Add HuggingFace datasets support (#3570)

* add huggingface to tasks list

* huggingface task for glue dataset support

* handle splits

* readme

* string labels instead of numerical labels

* identifying text, labels, cands for setup data

* update test:

* add huggingface datasets module to requirements

* abstract class

* abstract class and add args to specify what features for query and label in parlai

* add glue teacher

* requirements tqdm edit

* formatting

* update docutils for cleaninstall

* update docutils for cleaninstall

* abstracthuggingfaceteacher, instead of taking in info through args, move this to the development side

* add hugging face splits mapping property

* remove unused _path and moved load_dataset to setup_data

* add render_text_field

* add properties to abstract class and methods

* create glue task

* update the hf agent and tasklist

parlai/tasks/glue/README.md

@@ -0,0 +1,5 @@
+Task: Glue
+===============
+Description: GLUE, the General Language Understanding Evaluation benchmark is a collection of resources for training, evaluating, and analyzing natural language understanding systems. This task uses the `AbstractHuggingFaceTeacher` to load the dataset.
+
+Websites: https://huggingface.co/ and https://gluebenchmark.com/

parlai/tasks/glue/__init__.py

@@ -0,0 +1,5 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.

parlai/tasks/glue/agents.py

@@ -0,0 +1,19 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from parlai.tasks.huggingface.agents import AbstractHuggingFaceTeacher
+
+
+class ColaTeacher(AbstractHuggingFaceTeacher):
+    hf_path = 'glue'
+    hf_name = 'cola'
+    hf_text_fields = ['sentence']
+    hf_label_field = 'label'
+    hf_splits_mapping = {'train': 'train', 'valid': 'validation', 'test': 'test'}
+
+
+class DefaultTeacher(ColaTeacher):
+    pass

parlai/tasks/huggingface/README.md

@@ -0,0 +1,6 @@
+Task: HuggingFace
+===============
+Description: Can load HuggingFace datasets. 
+
+Website: https://huggingface.co/
+

parlai/tasks/huggingface/__init__.py

@@ -0,0 +1,5 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.

parlai/tasks/huggingface/agents.py

@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+
+# Copyright (c) Facebook, Inc. and its affiliates.
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from parlai.core.teachers import DialogTeacher
+from parlai.utils.data import DatatypeHelper
+from typing import Dict, Iterable, List, Optional, Tuple
+from typing_extensions import TypedDict
+import os
+
+# huggingface imports
+from datasets import load_dataset
+
+
+class SplitsMappingDict(TypedDict):
+    train: str
+    valid: str
+    test: str
+
+
+class AbstractHuggingFaceTeacher(DialogTeacher):
+    """
+    Abstract parent class for HuggingFace teachers. Extend this class and specify the
+    attributes below to use a different dataset.
+
+    hf_path = path parameter passed into hugging face load_dataset function
+    hf_name = name parameter passed into hugging face load_dataset function
+    hf_text_fields = list of names of the data fields from the dataset to be included in the text/query
+    hf_label_field = name of the data field from the hf dataset that specifies the label of the episode
+    hf_splits_mapping = dictionary mapping with the keys 'train', 'valid', and 'test', that map to the
+    names of the splits of the hf dataset.
+    render_text_field = bool where if True, will include the text field name in the query (e.g. "sentence: <sentence>")
+    """
+
+    def __init__(self, opt, shared=None):
+        self.fold = DatatypeHelper.fold(opt['datatype'])
+        self.hf_split = self.hf_splits_mapping[self.fold]
+        self.data_path = self._path(opt)
+        opt['datafile'] = self.data_path
+
+        self.id = "huggingface"
+        super().__init__(opt, shared)
+
+    def _path(self, opt):
+        if self.hf_name:
+            return os.path.join(
+                opt['datapath'], 'huggingface', self.hf_path, self.hf_name, self.fold
+            )
+        return os.path.join(opt['datapath'], 'huggingface', self.hf_path, self.fold)
+
+    @property
+    def hf_path(self) -> str:
+        raise NotImplementedError
+
+    @property
+    def hf_name(self) -> Optional[str]:
+        return None
+
+    @property
+    def hf_text_fields(self) -> List[str]:
+        raise NotImplementedError
+
+    @property
+    def hf_label_field(self) -> str:
+        raise NotImplementedError
+
+    @property
+    def hf_splits_mapping(self) -> SplitsMappingDict:
+        raise NotImplementedError
+
+    def _get_text_value(self, row) -> Tuple[str, Dict[str, str]]:
+        """
+        return the constructed text query and dict mapping text field names to values.
+        """
+        # construct text query from the hf_text_fields specified
+        text_dict = {}
+        for col in self.hf_text_fields:
+            text_part = row.get(col)
+            if text_part is None:
+                raise KeyError(f'Feature "{col}" not found in data.')
+            text_dict[col] = text_part
+        return '\n'.join(text_dict.values()), text_dict
+
+    def _get_label_value(self, row):
+        """
+        return the label value from the data row.
+        """
+        return row[self.hf_label_field]
+
+    def _get_label_candidates(self, row, label) -> str:
+        """
+        try to return the true label text value from the row and the candidates.
+        """
+        pre_candidates = self.dataset.features[self.hf_label_field].names
+        # construct label and candidates
+        if type(label) is int:
+            return pre_candidates[label], pre_candidates
+        if label in row:
+            return row[label], [row[l] for l in pre_candidates]
+        return label, pre_candidates
+
+    def setup_data(self, path: str) -> Iterable[tuple]:
+        """
+        Default implementation of setup_data.
+
+        Manually override if needed.
+        """
+        # load dataset from HuggingFace
+        self.dataset = load_dataset(
+            path=self.hf_path, name=self.hf_name, split=self.hf_split
+        )
+
+        for row in self.dataset:
+            query, text_dict = self._get_text_value(row)
+            label = self._get_label_value(row)
+            label, candidates = self._get_label_candidates(row, label)
+
+            episode_dict = text_dict
+            episode_dict['text'] = query
+            episode_dict['label'] = label
+            episode_dict['label_candidates'] = candidates
+            yield episode_dict, True
+
+
+class DefaultTeacher:
+    def __init__(self, opt):
+        raise NotImplementedError(
+            "There is no default teacher for HuggingFace datasets. Please use a specific one."
+        )

parlai/tasks/task_list.py

@@ -288,6 +288,19 @@
         ),
         "links": {"arXiv": "https://arxiv.org/abs/1706.05125"},
     },
+    {
+        "id": "Glue",
+        "display_name": "Glue",
+        "task": "glue",
+        "tags": [],
+        "description": (
+            "GLUE, the General Language Understanding Evaluation benchmark is a collection of resources for training, evaluating, and analyzing natural language understanding systems."
+        ),
+        "links": {
+            "website": "https://gluebenchmark.com/",
+            "website2": "https://huggingface.co/datasets/glue",
+        },
+    },
     {
         "id": "HotpotQA",
         "display_name": "HotpotQA",
@@ -303,6 +316,14 @@
         ),
         "links": {"arXiv": "https://arxiv.org/abs/1809.09600"},
     },
+    {
+        "id": "HuggingFace",
+        "display_name": "HuggingFace",
+        "task": "huggingface",
+        "tags": [],
+        "description": ("HuggingFace datasets"),
+        "links": {"website": "https://huggingface.co/"},
+    },
     {
         "id": "LIGHT-Dialogue",
         "display_name": "LIGHT-Dialogue",

requirements.txt

@@ -1,6 +1,7 @@
 boto3==1.9.246
 botocore==1.12.246
 coloredlogs==14.0
+datasets==1.4.1
 docutils<0.16,>=0.14
 emoji==0.5.4
 docformatter==1.3.0
@@ -42,7 +43,7 @@ tensorboardX==2.1
 tokenizers>=0.8.0
 torchtext>=0.5.0
 tornado==6.0.4
-tqdm==4.36.1
+tqdm~=4.36.1
 typing-extensions==3.7.4.1
 Unidecode==1.1.1
 urllib3~=1.25.9