Datasets:

Tasks:
Other
Languages:
English
Multilinguality:
monolingual
Size Categories:
100M<n<1B
ArXiv:
Tags:
License:
VictorSanh HF staff commited on
Commit
e3e6197
1 Parent(s): 11892f2

converging

Browse files
Files changed (2) hide show
  1. P3.py +59 -97
  2. print_data_split_sizes.py +30 -0
P3.py CHANGED
@@ -16,9 +16,8 @@
16
 
17
 
18
  import datasets
19
- import glob
20
  import json
21
- import os
22
  from collections import defaultdict
23
  import tensorflow as tf
24
 
@@ -27,7 +26,7 @@ _CITATION = """\
27
  TODO"""
28
 
29
  _DESCRIPTION = """\
30
- P3 (Pubic Pool of Prompts)is a collection of prompted English datasets covering a diverse set of NLP tasks. A prompt is the combination of an input template and a target template. The templates are functions mapping a data example into natural language for the input and target sequences. For example, in the case of an NLI dataset, the data example would include fields for *Premise, Hypothesis, Label*. An input template would be *If {Premise} is true, is it also true that {Hypothesis}?*, whereas a target template can be defined with the label choices *Choices[label]*. Here *Choices* is prompt-specific metadata that consists of the options *yes, maybe, no* corresponding to *label* being entailment (0), neutral (1) or contradiction (2).
31
 
32
  Prompts are collected using [Promptsource](https://github.com/bigscience-workshop/promptsource), an interface to interactively write prompts on datasets, and collect prompt-specific metadata such as evaluation metrics. As of October 13th, there are 2'000 prompts collected for 270+ data(sub)sets. The collection of prompts is publicly available on [Promptsource](https://github.com/bigscience-workshop/promptsource).
33
 
@@ -38,53 +37,11 @@ _LICENSE = "Apache License 2.0"
38
 
39
  _HOMEPAGE = "https://github.com/bigscience-workshop/promptsource"
40
 
41
- _DATA_PATH = "data"
42
-
43
-
44
- # def load_cached_task(cache_dir, split):
45
- # # TODO(Victor): this info.*.json is actually done twice... -> factorize
46
- # with tf.io.gfile.GFile(os.path.join(cache_dir, f"info.{split}.json")) as f:
47
- # split_info = json.load(f)
48
- # features = split_info["features"]
49
-
50
- # # Use `FixedLenSequenceFeature` for sequences with variable length.
51
- # def _feature_config(shape, dtype):
52
- # if dtype in ("int32", "bool"):
53
- # # int32 and bool are stored as int64 in the tf.train.Example protobuf.
54
- # dtype = "int64"
55
- # if shape and shape[0] is None:
56
- # return tf.io.FixedLenSequenceFeature(
57
- # shape[1:], dtype, allow_missing=True
58
- # )
59
- # return tf.io.FixedLenFeature(shape, dtype)
60
-
61
- # feature_description = {
62
- # feat: _feature_config(**desc) for feat, desc in features.items()
63
- # }
64
-
65
- # tfrecords = os.path.join(
66
- # cache_dir, f"{split}.tfrecord-*-of-*{split_info['num_shards']}"
67
- # )
68
- # ds = tf.data.TFRecordDataset(tf.io.gfile.glob(tfrecords))
69
- # ds = ds.map(
70
- # lambda pb: tf.io.parse_single_example(pb, feature_description),
71
- # num_parallel_calls=tf.data.experimental.AUTOTUNE
72
- # )
73
- # # Cast features back to the types from the info JSON since some features
74
- # # must be cast for storage (e.g., in32 is stored as int64).
75
- # ds = ds.map(
76
- # lambda x: {k: tf.cast(v, features[k]["dtype"]) for k, v in x.items()},
77
- # num_parallel_calls=tf.data.experimental.AUTOTUNE
78
- # )
79
- # return ds
80
-
81
- def load_cached_task(features_file, tfrecord):
82
- # # TODO(Victor): this info.*.json is actually done twice... -> factorize
83
- # with tf.io.gfile.GFile(os.path.join(cache_dir, f"info.{split}.json")) as f:
84
- with tf.io.gfile.GFile(features_file) as f:
85
- split_info = json.load(f)
86
- features = split_info["features"]
87
 
 
 
88
  # Use `FixedLenSequenceFeature` for sequences with variable length.
89
  def _feature_config(shape, dtype):
90
  if dtype in ("int32", "bool"):
@@ -97,81 +54,88 @@ def load_cached_task(features_file, tfrecord):
97
  return tf.io.FixedLenFeature(shape, dtype)
98
 
99
  feature_description = {
100
- feat: _feature_config(**desc) for feat, desc in features.items()
101
  }
102
 
103
- ds = tf.data.TFRecordDataset(tf.io.gfile.glob([tfrecord]))
104
  ds = ds.map(
105
  lambda pb: tf.io.parse_single_example(pb, feature_description),
106
  num_parallel_calls=tf.data.experimental.AUTOTUNE
107
  )
108
  # Cast features back to the types from the info JSON since some features
109
- # must be cast for storage (e.g., in32 is stored as int64).
110
  ds = ds.map(
111
- lambda x: {k: tf.cast(v, features[k]["dtype"]) for k, v in x.items()},
112
  num_parallel_calls=tf.data.experimental.AUTOTUNE
113
  )
114
  return ds
115
 
116
 
117
- def find_task_splits_and_features():
118
- """Find the available tasks under ./data and their available splits and features."""
119
- task_and_their_splits = defaultdict(dict)
120
- for stats in glob.glob(f"{_DATA_PATH}/*/stats.*.json"):
121
- folder_path = os.path.dirname(stats)
122
- task_name = folder_path.split("/")[-1]
123
- split_name = os.path.basename(stats).split(".")[1]
 
124
 
125
- if not os.path.exists(f"{folder_path}/COMPLETED"):
126
- continue
127
 
128
- with open(stats, "r") as f:
129
- split_stats = json.load(f)
130
- nb_examples = split_stats["examples"]
 
 
 
 
 
 
 
 
 
 
131
 
132
- if nb_examples > 0:
133
- with open(os.path.join(folder_path, f"info.{split_name}.json")) as f:
134
- split_info = json.load(f)
135
- features = split_info["features"]
136
- assert split_info["num_shards"] == 1
 
 
 
137
 
138
- # All splits under the same task have the same features dictionary (and thus the same features list)
139
- if task_and_their_splits[task_name] == {}:
140
- task_and_their_splits[task_name] = {
141
  "splits": [],
142
- "features": [],
143
  }
 
 
144
 
145
- task_and_their_splits[task_name]["splits"].append(split_name)
146
- if task_and_their_splits[task_name]["features"] == []:
147
- task_and_their_splits[task_name]["features"] = sorted(list(features.keys()))
148
- else:
149
- assert task_and_their_splits[task_name]["features"] == sorted(list(features.keys()))
150
- return task_and_their_splits
151
 
152
 
153
- _TASK_SPLITS_AND_FEATURES = find_task_splits_and_features()
154
  _URLs = {
155
  task_name: {
156
  split_name: {
157
  "tfrecord": f"{_DATA_PATH}/{task_name}/{split_name}.tfrecord-00000-of-00001",
158
- "features_file": f"{_DATA_PATH}/{task_name}/info.{split_name}.json",
159
  }
160
- for split_name in splits_and_features["splits"]
161
  }
162
- for task_name, splits_and_features in _TASK_SPLITS_AND_FEATURES.items()
163
  }
164
 
165
 
166
  class P3Config(datasets.BuilderConfig):
167
  """BuilderConfig for P3."""
168
 
169
- def __init__(self, splits, features, score_eval, **kwargs):
170
  """BuilderConfig for P3.
171
 
172
  Args:
173
  splits: `List[str]`, the lists of splits which are available for this task
174
- features: `List[str]`, the list of features for this task
175
  score_eval: `bool`, whether this is task formulated as a rank classification problem
176
  **kwargs: keyword arguments forwarded to super.
177
  """
@@ -179,7 +143,7 @@ class P3Config(datasets.BuilderConfig):
179
  # 0.1 initial commit
180
  super(P3Config, self).__init__(version=datasets.Version("0.1.0"), **kwargs)
181
  self.splits = splits
182
- self.features = features
183
  self.score_eval = score_eval
184
 
185
 
@@ -189,11 +153,11 @@ class P3(datasets.GeneratorBasedBuilder):
189
  BUILDER_CONFIGS = [
190
  P3Config(
191
  name=task_name,
192
- splits=splits_and_features["splits"],
193
- features=splits_and_features["features"],
194
  score_eval=task_name.endswith("score_eval")
195
  )
196
- for task_name, splits_and_features in _TASK_SPLITS_AND_FEATURES.items()
197
  ]
198
 
199
  def _info(self):
@@ -211,7 +175,7 @@ class P3(datasets.GeneratorBasedBuilder):
211
  }
212
 
213
  features = {}
214
- for feat_name in self.config.features:
215
  features[feat_name] = _FEAT_MAPPING[feat_name]
216
 
217
  return datasets.DatasetInfo(
@@ -233,7 +197,6 @@ class P3(datasets.GeneratorBasedBuilder):
233
  datasets.SplitGenerator(
234
  name=datasets.Split.TRAIN,
235
  gen_kwargs={
236
- "features_file": data_dir[task_name][split_name]["features_file"],
237
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
238
  }
239
  )
@@ -244,7 +207,6 @@ class P3(datasets.GeneratorBasedBuilder):
244
  datasets.SplitGenerator(
245
  name=datasets.Split.VALIDATION,
246
  gen_kwargs={
247
- "features_file": data_dir[task_name][split_name]["features_file"],
248
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
249
  }
250
  )
@@ -255,7 +217,6 @@ class P3(datasets.GeneratorBasedBuilder):
255
  datasets.SplitGenerator(
256
  name=datasets.Split.TEST,
257
  gen_kwargs={
258
- "features_file": data_dir[task_name][split_name]["features_file"],
259
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
260
  }
261
  )
@@ -267,7 +228,6 @@ class P3(datasets.GeneratorBasedBuilder):
267
  datasets.SplitGenerator(
268
  name=datasets.Split(special_split_name),
269
  gen_kwargs={
270
- "features_file": data_dir[task_name][special_split_name]["features_file"],
271
  "tfrecord": data_dir[task_name][special_split_name]["tfrecord"],
272
  }
273
  )
@@ -275,7 +235,7 @@ class P3(datasets.GeneratorBasedBuilder):
275
  return split_generators
276
 
277
 
278
- def _generate_examples(self, features_file, tfrecord):
279
  """This function returns the examples in the raw (text) form."""
280
  _FEAT_MAPPING_FUNCTIONS = {
281
  "answer_choices": lambda x: [choice.decode("utf-8") for choice in x],
@@ -289,7 +249,9 @@ class P3(datasets.GeneratorBasedBuilder):
289
  }
290
 
291
  key = 0
292
- ds = load_cached_task(features_file, tfrecord)
 
 
293
  for ex in ds.as_numpy_iterator():
294
  ex_dict = {}
295
  for feat_name, feat_value in ex.items():
 
16
 
17
 
18
  import datasets
 
19
  import json
20
+ import urllib
21
  from collections import defaultdict
22
  import tensorflow as tf
23
 
 
26
  TODO"""
27
 
28
  _DESCRIPTION = """\
29
+ P3 (Public Pool of Prompts)is a collection of prompted English datasets covering a diverse set of NLP tasks. A prompt is the combination of an input template and a target template. The templates are functions mapping a data example into natural language for the input and target sequences. For example, in the case of an NLI dataset, the data example would include fields for *Premise, Hypothesis, Label*. An input template would be *If {Premise} is true, is it also true that {Hypothesis}?*, whereas a target template can be defined with the label choices *Choices[label]*. Here *Choices* is prompt-specific metadata that consists of the options *yes, maybe, no* corresponding to *label* being entailment (0), neutral (1) or contradiction (2).
30
 
31
  Prompts are collected using [Promptsource](https://github.com/bigscience-workshop/promptsource), an interface to interactively write prompts on datasets, and collect prompt-specific metadata such as evaluation metrics. As of October 13th, there are 2'000 prompts collected for 270+ data(sub)sets. The collection of prompts is publicly available on [Promptsource](https://github.com/bigscience-workshop/promptsource).
32
 
 
37
 
38
  _HOMEPAGE = "https://github.com/bigscience-workshop/promptsource"
39
 
40
+ _DATA_PATH = "/home/hf/P3/data"
41
+ _HUB_PATH = "https://huggingface.co/datasets/bigscience/P3/raw/main"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
 
43
+
44
+ def load_cached_task(features_dict, tfrecord):
45
  # Use `FixedLenSequenceFeature` for sequences with variable length.
46
  def _feature_config(shape, dtype):
47
  if dtype in ("int32", "bool"):
 
54
  return tf.io.FixedLenFeature(shape, dtype)
55
 
56
  feature_description = {
57
+ feat: _feature_config(**desc) for feat, desc in features_dict.items()
58
  }
59
 
60
+ ds = tf.data.TFRecordDataset(tf.io.gfile.glob([tfrecord])) #TODO handle multiple shards
61
  ds = ds.map(
62
  lambda pb: tf.io.parse_single_example(pb, feature_description),
63
  num_parallel_calls=tf.data.experimental.AUTOTUNE
64
  )
65
  # Cast features back to the types from the info JSON since some features
66
+ # must be cast for storage (e.g., int32 is stored as int64).
67
  ds = ds.map(
68
+ lambda x: {k: tf.cast(v, features_dict[k]["dtype"]) for k, v in x.items()},
69
  num_parallel_calls=tf.data.experimental.AUTOTUNE
70
  )
71
  return ds
72
 
73
 
74
+ def read_from_url(url):
75
+ # TODO: there might be a better way to handle these downloads (especially regarding caching).
76
+ # TODO: Ultimately, we should rely on the cache if internet is not available.
77
+ try:
78
+ content = urllib.request.urlopen(url, timeout=10.0)
79
+ except urllib.error.URLError as e:
80
+ raise ConnectionError(e)
81
+ return content.read().decode("utf-8")
82
 
 
 
83
 
84
+ def find_task_splits_and_features_dict():
85
+ """Get the task available (list was pre-computed by `print_data_split_sizes.py`), and get the features for each task."""
86
+ task_splits_and_features = defaultdict(dict)
87
+
88
+ data = read_from_url(f"{_HUB_PATH}/data_split_sizes.csv")
89
+ data = [t.strip() for t in data.splitlines()]
90
+ data = data[1:]
91
+ data = [t.split("|") for t in data]
92
+ data = [(t[0], json.loads(t[1])) for t in data]
93
+
94
+ for task_name, split_sizes in data:
95
+ if "adversarial_qa" not in task_name: #TODO remove
96
+ continue
97
 
98
+ for split_name in split_sizes.keys():
99
+ split_info = json.loads(
100
+ read_from_url(
101
+ f"{_HUB_PATH}/data/{task_name}/info.{split_name}.json"
102
+ )
103
+ )
104
+ features_dict = split_info["features"]
105
+ assert split_info["num_shards"] == 1 #TODO -> change to multiple shards
106
 
107
+ if not task_splits_and_features[task_name]:
108
+ task_splits_and_features[task_name] = {
 
109
  "splits": [],
110
+ "features_dict": features_dict,
111
  }
112
+ task_splits_and_features[task_name]["splits"].append(split_name)
113
+ assert features_dict == task_splits_and_features[task_name]["features_dict"]
114
 
115
+ return task_splits_and_features
 
 
 
 
 
116
 
117
 
118
+ _TASK_SPLITS_AND_FEATURES_DICT = find_task_splits_and_features_dict()
119
  _URLs = {
120
  task_name: {
121
  split_name: {
122
  "tfrecord": f"{_DATA_PATH}/{task_name}/{split_name}.tfrecord-00000-of-00001",
 
123
  }
124
+ for split_name in splits_and_features_dict["splits"]
125
  }
126
+ for task_name, splits_and_features_dict in _TASK_SPLITS_AND_FEATURES_DICT.items()
127
  }
128
 
129
 
130
  class P3Config(datasets.BuilderConfig):
131
  """BuilderConfig for P3."""
132
 
133
+ def __init__(self, splits, features_dict, score_eval, **kwargs):
134
  """BuilderConfig for P3.
135
 
136
  Args:
137
  splits: `List[str]`, the lists of splits which are available for this task
138
+ features_dict: `dict`, the dict of features for this task
139
  score_eval: `bool`, whether this is task formulated as a rank classification problem
140
  **kwargs: keyword arguments forwarded to super.
141
  """
 
143
  # 0.1 initial commit
144
  super(P3Config, self).__init__(version=datasets.Version("0.1.0"), **kwargs)
145
  self.splits = splits
146
+ self.features_dict = features_dict
147
  self.score_eval = score_eval
148
 
149
 
 
153
  BUILDER_CONFIGS = [
154
  P3Config(
155
  name=task_name,
156
+ splits=splits_and_features_dict["splits"],
157
+ features_dict=splits_and_features_dict["features_dict"],
158
  score_eval=task_name.endswith("score_eval")
159
  )
160
+ for task_name, splits_and_features_dict in _TASK_SPLITS_AND_FEATURES_DICT.items()
161
  ]
162
 
163
  def _info(self):
 
175
  }
176
 
177
  features = {}
178
+ for feat_name in self.config.features_dict.keys():
179
  features[feat_name] = _FEAT_MAPPING[feat_name]
180
 
181
  return datasets.DatasetInfo(
 
197
  datasets.SplitGenerator(
198
  name=datasets.Split.TRAIN,
199
  gen_kwargs={
 
200
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
201
  }
202
  )
 
207
  datasets.SplitGenerator(
208
  name=datasets.Split.VALIDATION,
209
  gen_kwargs={
 
210
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
211
  }
212
  )
 
217
  datasets.SplitGenerator(
218
  name=datasets.Split.TEST,
219
  gen_kwargs={
 
220
  "tfrecord": data_dir[task_name][split_name]["tfrecord"],
221
  }
222
  )
 
228
  datasets.SplitGenerator(
229
  name=datasets.Split(special_split_name),
230
  gen_kwargs={
 
231
  "tfrecord": data_dir[task_name][special_split_name]["tfrecord"],
232
  }
233
  )
 
235
  return split_generators
236
 
237
 
238
+ def _generate_examples(self, tfrecord):
239
  """This function returns the examples in the raw (text) form."""
240
  _FEAT_MAPPING_FUNCTIONS = {
241
  "answer_choices": lambda x: [choice.decode("utf-8") for choice in x],
 
249
  }
250
 
251
  key = 0
252
+ features_dict = self.config.features_dict
253
+ ds = load_cached_task(features_dict, tfrecord)
254
+
255
  for ex in ds.as_numpy_iterator():
256
  ex_dict = {}
257
  for feat_name, feat_value in ex.items():
print_data_split_sizes.py ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import glob
2
+ import json
3
+ import os
4
+
5
+ from collections import defaultdict
6
+
7
+ _DATA_PATH = "data"
8
+
9
+ data_split_sizes = defaultdict(dict)
10
+
11
+ for stats in glob.glob(f"{_DATA_PATH}/*/stats.*.json"):
12
+ folder_path = os.path.dirname(stats)
13
+ task_name = folder_path.split("/")[-1]
14
+ split_name = os.path.basename(stats).split(".")[1]
15
+
16
+ if not os.path.exists(f"{folder_path}/COMPLETED"):
17
+ continue
18
+
19
+ with open(stats, "r") as f:
20
+ split_stats = json.load(f)
21
+ nb_examples = split_stats["examples"]
22
+
23
+ if nb_examples > 0:
24
+ data_split_sizes[task_name][split_name] = nb_examples
25
+
26
+ with open("data_split_sizes.csv", "w", encoding="utf=8") as f:
27
+ f.write("Data(sub)set|Number of examples per splits\n")
28
+ for task_name in sorted(list(data_split_sizes.keys())):
29
+ split_sizes_dict = json.dumps(data_split_sizes[task_name])
30
+ f.write(f"{task_name}|{split_sizes_dict}\n")