Add HuggingfaceDatasetSplitReader for using Huggingface datasets

Added a new reader to allow for reading huggingface datasets as instance
Mapped limited `datasets.features` to `allenlp.data.fields`

Verified for selective dataset and/or dataset configurations

New Dependency - "datasets==1.5.0"

Signed-off-by: Abhishek P (VMware) <pab@vmware.com>

CHANGELOG.md

@@ -8,7 +8,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## Unreleased
 
 ### Added
-
+- Add `HuggingfaceDatasetSplitReader` for using huggingface datasets in AllenNLP with limited support
 - Ported the following Huggingface `LambdaLR`-based schedulers: `ConstantLearningRateScheduler`, `ConstantWithWarmupLearningRateScheduler`, `CosineWithWarmupLearningRateScheduler`, `CosineHardRestartsWithWarmupLearningRateScheduler`.
 
 ### Changed
@@ -264,7 +264,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Added sampler class and parameter in beam search for non-deterministic search, with several
   implementations, including `MultinomialSampler`, `TopKSampler`, `TopPSampler`, and
   `GumbelSampler`. Utilizing `GumbelSampler` will give [Stochastic Beam Search](https://api.semanticscholar.org/CorpusID:76662039).
-
+    
 ### Changed
 
 - Pass batch metrics to `BatchCallback`.

allennlp/data/dataset_readers/hugging_face_datasets_reader.py

@@ -0,0 +1,170 @@
+from typing import Iterable, Optional
+
+from allennlp.data import DatasetReader, Token
+from allennlp.data.fields import TextField, LabelField, ListField
+from allennlp.data.instance import Instance
+from datasets import load_dataset
+from datasets.features import ClassLabel, Sequence, Translation, TranslationVariableLanguages
+from datasets.features import Value
+
+
+class HuggingfaceDatasetSplitReader(DatasetReader):
+    """
+        This reader implementation wraps the huggingface datasets package to utilize it's dataset management functionality
+        and load the information in AllenNLP friendly formats
+        Note: Reader works w.r.t to only one split of the dataset, i.e. you would need to create seperate reader for seperate splits
+
+        Following dataset and configurations have been verified and work with this reader
+
+                Dataset                       Dataset Configuration
+                `xnli`                        `ar`
+                `xnli`                        `en`
+                `xnli`                        `de`
+                `xnli`                        `all_languages`
+                `glue`                        `cola`
+                `glue`                        `mrpc`
+                `glue`                        `sst2`
+                `glue`                        `qqp`
+                `glue`                        `mnli`
+                `glue`                        `mnli_matched`
+                `universal_dependencies`      `en_lines`
+                `universal_dependencies`      `ko_kaist`
+                `universal_dependencies`      `af_afribooms`
+                `afrikaans_ner_corpus`        `NA`
+                `swahili`                     `NA`
+                `conll2003`                   `NA`
+                `dbpedia_14`                  `NA`
+                `trec`                        `NA`
+                `emotion`                     `NA`
+        """
+
+    def __init__(
+            self,
+            max_instances: Optional[int] = None,
+            manual_distributed_sharding: bool = False,
+            manual_multiprocess_sharding: bool = False,
+            serialization_dir: Optional[str] = None,
+            dataset_name: [str] = None,
+            split: str = 'train',
+            config_name: Optional[str] = None,
+    ) -> None:
+        super().__init__(max_instances, manual_distributed_sharding, manual_multiprocess_sharding, serialization_dir)
+
+        # It would be cleaner to create a separate reader object for different dataset
+        self.dataset = None
+        self.dataset_name = dataset_name
+        self.config_name = config_name
+        self.index = -1
+
+        if config_name:
+            self.dataset = load_dataset(self.dataset_name, self.config_name, split=split)
+        else:
+            self.dataset = load_dataset(self.dataset_name, split=split)
+
+    def _read(self, file_path) -> Iterable[Instance]:
+        """
+        Reads the dataset and converts the entry to AllenNLP friendly instance
+        """
+        for entry in self.dataset:
+            yield self.text_to_instance(entry)
+
+    def text_to_instance(self, *inputs) -> Instance:
+        """
+        Takes care of converting dataset entry into AllenNLP friendly instance
+        Currently it is implemented in an unseemly catch-up model where it converts datasets.features that are required
+        for the supported dataset, ideally it would require design where we cleanly map dataset.feature to an AllenNLP model
+        and then go ahead with converting it one by one
+        Doing that would provide the best chance of providing largest possible coverage with datasets
+
+        Currently this is how datasets.features types are mapped to AllenNLP Fields
+
+        dataset.feature type        allennlp.data.fields
+        `ClassLabel`                  `LabelField` in feature name namespace
+        `Value.string`                `TextField` with value as Token
+        `Value.*`                     `LabelField` with value being label in feature name namespace
+        `Sequence.string`             `ListField` of `TextField` with individual string as token
+        `Sequence.ClassLabel`         `ListField` of `ClassLabel` in feature name namespace
+        `Translation`                 `ListField` of 2 ListField (ClassLabel and TextField)
+        `TranslationVariableLanguages`                 `ListField` of 2 ListField (ClassLabel and TextField)
+        """
+
+        # features indicate the different information available in each entry from dataset
+        # feature types decide what type of information they are
+        # e.g. In a Sentiment an entry could have one feature indicating the text and another indica
+        features = self.dataset.features
+        fields = dict()
+
+        # TODO we need to support all different datasets features of https://huggingface.co/docs/datasets/features.html
+        for feature in features:
+            value = features[feature]
+
+            # datasets ClassLabel maps to LabelField
+            if isinstance(value, ClassLabel):
+                field = LabelField(inputs[0][feature], label_namespace=feature, skip_indexing=True)
+
+            # datasets Value can be of different types
+            elif isinstance(value, Value):
+
+                # String value maps to TextField
+                if value.dtype == 'string':
+                    # Since TextField has to be made of Tokens add whole text as a token
+                    # TODO Should we use simple heuristics to identify what is token and what is not?
+                    field = TextField([Token(inputs[0][feature])])
+
+                else:
+                    field = LabelField(inputs[0][feature], label_namespace=feature, skip_indexing=True)
+
+
+            elif isinstance(value, Sequence):
+                # datasets Sequence of strings to ListField of TextField
+                if value.feature.dtype == 'string':
+                    field_list = list()
+                    for item in inputs[0][feature]:
+                        item_field = TextField([Token(item)])
+                        field_list.append(item_field)
+                    if len(field_list) == 0:
+                        continue
+                    field = ListField(field_list)
+
+                # datasets Sequence of strings to ListField of LabelField
+                elif isinstance(value.feature, ClassLabel):
+                    field_list = list()
+                    for item in inputs[0][feature]:
+                        item_field = LabelField(label=item, label_namespace=feature, skip_indexing=True)
+                        field_list.append(item_field)
+                    if len(field_list) == 0:
+                        continue
+                    field = ListField(field_list)
+
+            # datasets Translation cannot be mapped directly but it's dict structure can be mapped to a ListField of 2 ListField
+            elif isinstance(value, Translation):
+                if value.dtype == "dict":
+                    input_dict = inputs[0][feature]
+                    langs = list(input_dict.keys())
+                    field_langs = [LabelField(lang, label_namespace="languages") for lang in langs]
+                    langs_field = ListField(field_langs)
+                    texts = list()
+                    for lang in langs:
+                        texts.append(TextField([Token(input_dict[lang])]))
+                    field = ListField([langs_field, ListField(texts)])
+
+            # TranslationVariableLanguages is functionally a pair of Lists and hence mapped to a ListField of 2 ListField
+            elif isinstance(value, TranslationVariableLanguages):
+                # Although it is indicated as dict made up of a pair of lists
+                if value.dtype == "dict":
+                    input_dict = inputs[0][feature]
+                    langs = input_dict["language"]
+                    field_langs = [LabelField(lang, label_namespace="languages") for lang in langs]
+                    langs_field = ListField(field_langs)
+                    texts = list()
+                    for lang in langs:
+                        index = langs.index(lang)
+                        texts.append(TextField([Token(input_dict["translation"][index])]))
+                    field = ListField([langs_field, ListField(texts)])
+
+            else:
+                raise ValueError(f"Datasets feature type {type(value)} is not supported yet.")
+
+            fields[feature] = field
+
+        return Instance(fields)

setup.py

@@ -73,6 +73,7 @@
         "lmdb",
         "more-itertools",
         "wandb>=0.10.0,<0.11.0",
+        "datasets==1.5.0"
     ],
     entry_points={"console_scripts": ["allennlp=allennlp.__main__:run"]},
     include_package_data=True,

tests/data/dataset_readers/huggingface_datasets_test.py

@@ -0,0 +1,45 @@
+import pytest
+
+from allennlp.data.dataset_readers.conll2003 import Conll2003DatasetReader
+from allennlp.data.dataset_readers.hugging_face_datasets_reader import HuggingfaceDatasetSplitReader
+from allennlp.common.checks import ConfigurationError
+from allennlp.common.util import ensure_list
+from allennlp.common.testing import AllenNlpTestCase
+import logging
+
+logger = logging.getLogger(__name__)
+
+# TODO these UTs are actually downloading the datasets and will be very very slow
+class HuggingfaceDatasetSplitReaderTest:
+
+
+    SUPPORTED_DATASETS_WITHOUT_CONFIG = ["afrikaans_ner_corpus", "dbpedia_14", "trec", "swahili", "conll2003", "emotion"]
+
+    """
+        Running the tests for supported datasets which do not require config name to be specified
+    """
+    @pytest.mark.parametrize("dataset", SUPPORTED_DATASETS_WITHOUT_CONFIG)
+    def test_read_for_datasets_without_config(self, dataset):
+        huggingface_reader = HuggingfaceDatasetSplitReader(dataset_name=dataset)
+        instances = list(huggingface_reader.read(None))
+        assert len(instances) == len(huggingface_reader.dataset)
+
+    # Not testing for all configurations only some
+    SUPPORTED_DATASET_CONFIGURATION = (
+        ("glue", "cola"),
+        ("universal_dependencies", "af_afribooms"),
+        ("xnli", "all_languages")
+    )
+
+    """
+        Running the tests for supported datasets which require config name to be specified
+    """
+    @pytest.mark.parametrize("dataset, config", SUPPORTED_DATASET_CONFIGURATION)
+    def test_read_for_datasets_requiring_config(self, dataset, config):
+        huggingface_reader = HuggingfaceDatasetSplitReader(dataset_name=dataset, config_name=config)
+        instances = list(huggingface_reader.read(None))
+        assert len(instances) == len(huggingface_reader.dataset)
+
+
+
+