Spaces:
Sleeping
Sleeping
Pradeep Kumar
commited on
Commit
•
d6fca6a
1
Parent(s):
17cab64
Upload classifier_data_lib.py
Browse files- classifier_data_lib.py +1612 -0
classifier_data_lib.py
ADDED
@@ -0,0 +1,1612 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 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 |
+
"""BERT library to process data for classification task."""
|
16 |
+
|
17 |
+
import collections
|
18 |
+
import csv
|
19 |
+
import importlib
|
20 |
+
import json
|
21 |
+
import os
|
22 |
+
|
23 |
+
from absl import logging
|
24 |
+
import tensorflow as tf, tf_keras
|
25 |
+
import tensorflow_datasets as tfds
|
26 |
+
|
27 |
+
from official.nlp.tools import tokenization
|
28 |
+
|
29 |
+
|
30 |
+
class InputExample(object):
|
31 |
+
"""A single training/test example for simple seq regression/classification."""
|
32 |
+
|
33 |
+
def __init__(self,
|
34 |
+
guid,
|
35 |
+
text_a,
|
36 |
+
text_b=None,
|
37 |
+
label=None,
|
38 |
+
weight=None,
|
39 |
+
example_id=None):
|
40 |
+
"""Constructs a InputExample.
|
41 |
+
|
42 |
+
Args:
|
43 |
+
guid: Unique id for the example.
|
44 |
+
text_a: string. The untokenized text of the first sequence. For single
|
45 |
+
sequence tasks, only this sequence must be specified.
|
46 |
+
text_b: (Optional) string. The untokenized text of the second sequence.
|
47 |
+
Only must be specified for sequence pair tasks.
|
48 |
+
label: (Optional) string for classification, float for regression. The
|
49 |
+
label of the example. This should be specified for train and dev
|
50 |
+
examples, but not for test examples.
|
51 |
+
weight: (Optional) float. The weight of the example to be used during
|
52 |
+
training.
|
53 |
+
example_id: (Optional) int. The int identification number of example in
|
54 |
+
the corpus.
|
55 |
+
"""
|
56 |
+
self.guid = guid
|
57 |
+
self.text_a = text_a
|
58 |
+
self.text_b = text_b
|
59 |
+
self.label = label
|
60 |
+
self.weight = weight
|
61 |
+
self.example_id = example_id
|
62 |
+
|
63 |
+
|
64 |
+
class InputFeatures(object):
|
65 |
+
"""A single set of features of data."""
|
66 |
+
|
67 |
+
def __init__(self,
|
68 |
+
input_ids,
|
69 |
+
input_mask,
|
70 |
+
segment_ids,
|
71 |
+
label_id,
|
72 |
+
is_real_example=True,
|
73 |
+
weight=None,
|
74 |
+
example_id=None):
|
75 |
+
self.input_ids = input_ids
|
76 |
+
self.input_mask = input_mask
|
77 |
+
self.segment_ids = segment_ids
|
78 |
+
self.label_id = label_id
|
79 |
+
self.is_real_example = is_real_example
|
80 |
+
self.weight = weight
|
81 |
+
self.example_id = example_id
|
82 |
+
|
83 |
+
|
84 |
+
class DataProcessor(object):
|
85 |
+
"""Base class for converters for seq regression/classification datasets."""
|
86 |
+
|
87 |
+
def __init__(self, process_text_fn=tokenization.convert_to_unicode):
|
88 |
+
self.process_text_fn = process_text_fn
|
89 |
+
self.is_regression = False
|
90 |
+
self.label_type = None
|
91 |
+
|
92 |
+
def get_train_examples(self, data_dir):
|
93 |
+
"""Gets a collection of `InputExample`s for the train set."""
|
94 |
+
raise NotImplementedError()
|
95 |
+
|
96 |
+
def get_dev_examples(self, data_dir):
|
97 |
+
"""Gets a collection of `InputExample`s for the dev set."""
|
98 |
+
raise NotImplementedError()
|
99 |
+
|
100 |
+
def get_test_examples(self, data_dir):
|
101 |
+
"""Gets a collection of `InputExample`s for prediction."""
|
102 |
+
raise NotImplementedError()
|
103 |
+
|
104 |
+
def get_labels(self):
|
105 |
+
"""Gets the list of labels for this data set."""
|
106 |
+
raise NotImplementedError()
|
107 |
+
|
108 |
+
@staticmethod
|
109 |
+
def get_processor_name():
|
110 |
+
"""Gets the string identifier of the processor."""
|
111 |
+
raise NotImplementedError()
|
112 |
+
|
113 |
+
@classmethod
|
114 |
+
def _read_tsv(cls, input_file, quotechar=None):
|
115 |
+
"""Reads a tab separated value file."""
|
116 |
+
with tf.io.gfile.GFile(input_file, "r") as f:
|
117 |
+
reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
|
118 |
+
lines = []
|
119 |
+
for line in reader:
|
120 |
+
lines.append(line)
|
121 |
+
return lines
|
122 |
+
|
123 |
+
@classmethod
|
124 |
+
def _read_jsonl(cls, input_file):
|
125 |
+
"""Reads a json line file."""
|
126 |
+
with tf.io.gfile.GFile(input_file, "r") as f:
|
127 |
+
lines = []
|
128 |
+
for json_str in f:
|
129 |
+
lines.append(json.loads(json_str))
|
130 |
+
return lines
|
131 |
+
|
132 |
+
def featurize_example(self, *kargs, **kwargs):
|
133 |
+
"""Converts a single `InputExample` into a single `InputFeatures`."""
|
134 |
+
return convert_single_example(*kargs, **kwargs)
|
135 |
+
|
136 |
+
|
137 |
+
class DefaultGLUEDataProcessor(DataProcessor):
|
138 |
+
"""Processor for the SuperGLUE dataset."""
|
139 |
+
|
140 |
+
def get_train_examples(self, data_dir):
|
141 |
+
"""See base class."""
|
142 |
+
return self._create_examples_tfds("train")
|
143 |
+
|
144 |
+
def get_dev_examples(self, data_dir):
|
145 |
+
"""See base class."""
|
146 |
+
return self._create_examples_tfds("validation")
|
147 |
+
|
148 |
+
def get_test_examples(self, data_dir):
|
149 |
+
"""See base class."""
|
150 |
+
return self._create_examples_tfds("test")
|
151 |
+
|
152 |
+
def _create_examples_tfds(self, set_type):
|
153 |
+
"""Creates examples for the training/dev/test sets."""
|
154 |
+
raise NotImplementedError()
|
155 |
+
|
156 |
+
|
157 |
+
class AxProcessor(DataProcessor):
|
158 |
+
"""Processor for the AX dataset (GLUE diagnostics dataset)."""
|
159 |
+
|
160 |
+
def get_train_examples(self, data_dir):
|
161 |
+
"""See base class."""
|
162 |
+
train_mnli_dataset = tfds.load(
|
163 |
+
"glue/mnli", split="train", try_gcs=True).as_numpy_iterator()
|
164 |
+
return self._create_examples_tfds(train_mnli_dataset, "train")
|
165 |
+
|
166 |
+
def get_dev_examples(self, data_dir):
|
167 |
+
"""See base class."""
|
168 |
+
val_mnli_dataset = tfds.load(
|
169 |
+
"glue/mnli", split="validation_matched",
|
170 |
+
try_gcs=True).as_numpy_iterator()
|
171 |
+
return self._create_examples_tfds(val_mnli_dataset, "validation")
|
172 |
+
|
173 |
+
def get_test_examples(self, data_dir):
|
174 |
+
"""See base class."""
|
175 |
+
test_ax_dataset = tfds.load(
|
176 |
+
"glue/ax", split="test", try_gcs=True).as_numpy_iterator()
|
177 |
+
return self._create_examples_tfds(test_ax_dataset, "test")
|
178 |
+
|
179 |
+
def get_labels(self):
|
180 |
+
"""See base class."""
|
181 |
+
return ["contradiction", "entailment", "neutral"]
|
182 |
+
|
183 |
+
@staticmethod
|
184 |
+
def get_processor_name():
|
185 |
+
"""See base class."""
|
186 |
+
return "AX"
|
187 |
+
|
188 |
+
def _create_examples_tfds(self, dataset, set_type):
|
189 |
+
"""Creates examples for the training/dev/test sets."""
|
190 |
+
dataset = list(dataset)
|
191 |
+
dataset.sort(key=lambda x: x["idx"])
|
192 |
+
examples = []
|
193 |
+
for i, example in enumerate(dataset):
|
194 |
+
guid = "%s-%s" % (set_type, i)
|
195 |
+
label = "contradiction"
|
196 |
+
text_a = self.process_text_fn(example["hypothesis"])
|
197 |
+
text_b = self.process_text_fn(example["premise"])
|
198 |
+
if set_type != "test":
|
199 |
+
label = self.get_labels()[example["label"]]
|
200 |
+
examples.append(
|
201 |
+
InputExample(
|
202 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
203 |
+
weight=None))
|
204 |
+
return examples
|
205 |
+
|
206 |
+
|
207 |
+
class ColaProcessor(DefaultGLUEDataProcessor):
|
208 |
+
"""Processor for the CoLA data set (GLUE version)."""
|
209 |
+
|
210 |
+
def get_labels(self):
|
211 |
+
"""See base class."""
|
212 |
+
return ["0", "1"]
|
213 |
+
|
214 |
+
@staticmethod
|
215 |
+
def get_processor_name():
|
216 |
+
"""See base class."""
|
217 |
+
return "COLA"
|
218 |
+
|
219 |
+
def _create_examples_tfds(self, set_type):
|
220 |
+
"""Creates examples for the training/dev/test sets."""
|
221 |
+
dataset = tfds.load(
|
222 |
+
"glue/cola", split=set_type, try_gcs=True).as_numpy_iterator()
|
223 |
+
dataset = list(dataset)
|
224 |
+
dataset.sort(key=lambda x: x["idx"])
|
225 |
+
examples = []
|
226 |
+
for i, example in enumerate(dataset):
|
227 |
+
guid = "%s-%s" % (set_type, i)
|
228 |
+
label = "0"
|
229 |
+
text_a = self.process_text_fn(example["sentence"])
|
230 |
+
if set_type != "test":
|
231 |
+
label = str(example["label"])
|
232 |
+
examples.append(
|
233 |
+
InputExample(
|
234 |
+
guid=guid, text_a=text_a, text_b=None, label=label, weight=None))
|
235 |
+
return examples
|
236 |
+
|
237 |
+
|
238 |
+
class ImdbProcessor(DataProcessor):
|
239 |
+
"""Processor for the IMDb dataset."""
|
240 |
+
|
241 |
+
def get_labels(self):
|
242 |
+
return ["neg", "pos"]
|
243 |
+
|
244 |
+
def get_train_examples(self, data_dir):
|
245 |
+
return self._create_examples(os.path.join(data_dir, "train"))
|
246 |
+
|
247 |
+
def get_dev_examples(self, data_dir):
|
248 |
+
return self._create_examples(os.path.join(data_dir, "test"))
|
249 |
+
|
250 |
+
@staticmethod
|
251 |
+
def get_processor_name():
|
252 |
+
"""See base class."""
|
253 |
+
return "IMDB"
|
254 |
+
|
255 |
+
def _create_examples(self, data_dir):
|
256 |
+
"""Creates examples."""
|
257 |
+
examples = []
|
258 |
+
for label in ["neg", "pos"]:
|
259 |
+
cur_dir = os.path.join(data_dir, label)
|
260 |
+
for filename in tf.io.gfile.listdir(cur_dir):
|
261 |
+
if not filename.endswith("txt"):
|
262 |
+
continue
|
263 |
+
|
264 |
+
if len(examples) % 1000 == 0:
|
265 |
+
logging.info("Loading dev example %d", len(examples))
|
266 |
+
|
267 |
+
path = os.path.join(cur_dir, filename)
|
268 |
+
with tf.io.gfile.GFile(path, "r") as f:
|
269 |
+
text = f.read().strip().replace("<br />", " ")
|
270 |
+
examples.append(
|
271 |
+
InputExample(
|
272 |
+
guid="unused_id", text_a=text, text_b=None, label=label))
|
273 |
+
return examples
|
274 |
+
|
275 |
+
|
276 |
+
class MnliProcessor(DataProcessor):
|
277 |
+
"""Processor for the MultiNLI data set (GLUE version)."""
|
278 |
+
|
279 |
+
def __init__(self,
|
280 |
+
mnli_type="matched",
|
281 |
+
process_text_fn=tokenization.convert_to_unicode):
|
282 |
+
super(MnliProcessor, self).__init__(process_text_fn)
|
283 |
+
self.dataset = tfds.load("glue/mnli", try_gcs=True)
|
284 |
+
if mnli_type not in ("matched", "mismatched"):
|
285 |
+
raise ValueError("Invalid `mnli_type`: %s" % mnli_type)
|
286 |
+
self.mnli_type = mnli_type
|
287 |
+
|
288 |
+
def get_train_examples(self, data_dir):
|
289 |
+
"""See base class."""
|
290 |
+
return self._create_examples_tfds("train")
|
291 |
+
|
292 |
+
def get_dev_examples(self, data_dir):
|
293 |
+
"""See base class."""
|
294 |
+
if self.mnli_type == "matched":
|
295 |
+
return self._create_examples_tfds("validation_matched")
|
296 |
+
else:
|
297 |
+
return self._create_examples_tfds("validation_mismatched")
|
298 |
+
|
299 |
+
def get_test_examples(self, data_dir):
|
300 |
+
"""See base class."""
|
301 |
+
if self.mnli_type == "matched":
|
302 |
+
return self._create_examples_tfds("test_matched")
|
303 |
+
else:
|
304 |
+
return self._create_examples_tfds("test_mismatched")
|
305 |
+
|
306 |
+
def get_labels(self):
|
307 |
+
"""See base class."""
|
308 |
+
return ["contradiction", "entailment", "neutral"]
|
309 |
+
|
310 |
+
@staticmethod
|
311 |
+
def get_processor_name():
|
312 |
+
"""See base class."""
|
313 |
+
return "MNLI"
|
314 |
+
|
315 |
+
def _create_examples_tfds(self, set_type):
|
316 |
+
"""Creates examples for the training/dev/test sets."""
|
317 |
+
dataset = tfds.load(
|
318 |
+
"glue/mnli", split=set_type, try_gcs=True).as_numpy_iterator()
|
319 |
+
dataset = list(dataset)
|
320 |
+
dataset.sort(key=lambda x: x["idx"])
|
321 |
+
examples = []
|
322 |
+
for i, example in enumerate(dataset):
|
323 |
+
guid = "%s-%s" % (set_type, i)
|
324 |
+
label = "contradiction"
|
325 |
+
text_a = self.process_text_fn(example["hypothesis"])
|
326 |
+
text_b = self.process_text_fn(example["premise"])
|
327 |
+
if set_type != "test":
|
328 |
+
label = self.get_labels()[example["label"]]
|
329 |
+
examples.append(
|
330 |
+
InputExample(
|
331 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
332 |
+
weight=None))
|
333 |
+
return examples
|
334 |
+
|
335 |
+
|
336 |
+
class MrpcProcessor(DefaultGLUEDataProcessor):
|
337 |
+
"""Processor for the MRPC data set (GLUE version)."""
|
338 |
+
|
339 |
+
def get_labels(self):
|
340 |
+
"""See base class."""
|
341 |
+
return ["0", "1"]
|
342 |
+
|
343 |
+
@staticmethod
|
344 |
+
def get_processor_name():
|
345 |
+
"""See base class."""
|
346 |
+
return "MRPC"
|
347 |
+
|
348 |
+
def _create_examples_tfds(self, set_type):
|
349 |
+
"""Creates examples for the training/dev/test sets."""
|
350 |
+
dataset = tfds.load(
|
351 |
+
"glue/mrpc", split=set_type, try_gcs=True).as_numpy_iterator()
|
352 |
+
dataset = list(dataset)
|
353 |
+
dataset.sort(key=lambda x: x["idx"])
|
354 |
+
examples = []
|
355 |
+
for i, example in enumerate(dataset):
|
356 |
+
guid = "%s-%s" % (set_type, i)
|
357 |
+
label = "0"
|
358 |
+
text_a = self.process_text_fn(example["sentence1"])
|
359 |
+
text_b = self.process_text_fn(example["sentence2"])
|
360 |
+
if set_type != "test":
|
361 |
+
label = str(example["label"])
|
362 |
+
examples.append(
|
363 |
+
InputExample(
|
364 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
365 |
+
weight=None))
|
366 |
+
return examples
|
367 |
+
|
368 |
+
|
369 |
+
class PawsxProcessor(DataProcessor):
|
370 |
+
"""Processor for the PAWS-X data set."""
|
371 |
+
supported_languages = ["de", "en", "es", "fr", "ja", "ko", "zh"]
|
372 |
+
|
373 |
+
def __init__(self,
|
374 |
+
language="en",
|
375 |
+
process_text_fn=tokenization.convert_to_unicode):
|
376 |
+
super(PawsxProcessor, self).__init__(process_text_fn)
|
377 |
+
if language == "all":
|
378 |
+
self.languages = PawsxProcessor.supported_languages
|
379 |
+
elif language not in PawsxProcessor.supported_languages:
|
380 |
+
raise ValueError("language %s is not supported for PAWS-X task." %
|
381 |
+
language)
|
382 |
+
else:
|
383 |
+
self.languages = [language]
|
384 |
+
|
385 |
+
def get_train_examples(self, data_dir):
|
386 |
+
"""See base class."""
|
387 |
+
lines = []
|
388 |
+
for language in self.languages:
|
389 |
+
if language == "en":
|
390 |
+
train_tsv = "train.tsv"
|
391 |
+
else:
|
392 |
+
train_tsv = "translated_train.tsv"
|
393 |
+
# Skips the header.
|
394 |
+
lines.extend(
|
395 |
+
self._read_tsv(os.path.join(data_dir, language, train_tsv))[1:])
|
396 |
+
|
397 |
+
examples = []
|
398 |
+
for i, line in enumerate(lines):
|
399 |
+
guid = "train-%d" % i
|
400 |
+
text_a = self.process_text_fn(line[1])
|
401 |
+
text_b = self.process_text_fn(line[2])
|
402 |
+
label = self.process_text_fn(line[3])
|
403 |
+
examples.append(
|
404 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
405 |
+
return examples
|
406 |
+
|
407 |
+
def get_dev_examples(self, data_dir):
|
408 |
+
"""See base class."""
|
409 |
+
lines = []
|
410 |
+
for lang in PawsxProcessor.supported_languages:
|
411 |
+
lines.extend(
|
412 |
+
self._read_tsv(os.path.join(data_dir, lang, "dev_2k.tsv"))[1:])
|
413 |
+
|
414 |
+
examples = []
|
415 |
+
for i, line in enumerate(lines):
|
416 |
+
guid = "dev-%d" % i
|
417 |
+
text_a = self.process_text_fn(line[1])
|
418 |
+
text_b = self.process_text_fn(line[2])
|
419 |
+
label = self.process_text_fn(line[3])
|
420 |
+
examples.append(
|
421 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
422 |
+
return examples
|
423 |
+
|
424 |
+
def get_test_examples(self, data_dir):
|
425 |
+
"""See base class."""
|
426 |
+
examples_by_lang = {k: [] for k in self.supported_languages}
|
427 |
+
for lang in self.supported_languages:
|
428 |
+
lines = self._read_tsv(os.path.join(data_dir, lang, "test_2k.tsv"))[1:]
|
429 |
+
for i, line in enumerate(lines):
|
430 |
+
guid = "test-%d" % i
|
431 |
+
text_a = self.process_text_fn(line[1])
|
432 |
+
text_b = self.process_text_fn(line[2])
|
433 |
+
label = self.process_text_fn(line[3])
|
434 |
+
examples_by_lang[lang].append(
|
435 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
436 |
+
return examples_by_lang
|
437 |
+
|
438 |
+
def get_labels(self):
|
439 |
+
"""See base class."""
|
440 |
+
return ["0", "1"]
|
441 |
+
|
442 |
+
@staticmethod
|
443 |
+
def get_processor_name():
|
444 |
+
"""See base class."""
|
445 |
+
return "XTREME-PAWS-X"
|
446 |
+
|
447 |
+
|
448 |
+
class QnliProcessor(DefaultGLUEDataProcessor):
|
449 |
+
"""Processor for the QNLI data set (GLUE version)."""
|
450 |
+
|
451 |
+
def get_labels(self):
|
452 |
+
"""See base class."""
|
453 |
+
return ["entailment", "not_entailment"]
|
454 |
+
|
455 |
+
@staticmethod
|
456 |
+
def get_processor_name():
|
457 |
+
"""See base class."""
|
458 |
+
return "QNLI"
|
459 |
+
|
460 |
+
def _create_examples_tfds(self, set_type):
|
461 |
+
"""Creates examples for the training/dev/test sets."""
|
462 |
+
dataset = tfds.load(
|
463 |
+
"glue/qnli", split=set_type, try_gcs=True).as_numpy_iterator()
|
464 |
+
dataset = list(dataset)
|
465 |
+
dataset.sort(key=lambda x: x["idx"])
|
466 |
+
examples = []
|
467 |
+
for i, example in enumerate(dataset):
|
468 |
+
guid = "%s-%s" % (set_type, i)
|
469 |
+
label = "entailment"
|
470 |
+
text_a = self.process_text_fn(example["question"])
|
471 |
+
text_b = self.process_text_fn(example["sentence"])
|
472 |
+
if set_type != "test":
|
473 |
+
label = self.get_labels()[example["label"]]
|
474 |
+
examples.append(
|
475 |
+
InputExample(
|
476 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
477 |
+
weight=None))
|
478 |
+
return examples
|
479 |
+
|
480 |
+
|
481 |
+
class QqpProcessor(DefaultGLUEDataProcessor):
|
482 |
+
"""Processor for the QQP data set (GLUE version)."""
|
483 |
+
|
484 |
+
def get_labels(self):
|
485 |
+
"""See base class."""
|
486 |
+
return ["0", "1"]
|
487 |
+
|
488 |
+
@staticmethod
|
489 |
+
def get_processor_name():
|
490 |
+
"""See base class."""
|
491 |
+
return "QQP"
|
492 |
+
|
493 |
+
def _create_examples_tfds(self, set_type):
|
494 |
+
"""Creates examples for the training/dev/test sets."""
|
495 |
+
dataset = tfds.load(
|
496 |
+
"glue/qqp", split=set_type, try_gcs=True).as_numpy_iterator()
|
497 |
+
dataset = list(dataset)
|
498 |
+
dataset.sort(key=lambda x: x["idx"])
|
499 |
+
examples = []
|
500 |
+
for i, example in enumerate(dataset):
|
501 |
+
guid = "%s-%s" % (set_type, i)
|
502 |
+
label = "0"
|
503 |
+
text_a = self.process_text_fn(example["question1"])
|
504 |
+
text_b = self.process_text_fn(example["question2"])
|
505 |
+
if set_type != "test":
|
506 |
+
label = str(example["label"])
|
507 |
+
examples.append(
|
508 |
+
InputExample(
|
509 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
510 |
+
weight=None))
|
511 |
+
return examples
|
512 |
+
|
513 |
+
|
514 |
+
class RteProcessor(DefaultGLUEDataProcessor):
|
515 |
+
"""Processor for the RTE data set (GLUE version)."""
|
516 |
+
|
517 |
+
def get_labels(self):
|
518 |
+
"""See base class."""
|
519 |
+
# All datasets are converted to 2-class split, where for 3-class datasets we
|
520 |
+
# collapse neutral and contradiction into not_entailment.
|
521 |
+
return ["entailment", "not_entailment"]
|
522 |
+
|
523 |
+
@staticmethod
|
524 |
+
def get_processor_name():
|
525 |
+
"""See base class."""
|
526 |
+
return "RTE"
|
527 |
+
|
528 |
+
def _create_examples_tfds(self, set_type):
|
529 |
+
"""Creates examples for the training/dev/test sets."""
|
530 |
+
dataset = tfds.load(
|
531 |
+
"glue/rte", split=set_type, try_gcs=True).as_numpy_iterator()
|
532 |
+
dataset = list(dataset)
|
533 |
+
dataset.sort(key=lambda x: x["idx"])
|
534 |
+
examples = []
|
535 |
+
for i, example in enumerate(dataset):
|
536 |
+
guid = "%s-%s" % (set_type, i)
|
537 |
+
label = "entailment"
|
538 |
+
text_a = self.process_text_fn(example["sentence1"])
|
539 |
+
text_b = self.process_text_fn(example["sentence2"])
|
540 |
+
if set_type != "test":
|
541 |
+
label = self.get_labels()[example["label"]]
|
542 |
+
examples.append(
|
543 |
+
InputExample(
|
544 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
545 |
+
weight=None))
|
546 |
+
return examples
|
547 |
+
|
548 |
+
|
549 |
+
class SstProcessor(DefaultGLUEDataProcessor):
|
550 |
+
"""Processor for the SST-2 data set (GLUE version)."""
|
551 |
+
|
552 |
+
def get_labels(self):
|
553 |
+
"""See base class."""
|
554 |
+
return ["0", "1"]
|
555 |
+
|
556 |
+
@staticmethod
|
557 |
+
def get_processor_name():
|
558 |
+
"""See base class."""
|
559 |
+
return "SST-2"
|
560 |
+
|
561 |
+
def _create_examples_tfds(self, set_type):
|
562 |
+
"""Creates examples for the training/dev/test sets."""
|
563 |
+
dataset = tfds.load(
|
564 |
+
"glue/sst2", split=set_type, try_gcs=True).as_numpy_iterator()
|
565 |
+
dataset = list(dataset)
|
566 |
+
dataset.sort(key=lambda x: x["idx"])
|
567 |
+
examples = []
|
568 |
+
for i, example in enumerate(dataset):
|
569 |
+
guid = "%s-%s" % (set_type, i)
|
570 |
+
label = "0"
|
571 |
+
text_a = self.process_text_fn(example["sentence"])
|
572 |
+
if set_type != "test":
|
573 |
+
label = str(example["label"])
|
574 |
+
examples.append(
|
575 |
+
InputExample(
|
576 |
+
guid=guid, text_a=text_a, text_b=None, label=label, weight=None))
|
577 |
+
return examples
|
578 |
+
|
579 |
+
|
580 |
+
class StsBProcessor(DefaultGLUEDataProcessor):
|
581 |
+
"""Processor for the STS-B data set (GLUE version)."""
|
582 |
+
|
583 |
+
def __init__(self, process_text_fn=tokenization.convert_to_unicode):
|
584 |
+
super(StsBProcessor, self).__init__(process_text_fn=process_text_fn)
|
585 |
+
self.is_regression = True
|
586 |
+
self.label_type = float
|
587 |
+
self._labels = None
|
588 |
+
|
589 |
+
def _create_examples_tfds(self, set_type):
|
590 |
+
"""Creates examples for the training/dev/test sets."""
|
591 |
+
dataset = tfds.load(
|
592 |
+
"glue/stsb", split=set_type, try_gcs=True).as_numpy_iterator()
|
593 |
+
dataset = list(dataset)
|
594 |
+
dataset.sort(key=lambda x: x["idx"])
|
595 |
+
examples = []
|
596 |
+
for i, example in enumerate(dataset):
|
597 |
+
guid = "%s-%s" % (set_type, i)
|
598 |
+
label = 0.0
|
599 |
+
text_a = self.process_text_fn(example["sentence1"])
|
600 |
+
text_b = self.process_text_fn(example["sentence2"])
|
601 |
+
if set_type != "test":
|
602 |
+
label = self.label_type(example["label"])
|
603 |
+
examples.append(
|
604 |
+
InputExample(
|
605 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
606 |
+
weight=None))
|
607 |
+
return examples
|
608 |
+
|
609 |
+
def get_labels(self):
|
610 |
+
"""See base class."""
|
611 |
+
return self._labels
|
612 |
+
|
613 |
+
@staticmethod
|
614 |
+
def get_processor_name():
|
615 |
+
"""See base class."""
|
616 |
+
return "STS-B"
|
617 |
+
|
618 |
+
|
619 |
+
class TfdsProcessor(DataProcessor):
|
620 |
+
"""Processor for generic text classification and regression TFDS data set.
|
621 |
+
|
622 |
+
The TFDS parameters are expected to be provided in the tfds_params string, in
|
623 |
+
a comma-separated list of parameter assignments.
|
624 |
+
Examples:
|
625 |
+
tfds_params="dataset=scicite,text_key=string"
|
626 |
+
tfds_params="dataset=imdb_reviews,test_split=,dev_split=test"
|
627 |
+
tfds_params="dataset=glue/cola,text_key=sentence"
|
628 |
+
tfds_params="dataset=glue/sst2,text_key=sentence"
|
629 |
+
tfds_params="dataset=glue/qnli,text_key=question,text_b_key=sentence"
|
630 |
+
tfds_params="dataset=glue/mrpc,text_key=sentence1,text_b_key=sentence2"
|
631 |
+
tfds_params="dataset=glue/stsb,text_key=sentence1,text_b_key=sentence2,"
|
632 |
+
"is_regression=true,label_type=float"
|
633 |
+
tfds_params="dataset=snli,text_key=premise,text_b_key=hypothesis,"
|
634 |
+
"skip_label=-1"
|
635 |
+
Possible parameters (please refer to the documentation of Tensorflow Datasets
|
636 |
+
(TFDS) for the meaning of individual parameters):
|
637 |
+
dataset: Required dataset name (potentially with subset and version number).
|
638 |
+
data_dir: Optional TFDS source root directory.
|
639 |
+
module_import: Optional Dataset module to import.
|
640 |
+
train_split: Name of the train split (defaults to `train`).
|
641 |
+
dev_split: Name of the dev split (defaults to `validation`).
|
642 |
+
test_split: Name of the test split (defaults to `test`).
|
643 |
+
text_key: Key of the text_a feature (defaults to `text`).
|
644 |
+
text_b_key: Key of the second text feature if available.
|
645 |
+
label_key: Key of the label feature (defaults to `label`).
|
646 |
+
test_text_key: Key of the text feature to use in test set.
|
647 |
+
test_text_b_key: Key of the second text feature to use in test set.
|
648 |
+
test_label: String to be used as the label for all test examples.
|
649 |
+
label_type: Type of the label key (defaults to `int`).
|
650 |
+
weight_key: Key of the float sample weight (is not used if not provided).
|
651 |
+
is_regression: Whether the task is a regression problem (defaults to False).
|
652 |
+
skip_label: Skip examples with given label (defaults to None).
|
653 |
+
"""
|
654 |
+
|
655 |
+
def __init__(self,
|
656 |
+
tfds_params,
|
657 |
+
process_text_fn=tokenization.convert_to_unicode):
|
658 |
+
super(TfdsProcessor, self).__init__(process_text_fn)
|
659 |
+
self._process_tfds_params_str(tfds_params)
|
660 |
+
if self.module_import:
|
661 |
+
importlib.import_module(self.module_import)
|
662 |
+
|
663 |
+
self.dataset, info = tfds.load(
|
664 |
+
self.dataset_name, data_dir=self.data_dir, with_info=True)
|
665 |
+
if self.is_regression:
|
666 |
+
self._labels = None
|
667 |
+
else:
|
668 |
+
self._labels = list(range(info.features[self.label_key].num_classes))
|
669 |
+
|
670 |
+
def _process_tfds_params_str(self, params_str):
|
671 |
+
"""Extracts TFDS parameters from a comma-separated assignments string."""
|
672 |
+
dtype_map = {"int": int, "float": float}
|
673 |
+
cast_str_to_bool = lambda s: s.lower() not in ["false", "0"]
|
674 |
+
|
675 |
+
tuples = [x.split("=") for x in params_str.split(",")]
|
676 |
+
d = {k.strip(): v.strip() for k, v in tuples}
|
677 |
+
self.dataset_name = d["dataset"] # Required.
|
678 |
+
self.data_dir = d.get("data_dir", None)
|
679 |
+
self.module_import = d.get("module_import", None)
|
680 |
+
self.train_split = d.get("train_split", "train")
|
681 |
+
self.dev_split = d.get("dev_split", "validation")
|
682 |
+
self.test_split = d.get("test_split", "test")
|
683 |
+
self.text_key = d.get("text_key", "text")
|
684 |
+
self.text_b_key = d.get("text_b_key", None)
|
685 |
+
self.label_key = d.get("label_key", "label")
|
686 |
+
self.test_text_key = d.get("test_text_key", self.text_key)
|
687 |
+
self.test_text_b_key = d.get("test_text_b_key", self.text_b_key)
|
688 |
+
self.test_label = d.get("test_label", "test_example")
|
689 |
+
self.label_type = dtype_map[d.get("label_type", "int")]
|
690 |
+
self.is_regression = cast_str_to_bool(d.get("is_regression", "False"))
|
691 |
+
self.weight_key = d.get("weight_key", None)
|
692 |
+
self.skip_label = d.get("skip_label", None)
|
693 |
+
if self.skip_label is not None:
|
694 |
+
self.skip_label = self.label_type(self.skip_label)
|
695 |
+
|
696 |
+
def get_train_examples(self, data_dir):
|
697 |
+
assert data_dir is None
|
698 |
+
return self._create_examples(self.train_split, "train")
|
699 |
+
|
700 |
+
def get_dev_examples(self, data_dir):
|
701 |
+
assert data_dir is None
|
702 |
+
return self._create_examples(self.dev_split, "dev")
|
703 |
+
|
704 |
+
def get_test_examples(self, data_dir):
|
705 |
+
assert data_dir is None
|
706 |
+
return self._create_examples(self.test_split, "test")
|
707 |
+
|
708 |
+
def get_labels(self):
|
709 |
+
return self._labels
|
710 |
+
|
711 |
+
def get_processor_name(self):
|
712 |
+
return "TFDS_" + self.dataset_name
|
713 |
+
|
714 |
+
def _create_examples(self, split_name, set_type):
|
715 |
+
"""Creates examples for the training/dev/test sets."""
|
716 |
+
if split_name not in self.dataset:
|
717 |
+
raise ValueError("Split {} not available.".format(split_name))
|
718 |
+
dataset = self.dataset[split_name].as_numpy_iterator()
|
719 |
+
examples = []
|
720 |
+
text_b, weight = None, None
|
721 |
+
for i, example in enumerate(dataset):
|
722 |
+
guid = "%s-%s" % (set_type, i)
|
723 |
+
if set_type == "test":
|
724 |
+
text_a = self.process_text_fn(example[self.test_text_key])
|
725 |
+
if self.test_text_b_key:
|
726 |
+
text_b = self.process_text_fn(example[self.test_text_b_key])
|
727 |
+
label = self.test_label
|
728 |
+
else:
|
729 |
+
text_a = self.process_text_fn(example[self.text_key])
|
730 |
+
if self.text_b_key:
|
731 |
+
text_b = self.process_text_fn(example[self.text_b_key])
|
732 |
+
label = self.label_type(example[self.label_key])
|
733 |
+
if self.skip_label is not None and label == self.skip_label:
|
734 |
+
continue
|
735 |
+
if self.weight_key:
|
736 |
+
weight = float(example[self.weight_key])
|
737 |
+
examples.append(
|
738 |
+
InputExample(
|
739 |
+
guid=guid,
|
740 |
+
text_a=text_a,
|
741 |
+
text_b=text_b,
|
742 |
+
label=label,
|
743 |
+
weight=weight))
|
744 |
+
return examples
|
745 |
+
|
746 |
+
|
747 |
+
class WnliProcessor(DefaultGLUEDataProcessor):
|
748 |
+
"""Processor for the WNLI data set (GLUE version)."""
|
749 |
+
|
750 |
+
def get_labels(self):
|
751 |
+
"""See base class."""
|
752 |
+
return ["0", "1"]
|
753 |
+
|
754 |
+
@staticmethod
|
755 |
+
def get_processor_name():
|
756 |
+
"""See base class."""
|
757 |
+
return "WNLI"
|
758 |
+
|
759 |
+
def _create_examples_tfds(self, set_type):
|
760 |
+
"""Creates examples for the training/dev/test sets."""
|
761 |
+
dataset = tfds.load(
|
762 |
+
"glue/wnli", split=set_type, try_gcs=True).as_numpy_iterator()
|
763 |
+
dataset = list(dataset)
|
764 |
+
dataset.sort(key=lambda x: x["idx"])
|
765 |
+
examples = []
|
766 |
+
for i, example in enumerate(dataset):
|
767 |
+
guid = "%s-%s" % (set_type, i)
|
768 |
+
label = "0"
|
769 |
+
text_a = self.process_text_fn(example["sentence1"])
|
770 |
+
text_b = self.process_text_fn(example["sentence2"])
|
771 |
+
if set_type != "test":
|
772 |
+
label = str(example["label"])
|
773 |
+
examples.append(
|
774 |
+
InputExample(
|
775 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label,
|
776 |
+
weight=None))
|
777 |
+
return examples
|
778 |
+
|
779 |
+
|
780 |
+
class XnliProcessor(DataProcessor):
|
781 |
+
"""Processor for the XNLI data set."""
|
782 |
+
supported_languages = [
|
783 |
+
"ar", "bg", "de", "el", "en", "es", "fr", "hi", "ru", "sw", "th", "tr",
|
784 |
+
"ur", "vi", "zh"
|
785 |
+
]
|
786 |
+
|
787 |
+
def __init__(self,
|
788 |
+
language="en",
|
789 |
+
process_text_fn=tokenization.convert_to_unicode):
|
790 |
+
super(XnliProcessor, self).__init__(process_text_fn)
|
791 |
+
if language == "all":
|
792 |
+
self.languages = XnliProcessor.supported_languages
|
793 |
+
elif language not in XnliProcessor.supported_languages:
|
794 |
+
raise ValueError("language %s is not supported for XNLI task." % language)
|
795 |
+
else:
|
796 |
+
self.languages = [language]
|
797 |
+
|
798 |
+
def get_train_examples(self, data_dir):
|
799 |
+
"""See base class."""
|
800 |
+
lines = []
|
801 |
+
for language in self.languages:
|
802 |
+
# Skips the header.
|
803 |
+
lines.extend(
|
804 |
+
self._read_tsv(
|
805 |
+
os.path.join(data_dir, "multinli",
|
806 |
+
"multinli.train.%s.tsv" % language))[1:])
|
807 |
+
|
808 |
+
examples = []
|
809 |
+
for i, line in enumerate(lines):
|
810 |
+
guid = "train-%d" % i
|
811 |
+
text_a = self.process_text_fn(line[0])
|
812 |
+
text_b = self.process_text_fn(line[1])
|
813 |
+
label = self.process_text_fn(line[2])
|
814 |
+
if label == self.process_text_fn("contradictory"):
|
815 |
+
label = self.process_text_fn("contradiction")
|
816 |
+
examples.append(
|
817 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
818 |
+
return examples
|
819 |
+
|
820 |
+
def get_dev_examples(self, data_dir):
|
821 |
+
"""See base class."""
|
822 |
+
lines = self._read_tsv(os.path.join(data_dir, "xnli.dev.tsv"))
|
823 |
+
examples = []
|
824 |
+
for i, line in enumerate(lines):
|
825 |
+
if i == 0:
|
826 |
+
continue
|
827 |
+
guid = "dev-%d" % i
|
828 |
+
text_a = self.process_text_fn(line[6])
|
829 |
+
text_b = self.process_text_fn(line[7])
|
830 |
+
label = self.process_text_fn(line[1])
|
831 |
+
examples.append(
|
832 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
833 |
+
return examples
|
834 |
+
|
835 |
+
def get_test_examples(self, data_dir):
|
836 |
+
"""See base class."""
|
837 |
+
lines = self._read_tsv(os.path.join(data_dir, "xnli.test.tsv"))
|
838 |
+
examples_by_lang = {k: [] for k in XnliProcessor.supported_languages}
|
839 |
+
for i, line in enumerate(lines):
|
840 |
+
if i == 0:
|
841 |
+
continue
|
842 |
+
guid = "test-%d" % i
|
843 |
+
language = self.process_text_fn(line[0])
|
844 |
+
text_a = self.process_text_fn(line[6])
|
845 |
+
text_b = self.process_text_fn(line[7])
|
846 |
+
label = self.process_text_fn(line[1])
|
847 |
+
examples_by_lang[language].append(
|
848 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
849 |
+
return examples_by_lang
|
850 |
+
|
851 |
+
def get_labels(self):
|
852 |
+
"""See base class."""
|
853 |
+
return ["contradiction", "entailment", "neutral"]
|
854 |
+
|
855 |
+
@staticmethod
|
856 |
+
def get_processor_name():
|
857 |
+
"""See base class."""
|
858 |
+
return "XNLI"
|
859 |
+
|
860 |
+
|
861 |
+
class XtremePawsxProcessor(DataProcessor):
|
862 |
+
"""Processor for the XTREME PAWS-X data set."""
|
863 |
+
supported_languages = ["de", "en", "es", "fr", "ja", "ko", "zh"]
|
864 |
+
|
865 |
+
def __init__(self,
|
866 |
+
process_text_fn=tokenization.convert_to_unicode,
|
867 |
+
translated_data_dir=None,
|
868 |
+
only_use_en_dev=True):
|
869 |
+
"""See base class.
|
870 |
+
|
871 |
+
Args:
|
872 |
+
process_text_fn: See base class.
|
873 |
+
translated_data_dir: If specified, will also include translated data in
|
874 |
+
the training and testing data.
|
875 |
+
only_use_en_dev: If True, only use english dev data. Otherwise, use dev
|
876 |
+
data from all languages.
|
877 |
+
"""
|
878 |
+
super(XtremePawsxProcessor, self).__init__(process_text_fn)
|
879 |
+
self.translated_data_dir = translated_data_dir
|
880 |
+
self.only_use_en_dev = only_use_en_dev
|
881 |
+
|
882 |
+
def get_train_examples(self, data_dir):
|
883 |
+
"""See base class."""
|
884 |
+
examples = []
|
885 |
+
if self.translated_data_dir is None:
|
886 |
+
lines = self._read_tsv(os.path.join(data_dir, "train-en.tsv"))
|
887 |
+
for i, line in enumerate(lines):
|
888 |
+
guid = "train-%d" % i
|
889 |
+
text_a = self.process_text_fn(line[0])
|
890 |
+
text_b = self.process_text_fn(line[1])
|
891 |
+
label = self.process_text_fn(line[2])
|
892 |
+
examples.append(
|
893 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
894 |
+
else:
|
895 |
+
for lang in self.supported_languages:
|
896 |
+
lines = self._read_tsv(
|
897 |
+
os.path.join(self.translated_data_dir, "translate-train",
|
898 |
+
f"en-{lang}-translated.tsv"))
|
899 |
+
for i, line in enumerate(lines):
|
900 |
+
guid = f"train-{lang}-{i}"
|
901 |
+
text_a = self.process_text_fn(line[2])
|
902 |
+
text_b = self.process_text_fn(line[3])
|
903 |
+
label = self.process_text_fn(line[4])
|
904 |
+
examples.append(
|
905 |
+
InputExample(
|
906 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
907 |
+
return examples
|
908 |
+
|
909 |
+
def get_dev_examples(self, data_dir):
|
910 |
+
"""See base class."""
|
911 |
+
examples = []
|
912 |
+
if self.only_use_en_dev:
|
913 |
+
lines = self._read_tsv(os.path.join(data_dir, "dev-en.tsv"))
|
914 |
+
for i, line in enumerate(lines):
|
915 |
+
guid = "dev-%d" % i
|
916 |
+
text_a = self.process_text_fn(line[0])
|
917 |
+
text_b = self.process_text_fn(line[1])
|
918 |
+
label = self.process_text_fn(line[2])
|
919 |
+
examples.append(
|
920 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
921 |
+
else:
|
922 |
+
for lang in self.supported_languages:
|
923 |
+
lines = self._read_tsv(os.path.join(data_dir, f"dev-{lang}.tsv"))
|
924 |
+
for i, line in enumerate(lines):
|
925 |
+
guid = f"dev-{lang}-{i}"
|
926 |
+
text_a = self.process_text_fn(line[0])
|
927 |
+
text_b = self.process_text_fn(line[1])
|
928 |
+
label = self.process_text_fn(line[2])
|
929 |
+
examples.append(
|
930 |
+
InputExample(
|
931 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
932 |
+
return examples
|
933 |
+
|
934 |
+
def get_test_examples(self, data_dir):
|
935 |
+
"""See base class."""
|
936 |
+
examples_by_lang = {}
|
937 |
+
for lang in self.supported_languages:
|
938 |
+
examples_by_lang[lang] = []
|
939 |
+
lines = self._read_tsv(os.path.join(data_dir, f"test-{lang}.tsv"))
|
940 |
+
for i, line in enumerate(lines):
|
941 |
+
guid = f"test-{lang}-{i}"
|
942 |
+
text_a = self.process_text_fn(line[0])
|
943 |
+
text_b = self.process_text_fn(line[1])
|
944 |
+
label = "0"
|
945 |
+
examples_by_lang[lang].append(
|
946 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
947 |
+
if self.translated_data_dir is not None:
|
948 |
+
for lang in self.supported_languages:
|
949 |
+
if lang == "en":
|
950 |
+
continue
|
951 |
+
examples_by_lang[f"{lang}-en"] = []
|
952 |
+
lines = self._read_tsv(
|
953 |
+
os.path.join(self.translated_data_dir, "translate-test",
|
954 |
+
f"test-{lang}-en-translated.tsv"))
|
955 |
+
for i, line in enumerate(lines):
|
956 |
+
guid = f"test-{lang}-en-{i}"
|
957 |
+
text_a = self.process_text_fn(line[2])
|
958 |
+
text_b = self.process_text_fn(line[3])
|
959 |
+
label = "0"
|
960 |
+
examples_by_lang[f"{lang}-en"].append(
|
961 |
+
InputExample(
|
962 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
963 |
+
return examples_by_lang
|
964 |
+
|
965 |
+
def get_labels(self):
|
966 |
+
"""See base class."""
|
967 |
+
return ["0", "1"]
|
968 |
+
|
969 |
+
@staticmethod
|
970 |
+
def get_processor_name():
|
971 |
+
"""See base class."""
|
972 |
+
return "XTREME-PAWS-X"
|
973 |
+
|
974 |
+
|
975 |
+
class XtremeXnliProcessor(DataProcessor):
|
976 |
+
"""Processor for the XTREME XNLI data set."""
|
977 |
+
supported_languages = [
|
978 |
+
"ar", "bg", "de", "el", "en", "es", "fr", "hi", "ru", "sw", "th", "tr",
|
979 |
+
"ur", "vi", "zh"
|
980 |
+
]
|
981 |
+
|
982 |
+
def __init__(self,
|
983 |
+
process_text_fn=tokenization.convert_to_unicode,
|
984 |
+
translated_data_dir=None,
|
985 |
+
only_use_en_dev=True):
|
986 |
+
"""See base class.
|
987 |
+
|
988 |
+
Args:
|
989 |
+
process_text_fn: See base class.
|
990 |
+
translated_data_dir: If specified, will also include translated data in
|
991 |
+
the training data.
|
992 |
+
only_use_en_dev: If True, only use english dev data. Otherwise, use dev
|
993 |
+
data from all languages.
|
994 |
+
"""
|
995 |
+
super(XtremeXnliProcessor, self).__init__(process_text_fn)
|
996 |
+
self.translated_data_dir = translated_data_dir
|
997 |
+
self.only_use_en_dev = only_use_en_dev
|
998 |
+
|
999 |
+
def get_train_examples(self, data_dir):
|
1000 |
+
"""See base class."""
|
1001 |
+
lines = self._read_tsv(os.path.join(data_dir, "train-en.tsv"))
|
1002 |
+
|
1003 |
+
examples = []
|
1004 |
+
if self.translated_data_dir is None:
|
1005 |
+
for i, line in enumerate(lines):
|
1006 |
+
guid = "train-%d" % i
|
1007 |
+
text_a = self.process_text_fn(line[0])
|
1008 |
+
text_b = self.process_text_fn(line[1])
|
1009 |
+
label = self.process_text_fn(line[2])
|
1010 |
+
if label == self.process_text_fn("contradictory"):
|
1011 |
+
label = self.process_text_fn("contradiction")
|
1012 |
+
examples.append(
|
1013 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1014 |
+
else:
|
1015 |
+
for lang in self.supported_languages:
|
1016 |
+
lines = self._read_tsv(
|
1017 |
+
os.path.join(self.translated_data_dir, "translate-train",
|
1018 |
+
f"en-{lang}-translated.tsv"))
|
1019 |
+
for i, line in enumerate(lines):
|
1020 |
+
guid = f"train-{lang}-{i}"
|
1021 |
+
text_a = self.process_text_fn(line[2])
|
1022 |
+
text_b = self.process_text_fn(line[3])
|
1023 |
+
label = self.process_text_fn(line[4])
|
1024 |
+
if label == self.process_text_fn("contradictory"):
|
1025 |
+
label = self.process_text_fn("contradiction")
|
1026 |
+
examples.append(
|
1027 |
+
InputExample(
|
1028 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1029 |
+
return examples
|
1030 |
+
|
1031 |
+
def get_dev_examples(self, data_dir):
|
1032 |
+
"""See base class."""
|
1033 |
+
examples = []
|
1034 |
+
if self.only_use_en_dev:
|
1035 |
+
lines = self._read_tsv(os.path.join(data_dir, "dev-en.tsv"))
|
1036 |
+
for i, line in enumerate(lines):
|
1037 |
+
guid = "dev-%d" % i
|
1038 |
+
text_a = self.process_text_fn(line[0])
|
1039 |
+
text_b = self.process_text_fn(line[1])
|
1040 |
+
label = self.process_text_fn(line[2])
|
1041 |
+
examples.append(
|
1042 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1043 |
+
else:
|
1044 |
+
for lang in self.supported_languages:
|
1045 |
+
lines = self._read_tsv(os.path.join(data_dir, f"dev-{lang}.tsv"))
|
1046 |
+
for i, line in enumerate(lines):
|
1047 |
+
guid = f"dev-{lang}-{i}"
|
1048 |
+
text_a = self.process_text_fn(line[0])
|
1049 |
+
text_b = self.process_text_fn(line[1])
|
1050 |
+
label = self.process_text_fn(line[2])
|
1051 |
+
if label == self.process_text_fn("contradictory"):
|
1052 |
+
label = self.process_text_fn("contradiction")
|
1053 |
+
examples.append(
|
1054 |
+
InputExample(
|
1055 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1056 |
+
return examples
|
1057 |
+
|
1058 |
+
def get_test_examples(self, data_dir):
|
1059 |
+
"""See base class."""
|
1060 |
+
examples_by_lang = {}
|
1061 |
+
for lang in self.supported_languages:
|
1062 |
+
examples_by_lang[lang] = []
|
1063 |
+
lines = self._read_tsv(os.path.join(data_dir, f"test-{lang}.tsv"))
|
1064 |
+
for i, line in enumerate(lines):
|
1065 |
+
guid = f"test-{lang}-{i}"
|
1066 |
+
text_a = self.process_text_fn(line[0])
|
1067 |
+
text_b = self.process_text_fn(line[1])
|
1068 |
+
label = "contradiction"
|
1069 |
+
examples_by_lang[lang].append(
|
1070 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1071 |
+
if self.translated_data_dir is not None:
|
1072 |
+
for lang in self.supported_languages:
|
1073 |
+
if lang == "en":
|
1074 |
+
continue
|
1075 |
+
examples_by_lang[f"{lang}-en"] = []
|
1076 |
+
lines = self._read_tsv(
|
1077 |
+
os.path.join(self.translated_data_dir, "translate-test",
|
1078 |
+
f"test-{lang}-en-translated.tsv"))
|
1079 |
+
for i, line in enumerate(lines):
|
1080 |
+
guid = f"test-{lang}-en-{i}"
|
1081 |
+
text_a = self.process_text_fn(line[2])
|
1082 |
+
text_b = self.process_text_fn(line[3])
|
1083 |
+
label = "contradiction"
|
1084 |
+
examples_by_lang[f"{lang}-en"].append(
|
1085 |
+
InputExample(
|
1086 |
+
guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1087 |
+
return examples_by_lang
|
1088 |
+
|
1089 |
+
def get_labels(self):
|
1090 |
+
"""See base class."""
|
1091 |
+
return ["contradiction", "entailment", "neutral"]
|
1092 |
+
|
1093 |
+
@staticmethod
|
1094 |
+
def get_processor_name():
|
1095 |
+
"""See base class."""
|
1096 |
+
return "XTREME-XNLI"
|
1097 |
+
|
1098 |
+
|
1099 |
+
def convert_single_example(ex_index, example, label_list, max_seq_length,
|
1100 |
+
tokenizer):
|
1101 |
+
"""Converts a single `InputExample` into a single `InputFeatures`."""
|
1102 |
+
label_map = {}
|
1103 |
+
if label_list:
|
1104 |
+
for (i, label) in enumerate(label_list):
|
1105 |
+
label_map[label] = i
|
1106 |
+
|
1107 |
+
tokens_a = tokenizer.tokenize(example.text_a)
|
1108 |
+
tokens_b = None
|
1109 |
+
if example.text_b:
|
1110 |
+
tokens_b = tokenizer.tokenize(example.text_b)
|
1111 |
+
|
1112 |
+
if tokens_b:
|
1113 |
+
# Modifies `tokens_a` and `tokens_b` in place so that the total
|
1114 |
+
# length is less than the specified length.
|
1115 |
+
# Account for [CLS], [SEP], [SEP] with "- 3"
|
1116 |
+
_truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
|
1117 |
+
else:
|
1118 |
+
# Account for [CLS] and [SEP] with "- 2"
|
1119 |
+
if len(tokens_a) > max_seq_length - 2:
|
1120 |
+
tokens_a = tokens_a[0:(max_seq_length - 2)]
|
1121 |
+
|
1122 |
+
seg_id_a = 0
|
1123 |
+
seg_id_b = 1
|
1124 |
+
seg_id_cls = 0
|
1125 |
+
seg_id_pad = 0
|
1126 |
+
|
1127 |
+
# The convention in BERT is:
|
1128 |
+
# (a) For sequence pairs:
|
1129 |
+
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
|
1130 |
+
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
|
1131 |
+
# (b) For single sequences:
|
1132 |
+
# tokens: [CLS] the dog is hairy . [SEP]
|
1133 |
+
# type_ids: 0 0 0 0 0 0 0
|
1134 |
+
#
|
1135 |
+
# Where "type_ids" are used to indicate whether this is the first
|
1136 |
+
# sequence or the second sequence. The embedding vectors for `type=0` and
|
1137 |
+
# `type=1` were learned during pre-training and are added to the wordpiece
|
1138 |
+
# embedding vector (and position vector). This is not *strictly* necessary
|
1139 |
+
# since the [SEP] token unambiguously separates the sequences, but it makes
|
1140 |
+
# it easier for the model to learn the concept of sequences.
|
1141 |
+
#
|
1142 |
+
# For classification tasks, the first vector (corresponding to [CLS]) is
|
1143 |
+
# used as the "sentence vector". Note that this only makes sense because
|
1144 |
+
# the entire model is fine-tuned.
|
1145 |
+
tokens = []
|
1146 |
+
segment_ids = []
|
1147 |
+
tokens.append("[CLS]")
|
1148 |
+
segment_ids.append(seg_id_cls)
|
1149 |
+
for token in tokens_a:
|
1150 |
+
tokens.append(token)
|
1151 |
+
segment_ids.append(seg_id_a)
|
1152 |
+
tokens.append("[SEP]")
|
1153 |
+
segment_ids.append(seg_id_a)
|
1154 |
+
|
1155 |
+
if tokens_b:
|
1156 |
+
for token in tokens_b:
|
1157 |
+
tokens.append(token)
|
1158 |
+
segment_ids.append(seg_id_b)
|
1159 |
+
tokens.append("[SEP]")
|
1160 |
+
segment_ids.append(seg_id_b)
|
1161 |
+
|
1162 |
+
input_ids = tokenizer.convert_tokens_to_ids(tokens)
|
1163 |
+
|
1164 |
+
# The mask has 1 for real tokens and 0 for padding tokens. Only real
|
1165 |
+
# tokens are attended to.
|
1166 |
+
input_mask = [1] * len(input_ids)
|
1167 |
+
|
1168 |
+
# Zero-pad up to the sequence length.
|
1169 |
+
while len(input_ids) < max_seq_length:
|
1170 |
+
input_ids.append(0)
|
1171 |
+
input_mask.append(0)
|
1172 |
+
segment_ids.append(seg_id_pad)
|
1173 |
+
|
1174 |
+
assert len(input_ids) == max_seq_length
|
1175 |
+
assert len(input_mask) == max_seq_length
|
1176 |
+
assert len(segment_ids) == max_seq_length
|
1177 |
+
|
1178 |
+
label_id = label_map[example.label] if label_map else example.label
|
1179 |
+
if ex_index < 5:
|
1180 |
+
logging.info("*** Example ***")
|
1181 |
+
logging.info("guid: %s", (example.guid))
|
1182 |
+
logging.info("tokens: %s",
|
1183 |
+
" ".join([tokenization.printable_text(x) for x in tokens]))
|
1184 |
+
logging.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
|
1185 |
+
logging.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
|
1186 |
+
logging.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
|
1187 |
+
logging.info("label: %s (id = %s)", example.label, str(label_id))
|
1188 |
+
logging.info("weight: %s", example.weight)
|
1189 |
+
logging.info("example_id: %s", example.example_id)
|
1190 |
+
|
1191 |
+
feature = InputFeatures(
|
1192 |
+
input_ids=input_ids,
|
1193 |
+
input_mask=input_mask,
|
1194 |
+
segment_ids=segment_ids,
|
1195 |
+
label_id=label_id,
|
1196 |
+
is_real_example=True,
|
1197 |
+
weight=example.weight,
|
1198 |
+
example_id=example.example_id)
|
1199 |
+
|
1200 |
+
return feature
|
1201 |
+
|
1202 |
+
|
1203 |
+
class AXgProcessor(DataProcessor):
|
1204 |
+
"""Processor for the AXg dataset (SuperGLUE diagnostics dataset)."""
|
1205 |
+
|
1206 |
+
def get_test_examples(self, data_dir):
|
1207 |
+
"""See base class."""
|
1208 |
+
return self._create_examples(
|
1209 |
+
self._read_jsonl(os.path.join(data_dir, "AX-g.jsonl")), "test")
|
1210 |
+
|
1211 |
+
def get_labels(self):
|
1212 |
+
"""See base class."""
|
1213 |
+
return ["entailment", "not_entailment"]
|
1214 |
+
|
1215 |
+
@staticmethod
|
1216 |
+
def get_processor_name():
|
1217 |
+
"""See base class."""
|
1218 |
+
return "AXg"
|
1219 |
+
|
1220 |
+
def _create_examples(self, lines, set_type):
|
1221 |
+
"""Creates examples for the training/dev/test sets."""
|
1222 |
+
examples = []
|
1223 |
+
for line in lines:
|
1224 |
+
guid = "%s-%s" % (set_type, self.process_text_fn(str(line["idx"])))
|
1225 |
+
text_a = self.process_text_fn(line["premise"])
|
1226 |
+
text_b = self.process_text_fn(line["hypothesis"])
|
1227 |
+
label = self.process_text_fn(line["label"])
|
1228 |
+
examples.append(
|
1229 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1230 |
+
return examples
|
1231 |
+
|
1232 |
+
|
1233 |
+
class BoolQProcessor(DefaultGLUEDataProcessor):
|
1234 |
+
"""Processor for the BoolQ dataset (SuperGLUE diagnostics dataset)."""
|
1235 |
+
|
1236 |
+
def get_labels(self):
|
1237 |
+
"""See base class."""
|
1238 |
+
return ["True", "False"]
|
1239 |
+
|
1240 |
+
@staticmethod
|
1241 |
+
def get_processor_name():
|
1242 |
+
"""See base class."""
|
1243 |
+
return "BoolQ"
|
1244 |
+
|
1245 |
+
def _create_examples_tfds(self, set_type):
|
1246 |
+
"""Creates examples for the training/dev/test sets."""
|
1247 |
+
dataset = tfds.load(
|
1248 |
+
"super_glue/boolq", split=set_type, try_gcs=True).as_numpy_iterator()
|
1249 |
+
examples = []
|
1250 |
+
for example in dataset:
|
1251 |
+
guid = "%s-%s" % (set_type, self.process_text_fn(str(example["idx"])))
|
1252 |
+
text_a = self.process_text_fn(example["question"])
|
1253 |
+
text_b = self.process_text_fn(example["passage"])
|
1254 |
+
label = "False"
|
1255 |
+
if set_type != "test":
|
1256 |
+
label = self.get_labels()[example["label"]]
|
1257 |
+
examples.append(
|
1258 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1259 |
+
return examples
|
1260 |
+
|
1261 |
+
|
1262 |
+
class CBProcessor(DefaultGLUEDataProcessor):
|
1263 |
+
"""Processor for the CB dataset (SuperGLUE diagnostics dataset)."""
|
1264 |
+
|
1265 |
+
def get_labels(self):
|
1266 |
+
"""See base class."""
|
1267 |
+
return ["entailment", "neutral", "contradiction"]
|
1268 |
+
|
1269 |
+
@staticmethod
|
1270 |
+
def get_processor_name():
|
1271 |
+
"""See base class."""
|
1272 |
+
return "CB"
|
1273 |
+
|
1274 |
+
def _create_examples_tfds(self, set_type):
|
1275 |
+
"""Creates examples for the training/dev/test sets."""
|
1276 |
+
dataset = tfds.load(
|
1277 |
+
"super_glue/cb", split=set_type, try_gcs=True).as_numpy_iterator()
|
1278 |
+
examples = []
|
1279 |
+
for example in dataset:
|
1280 |
+
guid = "%s-%s" % (set_type, self.process_text_fn(str(example["idx"])))
|
1281 |
+
text_a = self.process_text_fn(example["premise"])
|
1282 |
+
text_b = self.process_text_fn(example["hypothesis"])
|
1283 |
+
label = "entailment"
|
1284 |
+
if set_type != "test":
|
1285 |
+
label = self.get_labels()[example["label"]]
|
1286 |
+
examples.append(
|
1287 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1288 |
+
return examples
|
1289 |
+
|
1290 |
+
|
1291 |
+
class SuperGLUERTEProcessor(DefaultGLUEDataProcessor):
|
1292 |
+
"""Processor for the RTE dataset (SuperGLUE version)."""
|
1293 |
+
|
1294 |
+
def get_labels(self):
|
1295 |
+
"""See base class."""
|
1296 |
+
# All datasets are converted to 2-class split, where for 3-class datasets we
|
1297 |
+
# collapse neutral and contradiction into not_entailment.
|
1298 |
+
return ["entailment", "not_entailment"]
|
1299 |
+
|
1300 |
+
@staticmethod
|
1301 |
+
def get_processor_name():
|
1302 |
+
"""See base class."""
|
1303 |
+
return "RTESuperGLUE"
|
1304 |
+
|
1305 |
+
def _create_examples_tfds(self, set_type):
|
1306 |
+
"""Creates examples for the training/dev/test sets."""
|
1307 |
+
examples = []
|
1308 |
+
dataset = tfds.load(
|
1309 |
+
"super_glue/rte", split=set_type, try_gcs=True).as_numpy_iterator()
|
1310 |
+
for example in dataset:
|
1311 |
+
guid = "%s-%s" % (set_type, self.process_text_fn(str(example["idx"])))
|
1312 |
+
text_a = self.process_text_fn(example["premise"])
|
1313 |
+
text_b = self.process_text_fn(example["hypothesis"])
|
1314 |
+
label = "entailment"
|
1315 |
+
if set_type != "test":
|
1316 |
+
label = self.get_labels()[example["label"]]
|
1317 |
+
examples.append(
|
1318 |
+
InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
|
1319 |
+
return examples
|
1320 |
+
|
1321 |
+
|
1322 |
+
class WiCInputExample(InputExample):
|
1323 |
+
"""Processor for the WiC dataset (SuperGLUE version)."""
|
1324 |
+
|
1325 |
+
def __init__(self,
|
1326 |
+
guid,
|
1327 |
+
text_a,
|
1328 |
+
text_b=None,
|
1329 |
+
label=None,
|
1330 |
+
word=None,
|
1331 |
+
weight=None,
|
1332 |
+
example_id=None):
|
1333 |
+
"""A single training/test example for simple seq regression/classification."""
|
1334 |
+
super(WiCInputExample, self).__init__(guid, text_a, text_b, label, weight,
|
1335 |
+
example_id)
|
1336 |
+
self.word = word
|
1337 |
+
|
1338 |
+
|
1339 |
+
class WiCProcessor(DefaultGLUEDataProcessor):
|
1340 |
+
"""Processor for the RTE dataset (SuperGLUE version)."""
|
1341 |
+
|
1342 |
+
def get_labels(self):
|
1343 |
+
"""Not used."""
|
1344 |
+
return []
|
1345 |
+
|
1346 |
+
@staticmethod
|
1347 |
+
def get_processor_name():
|
1348 |
+
"""See base class."""
|
1349 |
+
return "RTESuperGLUE"
|
1350 |
+
|
1351 |
+
def _create_examples_tfds(self, set_type):
|
1352 |
+
"""Creates examples for the training/dev/test sets."""
|
1353 |
+
examples = []
|
1354 |
+
dataset = tfds.load(
|
1355 |
+
"super_glue/wic", split=set_type, try_gcs=True).as_numpy_iterator()
|
1356 |
+
for example in dataset:
|
1357 |
+
guid = "%s-%s" % (set_type, self.process_text_fn(str(example["idx"])))
|
1358 |
+
text_a = self.process_text_fn(example["sentence1"])
|
1359 |
+
text_b = self.process_text_fn(example["sentence2"])
|
1360 |
+
word = self.process_text_fn(example["word"])
|
1361 |
+
label = 0
|
1362 |
+
if set_type != "test":
|
1363 |
+
label = example["label"]
|
1364 |
+
examples.append(
|
1365 |
+
WiCInputExample(
|
1366 |
+
guid=guid, text_a=text_a, text_b=text_b, word=word, label=label))
|
1367 |
+
return examples
|
1368 |
+
|
1369 |
+
def featurize_example(self, ex_index, example, label_list, max_seq_length,
|
1370 |
+
tokenizer):
|
1371 |
+
"""Here we concate sentence1, sentence2, word together with [SEP] tokens."""
|
1372 |
+
del label_list
|
1373 |
+
tokens_a = tokenizer.tokenize(example.text_a)
|
1374 |
+
tokens_b = tokenizer.tokenize(example.text_b)
|
1375 |
+
tokens_word = tokenizer.tokenize(example.word)
|
1376 |
+
|
1377 |
+
# Modifies `tokens_a` and `tokens_b` in place so that the total
|
1378 |
+
# length is less than the specified length.
|
1379 |
+
# Account for [CLS], [SEP], [SEP], [SEP] with "- 4"
|
1380 |
+
# Here we only pop out the first two sentence tokens.
|
1381 |
+
_truncate_seq_pair(tokens_a, tokens_b,
|
1382 |
+
max_seq_length - 4 - len(tokens_word))
|
1383 |
+
|
1384 |
+
seg_id_a = 0
|
1385 |
+
seg_id_b = 1
|
1386 |
+
seg_id_c = 2
|
1387 |
+
seg_id_cls = 0
|
1388 |
+
seg_id_pad = 0
|
1389 |
+
|
1390 |
+
tokens = []
|
1391 |
+
segment_ids = []
|
1392 |
+
tokens.append("[CLS]")
|
1393 |
+
segment_ids.append(seg_id_cls)
|
1394 |
+
for token in tokens_a:
|
1395 |
+
tokens.append(token)
|
1396 |
+
segment_ids.append(seg_id_a)
|
1397 |
+
tokens.append("[SEP]")
|
1398 |
+
segment_ids.append(seg_id_a)
|
1399 |
+
|
1400 |
+
for token in tokens_b:
|
1401 |
+
tokens.append(token)
|
1402 |
+
segment_ids.append(seg_id_b)
|
1403 |
+
|
1404 |
+
tokens.append("[SEP]")
|
1405 |
+
segment_ids.append(seg_id_b)
|
1406 |
+
|
1407 |
+
for token in tokens_word:
|
1408 |
+
tokens.append(token)
|
1409 |
+
segment_ids.append(seg_id_c)
|
1410 |
+
|
1411 |
+
tokens.append("[SEP]")
|
1412 |
+
segment_ids.append(seg_id_c)
|
1413 |
+
|
1414 |
+
input_ids = tokenizer.convert_tokens_to_ids(tokens)
|
1415 |
+
|
1416 |
+
# The mask has 1 for real tokens and 0 for padding tokens. Only real
|
1417 |
+
# tokens are attended to.
|
1418 |
+
input_mask = [1] * len(input_ids)
|
1419 |
+
|
1420 |
+
# Zero-pad up to the sequence length.
|
1421 |
+
while len(input_ids) < max_seq_length:
|
1422 |
+
input_ids.append(0)
|
1423 |
+
input_mask.append(0)
|
1424 |
+
segment_ids.append(seg_id_pad)
|
1425 |
+
|
1426 |
+
assert len(input_ids) == max_seq_length
|
1427 |
+
assert len(input_mask) == max_seq_length
|
1428 |
+
assert len(segment_ids) == max_seq_length
|
1429 |
+
|
1430 |
+
label_id = example.label
|
1431 |
+
if ex_index < 5:
|
1432 |
+
logging.info("*** Example ***")
|
1433 |
+
logging.info("guid: %s", (example.guid))
|
1434 |
+
logging.info("tokens: %s",
|
1435 |
+
" ".join([tokenization.printable_text(x) for x in tokens]))
|
1436 |
+
logging.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
|
1437 |
+
logging.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
|
1438 |
+
logging.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
|
1439 |
+
logging.info("label: %s (id = %s)", example.label, str(label_id))
|
1440 |
+
logging.info("weight: %s", example.weight)
|
1441 |
+
logging.info("example_id: %s", example.example_id)
|
1442 |
+
|
1443 |
+
feature = InputFeatures(
|
1444 |
+
input_ids=input_ids,
|
1445 |
+
input_mask=input_mask,
|
1446 |
+
segment_ids=segment_ids,
|
1447 |
+
label_id=label_id,
|
1448 |
+
is_real_example=True,
|
1449 |
+
weight=example.weight,
|
1450 |
+
example_id=example.example_id)
|
1451 |
+
|
1452 |
+
return feature
|
1453 |
+
|
1454 |
+
|
1455 |
+
def file_based_convert_examples_to_features(examples,
|
1456 |
+
label_list,
|
1457 |
+
max_seq_length,
|
1458 |
+
tokenizer,
|
1459 |
+
output_file,
|
1460 |
+
label_type=None,
|
1461 |
+
featurize_fn=None):
|
1462 |
+
"""Convert a set of `InputExample`s to a TFRecord file."""
|
1463 |
+
|
1464 |
+
tf.io.gfile.makedirs(os.path.dirname(output_file))
|
1465 |
+
writer = tf.io.TFRecordWriter(output_file)
|
1466 |
+
|
1467 |
+
for ex_index, example in enumerate(examples):
|
1468 |
+
if ex_index % 10000 == 0:
|
1469 |
+
logging.info("Writing example %d of %d", ex_index, len(examples))
|
1470 |
+
|
1471 |
+
if featurize_fn:
|
1472 |
+
feature = featurize_fn(ex_index, example, label_list, max_seq_length,
|
1473 |
+
tokenizer)
|
1474 |
+
else:
|
1475 |
+
feature = convert_single_example(ex_index, example, label_list,
|
1476 |
+
max_seq_length, tokenizer)
|
1477 |
+
|
1478 |
+
def create_int_feature(values):
|
1479 |
+
f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))
|
1480 |
+
return f
|
1481 |
+
|
1482 |
+
def create_float_feature(values):
|
1483 |
+
f = tf.train.Feature(float_list=tf.train.FloatList(value=list(values)))
|
1484 |
+
return f
|
1485 |
+
|
1486 |
+
features = collections.OrderedDict()
|
1487 |
+
features["input_ids"] = create_int_feature(feature.input_ids)
|
1488 |
+
features["input_mask"] = create_int_feature(feature.input_mask)
|
1489 |
+
features["segment_ids"] = create_int_feature(feature.segment_ids)
|
1490 |
+
if label_type is not None and label_type == float:
|
1491 |
+
features["label_ids"] = create_float_feature([feature.label_id])
|
1492 |
+
elif feature.label_id is not None:
|
1493 |
+
features["label_ids"] = create_int_feature([feature.label_id])
|
1494 |
+
features["is_real_example"] = create_int_feature(
|
1495 |
+
[int(feature.is_real_example)])
|
1496 |
+
if feature.weight is not None:
|
1497 |
+
features["weight"] = create_float_feature([feature.weight])
|
1498 |
+
if feature.example_id is not None:
|
1499 |
+
features["example_id"] = create_int_feature([feature.example_id])
|
1500 |
+
else:
|
1501 |
+
features["example_id"] = create_int_feature([ex_index])
|
1502 |
+
|
1503 |
+
tf_example = tf.train.Example(features=tf.train.Features(feature=features))
|
1504 |
+
writer.write(tf_example.SerializeToString())
|
1505 |
+
writer.close()
|
1506 |
+
|
1507 |
+
|
1508 |
+
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
|
1509 |
+
"""Truncates a sequence pair in place to the maximum length."""
|
1510 |
+
|
1511 |
+
# This is a simple heuristic which will always truncate the longer sequence
|
1512 |
+
# one token at a time. This makes more sense than truncating an equal percent
|
1513 |
+
# of tokens from each, since if one sequence is very short then each token
|
1514 |
+
# that's truncated likely contains more information than a longer sequence.
|
1515 |
+
while True:
|
1516 |
+
total_length = len(tokens_a) + len(tokens_b)
|
1517 |
+
if total_length <= max_length:
|
1518 |
+
break
|
1519 |
+
if len(tokens_a) > len(tokens_b):
|
1520 |
+
tokens_a.pop()
|
1521 |
+
else:
|
1522 |
+
tokens_b.pop()
|
1523 |
+
|
1524 |
+
|
1525 |
+
def generate_tf_record_from_data_file(processor,
|
1526 |
+
data_dir,
|
1527 |
+
tokenizer,
|
1528 |
+
train_data_output_path=None,
|
1529 |
+
eval_data_output_path=None,
|
1530 |
+
test_data_output_path=None,
|
1531 |
+
max_seq_length=128):
|
1532 |
+
"""Generates and saves training data into a tf record file.
|
1533 |
+
|
1534 |
+
Args:
|
1535 |
+
processor: Input processor object to be used for generating data. Subclass
|
1536 |
+
of `DataProcessor`.
|
1537 |
+
data_dir: Directory that contains train/eval/test data to process.
|
1538 |
+
tokenizer: The tokenizer to be applied on the data.
|
1539 |
+
train_data_output_path: Output to which processed tf record for training
|
1540 |
+
will be saved.
|
1541 |
+
eval_data_output_path: Output to which processed tf record for evaluation
|
1542 |
+
will be saved.
|
1543 |
+
test_data_output_path: Output to which processed tf record for testing
|
1544 |
+
will be saved. Must be a pattern template with {} if processor has
|
1545 |
+
language specific test data.
|
1546 |
+
max_seq_length: Maximum sequence length of the to be generated
|
1547 |
+
training/eval data.
|
1548 |
+
|
1549 |
+
Returns:
|
1550 |
+
A dictionary containing input meta data.
|
1551 |
+
"""
|
1552 |
+
assert train_data_output_path or eval_data_output_path
|
1553 |
+
|
1554 |
+
label_list = processor.get_labels()
|
1555 |
+
label_type = getattr(processor, "label_type", None)
|
1556 |
+
is_regression = getattr(processor, "is_regression", False)
|
1557 |
+
has_sample_weights = getattr(processor, "weight_key", False)
|
1558 |
+
|
1559 |
+
num_training_data = 0
|
1560 |
+
if train_data_output_path:
|
1561 |
+
train_input_data_examples = processor.get_train_examples(data_dir)
|
1562 |
+
file_based_convert_examples_to_features(train_input_data_examples,
|
1563 |
+
label_list, max_seq_length,
|
1564 |
+
tokenizer, train_data_output_path,
|
1565 |
+
label_type,
|
1566 |
+
processor.featurize_example)
|
1567 |
+
num_training_data = len(train_input_data_examples)
|
1568 |
+
|
1569 |
+
if eval_data_output_path:
|
1570 |
+
eval_input_data_examples = processor.get_dev_examples(data_dir)
|
1571 |
+
file_based_convert_examples_to_features(eval_input_data_examples,
|
1572 |
+
label_list, max_seq_length,
|
1573 |
+
tokenizer, eval_data_output_path,
|
1574 |
+
label_type,
|
1575 |
+
processor.featurize_example)
|
1576 |
+
|
1577 |
+
meta_data = {
|
1578 |
+
"processor_type": processor.get_processor_name(),
|
1579 |
+
"train_data_size": num_training_data,
|
1580 |
+
"max_seq_length": max_seq_length,
|
1581 |
+
}
|
1582 |
+
|
1583 |
+
if test_data_output_path:
|
1584 |
+
test_input_data_examples = processor.get_test_examples(data_dir)
|
1585 |
+
if isinstance(test_input_data_examples, dict):
|
1586 |
+
for language, examples in test_input_data_examples.items():
|
1587 |
+
file_based_convert_examples_to_features(
|
1588 |
+
examples, label_list, max_seq_length, tokenizer,
|
1589 |
+
test_data_output_path.format(language), label_type,
|
1590 |
+
processor.featurize_example)
|
1591 |
+
meta_data["test_{}_data_size".format(language)] = len(examples)
|
1592 |
+
else:
|
1593 |
+
file_based_convert_examples_to_features(test_input_data_examples,
|
1594 |
+
label_list, max_seq_length,
|
1595 |
+
tokenizer, test_data_output_path,
|
1596 |
+
label_type,
|
1597 |
+
processor.featurize_example)
|
1598 |
+
meta_data["test_data_size"] = len(test_input_data_examples)
|
1599 |
+
|
1600 |
+
if is_regression:
|
1601 |
+
meta_data["task_type"] = "bert_regression"
|
1602 |
+
meta_data["label_type"] = {int: "int", float: "float"}[label_type]
|
1603 |
+
else:
|
1604 |
+
meta_data["task_type"] = "bert_classification"
|
1605 |
+
meta_data["num_labels"] = len(processor.get_labels())
|
1606 |
+
if has_sample_weights:
|
1607 |
+
meta_data["has_sample_weights"] = True
|
1608 |
+
|
1609 |
+
if eval_data_output_path:
|
1610 |
+
meta_data["eval_data_size"] = len(eval_input_data_examples)
|
1611 |
+
|
1612 |
+
return meta_data
|