|
| 1 | +import numpy as np |
| 2 | +import os |
| 3 | +import json |
| 4 | +import h5py |
| 5 | +from tqdm import tqdm |
| 6 | + |
| 7 | +from torchmeta.utils.data import Dataset, ClassDataset, CombinationMetaDataset |
| 8 | +from torchmeta.datasets.utils import get_asset |
| 9 | + |
| 10 | + |
| 11 | +class Letter(CombinationMetaDataset): |
| 12 | + """The Letter Image Recognition Dataset """ |
| 13 | + def __init__(self, root, num_classes_per_task=None, meta_train=False, meta_val=False, meta_test=False, |
| 14 | + meta_split=None, transform=None, target_transform=None, dataset_transform=None, |
| 15 | + class_augmentations=None, download=False): |
| 16 | + """ |
| 17 | + Letter Image Recognition Data [1]: |
| 18 | +
|
| 19 | + https://archive.ics.uci.edu/ml/datasets/Letter+Recognition - 01-01-1991 |
| 20 | +
|
| 21 | + The objective is to identify each of a large number of black-and-white |
| 22 | + rectangular pixel displays as one of the 26 capital letters in the English |
| 23 | + alphabet. The character images were based on 20 different fonts and each |
| 24 | + letter within these 20 fonts was randomly distorted to produce a file of |
| 25 | + 20,000 unique stimuli. Each stimulus was converted into 16 primitive |
| 26 | + numerical attributes (statistical moments and edge counts) which were then |
| 27 | + scaled to fit into a range of integer values from 0 through 15. We |
| 28 | + typically train on the first 16000 items and then use the resulting model |
| 29 | + to predict the letter category for the remaining 4000. See the article |
| 30 | + cited above for more details. |
| 31 | +
|
| 32 | + The dataset is loaded and processed with benchlib. Originally it is from open-ml. |
| 33 | + https://www.openml.org/d/6 |
| 34 | +
|
| 35 | + Parameters |
| 36 | + ---------- |
| 37 | + root : string |
| 38 | + Root directory where the dataset folder `letter_task_id_6` exists. |
| 39 | +
|
| 40 | + num_classes_per_task : int |
| 41 | + Number of classes per tasks. This corresponds to "N" in "N-way" |
| 42 | + classification. |
| 43 | +
|
| 44 | + meta_train : bool (default: `False`) |
| 45 | + Use the meta-train split of the dataset. If set to `True`, then the |
| 46 | + arguments `meta_val` and `meta_test` must be set to `False`. Exactly one |
| 47 | + of these three arguments must be set to `True`. |
| 48 | +
|
| 49 | + meta_val : bool (default: `False`) |
| 50 | + Use the meta-validation split of the dataset. If set to `True`, then the |
| 51 | + arguments `meta_train` and `meta_test` must be set to `False`. Exactly |
| 52 | + one of these three arguments must be set to `True`. |
| 53 | +
|
| 54 | + meta_test : bool (default: `False`) |
| 55 | + Use the meta-test split of the dataset. If set to `True`, then the |
| 56 | + arguments `meta_train` and `meta_val` must be set to `False`. Exactly |
| 57 | + one of these three arguments must be set to `True`. |
| 58 | +
|
| 59 | + meta_split : string in {'train', 'val', 'test'}, optional |
| 60 | + Name of the split to use. This overrides the arguments `meta_train`, |
| 61 | + `meta_val` and `meta_test` if all three are set to `False`. |
| 62 | +
|
| 63 | + transform : callable, optional |
| 64 | + A function/transform that takes a numpy array or a pytorch array |
| 65 | + (depending when the transforms is applied), and returns a transformed |
| 66 | + version. |
| 67 | +
|
| 68 | + target_transform : callable, optional |
| 69 | + A function/transform that takes a target, and returns a transformed |
| 70 | + version. |
| 71 | +
|
| 72 | + dataset_transform : callable, optional |
| 73 | + A function/transform that takes a dataset (ie. a task), and returns a |
| 74 | + transformed version of it. E.g. `torchmeta.transforms.ClassSplitter()`. |
| 75 | +
|
| 76 | + class_augmentations : list of callable, optional |
| 77 | + A list of functions that augment the dataset with new classes. These |
| 78 | + classes are transformations of existing classes. |
| 79 | +
|
| 80 | + download : bool (default: `False`) |
| 81 | + If `True`, downloads the original files and processes the dataset in the |
| 82 | + root directory (under the `letter_task_id_6` folder). If the dataset |
| 83 | + is already available, this does not download/process the dataset again. |
| 84 | +
|
| 85 | + References |
| 86 | + ----- |
| 87 | + [1] P. W. Frey and D. J. Slate. "Letter Recognition Using Holland-style |
| 88 | + Adaptive Classifiers". Machine Learning 6(2), 1991 |
| 89 | + """ |
| 90 | + dataset = LetterClassDataset(root, |
| 91 | + meta_train=meta_train, |
| 92 | + meta_val=meta_val, |
| 93 | + meta_test=meta_test, |
| 94 | + meta_split=meta_split, |
| 95 | + transform=transform, |
| 96 | + class_augmentations=class_augmentations, |
| 97 | + download=download) |
| 98 | + super(Letter, self).__init__(dataset, |
| 99 | + num_classes_per_task, |
| 100 | + target_transform=target_transform, |
| 101 | + dataset_transform=dataset_transform) |
| 102 | + |
| 103 | + |
| 104 | +class LetterClassDataset(ClassDataset): |
| 105 | + |
| 106 | + benchlib_namespace = "openml_datasets" |
| 107 | + benchlib_dataset_name = "letter_task_id_6" |
| 108 | + |
| 109 | + folder = "letter_task_id_6" |
| 110 | + filename = '{0}_data.hdf5' |
| 111 | + filename_labels = '{0}_labels.json' |
| 112 | + |
| 113 | + def __init__(self, root, meta_train=False, meta_val=False, meta_test=False, meta_split=None, transform=None, |
| 114 | + class_augmentations=None, download=False): |
| 115 | + super(LetterClassDataset, self).__init__(meta_train=meta_train, meta_val=meta_val, meta_test=meta_test, |
| 116 | + meta_split=meta_split, class_augmentations=class_augmentations) |
| 117 | + |
| 118 | + self.root = os.path.join(os.path.expanduser(root), self.folder) |
| 119 | + self.transform = transform |
| 120 | + |
| 121 | + self.split_filename = os.path.join(self.root, self.filename.format(self.meta_split)) |
| 122 | + self.split_filename_labels = os.path.join(self.root, self.filename_labels.format(self.meta_split)) |
| 123 | + |
| 124 | + self._data_file = None |
| 125 | + self._data = None |
| 126 | + self._labels = None |
| 127 | + |
| 128 | + if download: |
| 129 | + self.download() |
| 130 | + |
| 131 | + if not self._check_integrity(): |
| 132 | + raise RuntimeError('Letter integrity check failed') |
| 133 | + self._num_classes = len(self.labels) |
| 134 | + |
| 135 | + def __getitem__(self, index): |
| 136 | + label = self.labels[index % self.num_classes] |
| 137 | + data = self.data[label] |
| 138 | + transform = self.get_transform(index, self.transform) |
| 139 | + target_transform = self.get_target_transform(index) |
| 140 | + |
| 141 | + return LetterDataset(index, data, label, transform=transform, target_transform=target_transform) |
| 142 | + |
| 143 | + @property |
| 144 | + def num_classes(self): |
| 145 | + return self._num_classes |
| 146 | + |
| 147 | + @property |
| 148 | + def data(self): |
| 149 | + if self._data is None: |
| 150 | + self._data_file = h5py.File(self.split_filename, 'r') |
| 151 | + self._data = self._data_file['datasets'] |
| 152 | + return self._data |
| 153 | + |
| 154 | + @property |
| 155 | + def labels(self): |
| 156 | + if self._labels is None: |
| 157 | + with open(self.split_filename_labels, 'r') as f: |
| 158 | + self._labels = json.load(f) |
| 159 | + return self._labels |
| 160 | + |
| 161 | + def _check_integrity(self): |
| 162 | + return (os.path.isfile(self.split_filename) |
| 163 | + and os.path.isfile(self.split_filename_labels)) |
| 164 | + |
| 165 | + def close(self): |
| 166 | + if self._data is not None: |
| 167 | + self._data.close() |
| 168 | + self._data = None |
| 169 | + |
| 170 | + def download(self): |
| 171 | + |
| 172 | + if self._check_integrity(): |
| 173 | + return |
| 174 | + |
| 175 | + from benchlib.datasets.syne_datasets import get_syne_dataset |
| 176 | + from benchlib.datasets.data_detergent import DataDetergent |
| 177 | + |
| 178 | + # feature transforms are performed by the DataDetergent |
| 179 | + d = DataDetergent(get_syne_dataset(namespace=self.benchlib_namespace, |
| 180 | + dataset_name=self.benchlib_dataset_name + "/"), |
| 181 | + do_impute_nans=True, |
| 182 | + do_normalize_cols=True, |
| 183 | + do_remove_const_features=True) |
| 184 | + |
| 185 | + # stack the features and targets into one big numpy array each, since we want a new split. |
| 186 | + features = [] |
| 187 | + targets = [] |
| 188 | + for split in ['train', 'val', 'test']: |
| 189 | + if split == 'train': |
| 190 | + data = d.get_training_data() |
| 191 | + elif split == 'val': |
| 192 | + data = d.get_validation_data() |
| 193 | + elif split == 'test': |
| 194 | + data = d.get_test_data() |
| 195 | + else: |
| 196 | + raise ValueError(f"split {split} not found.") |
| 197 | + features.append(data[0]) |
| 198 | + targets.append(data[1]) |
| 199 | + data = None |
| 200 | + features = np.concatenate(features, axis=0) |
| 201 | + targets = np.concatenate(targets, axis=0) |
| 202 | + |
| 203 | + # for each meta-data-split, get the labels, then check which data-point belongs to the set (via a mask). |
| 204 | + # then, retrieve the features and targets belonging to the set. Then create hdf5 file for these features. |
| 205 | + for s, split in enumerate(['train', 'val', 'test']): |
| 206 | + label_set = get_asset(self.folder, '{0}.json'.format(split)) |
| 207 | + label_set_integers = [int(l) for l in label_set] |
| 208 | + |
| 209 | + is_in_set = [t in label_set_integers for t in targets] |
| 210 | + features_set = features[is_in_set, :] |
| 211 | + targets_set = targets[is_in_set] |
| 212 | + assert targets_set.shape[0] == features_set.shape[0] |
| 213 | + |
| 214 | + unique_targets_set = np.sort(np.unique(targets_set)) |
| 215 | + if len(label_set_integers) > unique_targets_set.shape[0]: |
| 216 | + print(f"unique set of labels is smaller ({len(unique_targets_set.shape[0])}) than set of labels " |
| 217 | + f"given by assets ({len(label_set_integers)}). Proceeding with unique set of labels.") |
| 218 | + |
| 219 | + # write unique targets with enough data to json file. this is not necessarily the same as the tag set |
| 220 | + len_str = int(np.ceil(np.log10(unique_targets_set.shape[0] + 1))) |
| 221 | + unique_targets_str = [str(i).zfill(len_str) for i in unique_targets_set] |
| 222 | + |
| 223 | + labels_filename = os.path.join(self.root, self.filename_labels.format(split)) |
| 224 | + with open(labels_filename, 'w') as f: |
| 225 | + json.dump(unique_targets_str, f) |
| 226 | + |
| 227 | + # write data (features and class labels) |
| 228 | + filename = os.path.join(self.root, self.filename.format(split)) |
| 229 | + with h5py.File(filename, 'w') as f: |
| 230 | + group = f.create_group('datasets') |
| 231 | + dtype = h5py.special_dtype(vlen=np.float64) |
| 232 | + |
| 233 | + for i, label in enumerate(tqdm(unique_targets_str, desc=filename)): |
| 234 | + data_class = features_set[targets_set == int(label), :] |
| 235 | + group.create_dataset(label, data=data_class) # , dtype=dtype) |
| 236 | + |
| 237 | + |
| 238 | +class LetterDataset(Dataset): |
| 239 | + def __init__(self, index, data, label, transform=None, target_transform=None): |
| 240 | + super(LetterDataset, self).__init__(index, transform=transform, target_transform=target_transform) |
| 241 | + self.data = data |
| 242 | + self.label = label |
| 243 | + |
| 244 | + def __len__(self): |
| 245 | + return len(self.data) |
| 246 | + |
| 247 | + def __getitem__(self, index): |
| 248 | + features = self.data[index, :] |
| 249 | + target = self.label |
| 250 | + |
| 251 | + if self.transform is not None: |
| 252 | + features = self.transform(features) |
| 253 | + |
| 254 | + if self.target_transform is not None: |
| 255 | + target = self.target_transform(target) |
| 256 | + |
| 257 | + return features, target |
| 258 | + |
| 259 | + |
| 260 | +def create_asset(root='data', number_of_classes_per_split=None, numpy_seed=42): |
| 261 | + """This methods creates the assets of the letter dataset. These are the meta-dataset splits from the |
| 262 | + original data. Only run this method in case you want to create new assets. Once created, copy the assets to |
| 263 | + this directory: torchmeta.datasets.assets.letter_task_id_6. You can also manually change the assets.""" |
| 264 | + |
| 265 | + # split fractions: train, valid, tes |
| 266 | + if number_of_classes_per_split is None: |
| 267 | + number_of_classes_per_split = {"train": 15, |
| 268 | + "val": 5, |
| 269 | + "test": 6} |
| 270 | + num_classes = 0 |
| 271 | + for key in number_of_classes_per_split: |
| 272 | + num_classes += number_of_classes_per_split[key] |
| 273 | + assert num_classes == 26 |
| 274 | + |
| 275 | + def make_split(num_classes, number_of_classes_per_split): |
| 276 | + """get permutation of labels and split according to number of classes per split""" |
| 277 | + np.random.seed(numpy_seed) |
| 278 | + |
| 279 | + perm = np.random.permutation(num_classes) |
| 280 | + class_splits = {} |
| 281 | + start = 0 |
| 282 | + for split in ["train", "val", "test"]: |
| 283 | + num_c = number_of_classes_per_split[split] |
| 284 | + |
| 285 | + class_splits[split] = [str(i) for i in perm[start:start+num_c]] |
| 286 | + start += num_c |
| 287 | + return class_splits |
| 288 | + |
| 289 | + # Split the classes according to the number of classes per split, and store the splits in the data directory. |
| 290 | + class_splits = make_split(num_classes, number_of_classes_per_split) |
| 291 | + print(class_splits) |
| 292 | + root_path = os.path.join(os.path.expanduser(root), LetterClassDataset.folder) |
| 293 | + for split in ["train", "val", "test"]: |
| 294 | + asset_filename = os.path.join(root_path, "{0}.json".format(split)) |
| 295 | + with open(asset_filename, 'w') as f: |
| 296 | + json.dump(class_splits[split], f) |
0 commit comments