# coding=utf-8
# Copyright (c) 2020 Alibaba PAI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from .base import BaseDataset
from ..model_zoo import AutoTokenizer
from ..utils import io
from ..model_zoo.models.cnn import TextCNNTokenizer
[docs]class BertClassificationDataset(BaseDataset):
def __init__(self,
pretrained_model_name_or_path,
data_file,
max_seq_length,
input_schema,
first_sequence,
label_name=None,
second_sequence=None,
label_enumerate_values=None,
multi_label=False,
*args,
**kwargs):
super(BertClassificationDataset, self).__init__(data_file, input_schema=input_schema,
output_format="dict", *args, **kwargs)
self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path)
self.max_seq_length = max_seq_length
self.multi_label = multi_label
if label_enumerate_values is None:
self._label_enumerate_values = "0,1".split(",")
else:
if io.exists(label_enumerate_values):
with io.open(label_enumerate_values) as f:
self._label_enumerate_values = [line.strip() for line in f]
else:
self._label_enumerate_values = label_enumerate_values.split(",")
self.max_num_labels = len(self._label_enumerate_values)
assert first_sequence in self.column_names, \
"Column name %s needs to be included in columns" % first_sequence
self.first_sequence = first_sequence
if second_sequence:
assert second_sequence in self.column_names, \
"Column name %s needs to be included in columns" % second_sequence
self.second_sequence = second_sequence
else:
self.second_sequence = None
if label_name:
assert label_name in self.column_names, \
"Column name %s needs to be included in columns" % label_name
self.label_name = label_name
else:
self.label_name = None
self.label_map = dict({value: idx for idx, value in enumerate(self.label_enumerate_values)})
@property
def eval_metrics(self):
"""
Returns the evaluation metrics.
"""
return ("accuracy", "f1")
@property
def label_enumerate_values(self):
"""
Returns the label enumerate values.
"""
return self._label_enumerate_values
[docs] def convert_single_row_to_example(self, row):
"""Convert sample token to indices.
Args:
row: contains sequence and label.
text_a: the first sequence in row.
text_b: the second sequence in row if self.second_sequence is true.
label: label token if self.label_name is true.
Returns: sing example
encoding: an example contains token indices.
"""
text_a = row[self.first_sequence]
text_b = row[self.second_sequence] if self.second_sequence else None
label = row[self.label_name] if self.label_name else None
encoding = self.tokenizer(text_a, text_b, padding='max_length', truncation=True,
max_length=self.max_seq_length)
if not self.multi_label:
encoding['label_ids'] = self.label_map[label]
return encoding
else:
label_id = [self.label_map[x] for x in label.split(",") if x]
new_label_id = [0] * self.max_num_labels
for idx in label_id:
new_label_id[idx] = 1
encoding['label_ids'] = new_label_id
return encoding
[docs] def batch_fn(self, features):
"""
Divide examples into batches.
"""
return {k: torch.tensor([dic[k] for dic in features], dtype=torch.long) for k in features[0]}
[docs]class GLUEDataset(BaseDataset):
def __init__(self,
pretrained_model_name_or_path,
data_file,
max_seq_length,
task_name,
**kwargs):
super(GLUEDataset, self).__init__(data_file, skip_first_line=True, **kwargs)
self.task_name = task_name.lower()
self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path)
self.max_seq_length = max_seq_length
self.label_map = dict({value: idx for idx, value in enumerate(self.label_enumerate_values)})
@property
def eval_metrics(self):
task_eval_metrics = {
"cola": ("matthews_corrcoef",),
"sst-2": ("accuracy",),
"mrpc": ("accuracy", "f1"),
"sts-b": ("pearson_and_spearman",),
"qqp": ("accuracy", "f1"),
"mnli": ("accuracy",),
"mnli-mm": ("accuracy",),
"qnli": ("accuracy",),
"rte": ("accuracy",),
"wnli": ("accuracy",),
}
return task_eval_metrics[self.task_name]
@property
def label_enumerate_values(self):
task_label_enumerate_values = {
"cola": ["0", "1"],
"mnli": ["contradiction", "entailment", "neutral"],
"mrpc": ["0", "1"],
"sst-2": ["0", "1"],
"sts-b": [None],
"qqp": ["0", "1"],
"qnli": ["entailment", "not_entailment"],
"rte": ["entailment", "not_entailment"],
"wnli": ["0", "1"]
}
return task_label_enumerate_values[self.task_name]
[docs] def convert_single_row_to_example(self, row):
line = row.strip().split("\t")
task_name = self.task_name
if task_name == "cola":
text_a, text_b, label = line[3], None, line[1]
elif task_name == "mnli":
text_a, text_b, label = line[8], line[9], line[-1]
elif task_name == "mrpc":
text_a, text_b, label = line[3], line[4], line[0]
elif task_name == "sst-2":
text_a, text_b, label = line[0], None, line[1]
elif task_name == "sts-b":
text_a, text_b, label = line[7], line[8], line[-1]
elif task_name == "qqp":
try:
text_a, text_b, label = line[3], line[4], line[5]
except IndexError:
text_a, text_b, label = "", "", "1"
elif task_name == "qnli":
text_a, text_b, label = line[1], line[2], line[-1]
elif task_name == "rte":
text_a, text_b, label = line[1], line[2], line[-1]
elif task_name == "wnli":
text_a, text_b, label = line[1], line[2], line[-1]
else:
raise NotImplementedError
encoding = self.tokenizer(text_a, text_b, padding='max_length', truncation=True,
max_length=self.max_seq_length)
encoding['label_ids'] = self.label_map[label]
return encoding
[docs] def batch_fn(self, features):
return {k: torch.tensor([dic[k] for dic in features], dtype=torch.long) for k in features[0]}
[docs]class CNNClassificationDataset(BaseDataset):
def __init__(self,
pretrained_model_name_or_path,
data_file,
max_seq_length,
input_schema,
first_sequence,
label_name=None,
second_sequence=None,
label_enumerate_values=None,
multi_label=False,
*args,
**kwargs):
super(CNNClassificationDataset, self).__init__(data_file, input_schema=input_schema,
output_format="dict", *args, **kwargs)
self.tokenizer = TextCNNTokenizer.from_pretrained(pretrained_model_name_or_path)
self.max_seq_length = max_seq_length
self.multi_label = multi_label
if label_enumerate_values is None:
self._label_enumerate_values = "0,1".split(",")
else:
if io.exists(label_enumerate_values):
with io.open(label_enumerate_values) as f:
self._label_enumerate_values = [line.strip() for line in f]
else:
self._label_enumerate_values = label_enumerate_values.split(",")
self.max_num_labels = len(self._label_enumerate_values)
assert first_sequence in self.column_names, \
"Column name %s needs to be included in columns" % first_sequence
self.first_sequence = first_sequence
if second_sequence:
assert second_sequence in self.column_names, \
"Column name %s needs to be included in columns" % second_sequence
self.second_sequence = second_sequence
else:
self.second_sequence = None
if label_name:
assert label_name in self.column_names, \
"Column name %s needs to be included in columns" % label_name
self.label_name = label_name
else:
self.label_name = None
self.label_map = dict({value: idx for idx, value in enumerate(self.label_enumerate_values)})
@property
def eval_metrics(self):
"""
Returns the evaluation metrics.
"""
return ("accuracy", "f1")
@property
def label_enumerate_values(self):
"""
Returns the label enumerate values.
"""
return self._label_enumerate_values
[docs] def convert_single_row_to_example(self, row):
"""Convert sample token to indices.
Args:
row: contains sequence and label.
text_a: the first sequence in row.
text_b: the second sequence in row if self.second_sequence is true.
label: label token if self.label_name is true.
Returns: sing example
encoding: an example contains token indices.
"""
text_a = row[self.first_sequence]
text_b = row[self.second_sequence] if self.second_sequence else None
label = row[self.label_name] if self.label_name else None
encoding = self.tokenizer(text_a, text_b, padding='max_length', truncation=True,
max_length=self.max_seq_length)
if not self.multi_label:
encoding['label_ids'] = self.label_map[label]
return encoding
else:
label_id = [self.label_map[x] for x in label.split(",") if x]
new_label_id = [0] * self.max_num_labels
for idx in label_id:
new_label_id[idx] = 1
encoding['label_ids'] = new_label_id
return encoding
#
# def batch_fn(self, features):
# return {k: torch.tensor([dic[k] for dic in features], dtype=torch.long) for k in features[0]}
#
[docs] def batch_fn(self, features):
"""
Divide examples into batches.
"""
inputs = {
"input_ids": torch.tensor([f['input_ids'] for f in features], dtype=torch.long),
"label_ids": torch.tensor([f['label_ids'] for f in features], dtype=torch.long)
}
return inputs