 follow_links=False): 41 """Generates a `tf.data.Dataset` from text files in a directory. 42 43 If your directory structure is: 44 45 ``` 46 main_directory/ 47 ...class_a/ 48 ......a_text_1.txt 49 ......a_text_2.txt 50 ...class_b/ 51 ......b_text_1.txt 52 ......b_text_2.txt 53 ``` 54 55 Then calling `text_dataset_from_directory(main_directory, labels='inferred')` 56 will return a `tf.data.Dataset` that yields batches of texts from 57 the subdirectories `class_a` and `class_b`, together with labels 58 0 and 1 (0 corresponding to `class_a` and 1 corresponding to `class_b`). 59 60 Only `.txt` files are supported at this time. 61 62 Args: 63 directory: Directory where the data is located. 64 If `labels` is "inferred", it should contain 65 subdirectories, each containing text files for a class. 66 Otherwise, the directory structure is ignored. 67 labels: Either "inferred" 68 (labels are generated from the directory structure), 69 None (no labels), 70 or a list/tuple of integer labels of the same size as the number of 71 text files found in the directory. Labels should be sorted according 72 to the alphanumeric order of the text file paths 73 (obtained via `os.walk(directory)` in Python). 74 label_mode: 75 - 'int': means that the labels are encoded as integers 76 (e.g. for `sparse_categorical_crossentropy` loss). 77 - 'categorical' means that the labels are 78 encoded as a categorical vector 79 (e.g. for `categorical_crossentropy` loss). 80 - 'binary' means that the labels (there can be only 2) 81 are encoded as `float32` scalars with values 0 or 1 82 (e.g. for `binary_crossentropy`). 83 - None (no labels). 84 class_names: Only valid if "labels" is "inferred". This is the explict 85 list of class names (must match names of subdirectories). Used 86 to control the order of the classes 87 (otherwise alphanumerical order is used). 88 batch_size: Size of the batches of data. Default: 32. 89 max_length: Maximum size of a text string. Texts longer than this will 90 be truncated to `max_length`. 91 shuffle: Whether to shuffle the data. Default: True. 92 If set to False, sorts the data in alphanumeric order. 93 seed: Optional random seed for shuffling and transformations. 94 validation_split: Optional float between 0 and 1, 95 fraction of data to reserve for validation. 96 subset: One of "training" or "validation". 97 Only used if `validation_split` is set. 98 follow_links: Whether to visits subdirectories pointed to by symlinks. 99 Defaults to False. 100 101 Returns: 102 A `tf.data.Dataset` object. 103 - If `label_mode` is None, it yields `string` tensors of shape 104 `(batch_size,)`, containing the contents of a batch of text files. 105 - Otherwise, it yields a tuple `(texts, labels)`, where `texts` 106 has shape `(batch_size,)` and `labels` follows the format described 107 below. 108 109 Rules regarding labels format: 110 - if `label_mode` is `int`, the labels are an `int32` tensor of shape 111 `(batch_size,)`. 112 - if `label_mode` is `binary`, the labels are a `float32` tensor of 113 1s and 0s of shape `(batch_size, 1)`. 114 - if `label_mode` is `categorial`, the labels are a `float32` tensor 115 of shape `(batch_size, num_classes)`, representing a one-hot 116 encoding of the class index. 117 """ 118 if labels not in ('inferred', None): 119 if not isinstance(labels, (list, tuple)): 120 raise ValueError( 121 '`labels` argument should be a list/tuple of integer labels, of ' 122 'the same size as the number of text files in the target ' 123 'directory. If you wish to infer the labels from the subdirectory ' 124 'names in the target directory, pass `labels="inferred"`. ' 125 'If you wish to get a dataset that only contains text samples ' 126 '(no labels), pass `labels=None`.') 127 if class_names: 128 raise ValueError('You can only pass `class_names` if the labels are ' 129 'inferred from the subdirectory names in the target ' 130 'directory (`labels="inferred"`).') 131 if label_mode not in {'int', 'categorical', 'binary', None}: 132 raise ValueError( 133 '`label_mode` argument must be one of "int", "categorical", "binary", ' 134 'or None. Received: %s' % (label_mode,)) 135 if labels is None or label_mode is None: 136 labels = None 137 label_mode = None 138 dataset_utils.check_validation_split_arg( 139 validation_split, subset, shuffle, seed) 140 141 if seed is None: 142 seed = np.random.randint(1e6) 143 file_paths, labels, class_names = dataset_utils.index_directory( 144 directory, 145 labels, 146 formats=('.txt',), 147 class_names=class_names, 148 shuffle=shuffle, 149 seed=seed, 150 follow_links=follow_links) 151 152 if label_mode == 'binary' and len(class_names) != 2: 153 raise ValueError( 154 'When passing `label_mode="binary", there must exactly 2 classes. ' 155 'Found the following classes: %s' % (class_names,)) 156 157 file_paths, labels = dataset_utils.get_training_or_validation_split( 158 file_paths, labels, validation_split, subset) 159 if not file_paths: 160 raise ValueError('No text files found.') 161 162 dataset = paths_and_labels_to_dataset( 163 file_paths=file_paths, 164 labels=labels, 165 label_mode=label_mode, 166 num_classes=len(class_names), 167 max_length=max_length) 168 if shuffle: 169 # Shuffle locally at each iteration 170 dataset = dataset.shuffle(buffer_size=batch_size * 8, seed=seed) 171 dataset = dataset.batch(batch_size) 172 # Users may need to reference `class_names`. 173 dataset.class_names = class_names 174 return dataset 175 176 177def paths_and_labels_to_dataset(file_paths, 178 labels, 179 label_mode, 180 num_classes, 181 max_length): 182 """Constructs a dataset of text strings and labels.""" 183 path_ds = dataset_ops.Dataset.from_tensor_slices(file_paths) 184 string_ds = path_ds.map( 185 lambda x: path_to_string_content(x, max_length)) 186 if label_mode: 187 label_ds = dataset_utils.labels_to_dataset(labels, label_mode, num_classes) 188 string_ds = dataset_ops.Dataset.zip((string_ds, label_ds)) 189 return string_ds 190 191 192def path_to_string_content(path, max_length): 193 txt = io_ops.read_file(path) 194 if max_length is not None: 195 txt = string_ops.substr(txt, 0, max_length) 196 return txt 197