Spaces:
Running
Running
Pradeep Kumar
commited on
Delete dual_encoder_dataloader.py
Browse files- dual_encoder_dataloader.py +0 -147
dual_encoder_dataloader.py
DELETED
@@ -1,147 +0,0 @@
|
|
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 |
-
"""Loads dataset for the dual encoder (retrieval) task."""
|
16 |
-
import dataclasses
|
17 |
-
import functools
|
18 |
-
import itertools
|
19 |
-
from typing import Iterable, Mapping, Optional, Tuple
|
20 |
-
|
21 |
-
import tensorflow as tf, tf_keras
|
22 |
-
import tensorflow_hub as hub
|
23 |
-
|
24 |
-
from official.common import dataset_fn
|
25 |
-
from official.core import config_definitions as cfg
|
26 |
-
from official.core import input_reader
|
27 |
-
from official.nlp.data import data_loader
|
28 |
-
from official.nlp.data import data_loader_factory
|
29 |
-
from official.nlp.modeling import layers
|
30 |
-
|
31 |
-
|
32 |
-
@dataclasses.dataclass
|
33 |
-
class DualEncoderDataConfig(cfg.DataConfig):
|
34 |
-
"""Data config for dual encoder task (tasks/dual_encoder)."""
|
35 |
-
# Either set `input_path`...
|
36 |
-
input_path: str = ''
|
37 |
-
# ...or `tfds_name` and `tfds_split` to specify input.
|
38 |
-
tfds_name: str = ''
|
39 |
-
tfds_split: str = ''
|
40 |
-
global_batch_size: int = 32
|
41 |
-
# Either build preprocessing with Python code by specifying these values...
|
42 |
-
vocab_file: str = ''
|
43 |
-
lower_case: bool = True
|
44 |
-
# ...or load preprocessing from a SavedModel at this location.
|
45 |
-
preprocessing_hub_module_url: str = ''
|
46 |
-
|
47 |
-
left_text_fields: Tuple[str] = ('left_input',)
|
48 |
-
right_text_fields: Tuple[str] = ('right_input',)
|
49 |
-
is_training: bool = True
|
50 |
-
seq_length: int = 128
|
51 |
-
file_type: str = 'tfrecord'
|
52 |
-
|
53 |
-
|
54 |
-
@data_loader_factory.register_data_loader_cls(DualEncoderDataConfig)
|
55 |
-
class DualEncoderDataLoader(data_loader.DataLoader):
|
56 |
-
"""A class to load dataset for dual encoder task (tasks/dual_encoder)."""
|
57 |
-
|
58 |
-
def __init__(self, params):
|
59 |
-
if bool(params.tfds_name) == bool(params.input_path):
|
60 |
-
raise ValueError('Must specify either `tfds_name` and `tfds_split` '
|
61 |
-
'or `input_path`.')
|
62 |
-
if bool(params.vocab_file) == bool(params.preprocessing_hub_module_url):
|
63 |
-
raise ValueError('Must specify exactly one of vocab_file (with matching '
|
64 |
-
'lower_case flag) or preprocessing_hub_module_url.')
|
65 |
-
self._params = params
|
66 |
-
self._seq_length = params.seq_length
|
67 |
-
self._left_text_fields = params.left_text_fields
|
68 |
-
self._right_text_fields = params.right_text_fields
|
69 |
-
|
70 |
-
if params.preprocessing_hub_module_url:
|
71 |
-
preprocessing_hub_module = hub.load(params.preprocessing_hub_module_url)
|
72 |
-
self._tokenizer = preprocessing_hub_module.tokenize
|
73 |
-
self._pack_inputs = functools.partial(
|
74 |
-
preprocessing_hub_module.bert_pack_inputs,
|
75 |
-
seq_length=params.seq_length)
|
76 |
-
else:
|
77 |
-
self._tokenizer = layers.BertTokenizer(
|
78 |
-
vocab_file=params.vocab_file, lower_case=params.lower_case)
|
79 |
-
self._pack_inputs = layers.BertPackInputs(
|
80 |
-
seq_length=params.seq_length,
|
81 |
-
special_tokens_dict=self._tokenizer.get_special_tokens_dict())
|
82 |
-
|
83 |
-
def _decode(self, record: tf.Tensor):
|
84 |
-
"""Decodes a serialized tf.Example."""
|
85 |
-
name_to_features = {
|
86 |
-
x: tf.io.FixedLenFeature([], tf.string)
|
87 |
-
for x in itertools.chain(
|
88 |
-
*[self._left_text_fields, self._right_text_fields])
|
89 |
-
}
|
90 |
-
example = tf.io.parse_single_example(record, name_to_features)
|
91 |
-
|
92 |
-
# tf.Example only supports tf.int64, but the TPU only supports tf.int32.
|
93 |
-
# So cast all int64 to int32.
|
94 |
-
for name in example:
|
95 |
-
t = example[name]
|
96 |
-
if t.dtype == tf.int64:
|
97 |
-
t = tf.cast(t, tf.int32)
|
98 |
-
example[name] = t
|
99 |
-
|
100 |
-
return example
|
101 |
-
|
102 |
-
def _bert_tokenize(
|
103 |
-
self, record: Mapping[str, tf.Tensor],
|
104 |
-
text_fields: Iterable[str]) -> Tuple[tf.Tensor, tf.Tensor, tf.Tensor]:
|
105 |
-
"""Tokenize the input in text_fields using BERT tokenizer.
|
106 |
-
|
107 |
-
Args:
|
108 |
-
record: A tfexample record contains the features.
|
109 |
-
text_fields: A list of fields to be tokenzied.
|
110 |
-
|
111 |
-
Returns:
|
112 |
-
The tokenized features in a tuple of (input_word_ids, input_mask,
|
113 |
-
input_type_ids).
|
114 |
-
"""
|
115 |
-
segments_text = [record[x] for x in text_fields]
|
116 |
-
segments_tokens = [self._tokenizer(s) for s in segments_text]
|
117 |
-
segments = [tf.cast(x.merge_dims(1, 2), tf.int32) for x in segments_tokens]
|
118 |
-
return self._pack_inputs(segments)
|
119 |
-
|
120 |
-
def _bert_preprocess(
|
121 |
-
self, record: Mapping[str, tf.Tensor]) -> Mapping[str, tf.Tensor]:
|
122 |
-
"""Perform the bert word piece tokenization for left and right inputs."""
|
123 |
-
|
124 |
-
def _switch_prefix(string, old, new):
|
125 |
-
if string.startswith(old): return new + string[len(old):]
|
126 |
-
raise ValueError('Expected {} to start with {}'.format(string, old))
|
127 |
-
|
128 |
-
def _switch_key_prefix(d, old, new):
|
129 |
-
return {_switch_prefix(key, old, new): value for key, value in d.items()} # pytype: disable=attribute-error # trace-all-classes
|
130 |
-
|
131 |
-
model_inputs = _switch_key_prefix(
|
132 |
-
self._bert_tokenize(record, self._left_text_fields),
|
133 |
-
'input_', 'left_')
|
134 |
-
model_inputs.update(_switch_key_prefix(
|
135 |
-
self._bert_tokenize(record, self._right_text_fields),
|
136 |
-
'input_', 'right_'))
|
137 |
-
return model_inputs
|
138 |
-
|
139 |
-
def load(self, input_context: Optional[tf.distribute.InputContext] = None):
|
140 |
-
"""Returns a tf.dataset.Dataset."""
|
141 |
-
reader = input_reader.InputReader(
|
142 |
-
params=self._params,
|
143 |
-
# Skip `decoder_fn` for tfds input.
|
144 |
-
decoder_fn=self._decode if self._params.input_path else None,
|
145 |
-
dataset_fn=dataset_fn.pick_dataset_fn(self._params.file_type),
|
146 |
-
postprocess_fn=self._bert_preprocess)
|
147 |
-
return reader.read(input_context)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|