Upload run_classification.py
Browse files- run_classification.py +763 -0
run_classification.py
ADDED
@@ -0,0 +1,763 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
# Copyright 2020 The HuggingFace Inc. team. All rights reserved.
|
4 |
+
#
|
5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
+
# you may not use this file except in compliance with the License.
|
7 |
+
# You may obtain a copy of the License at
|
8 |
+
#
|
9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
10 |
+
#
|
11 |
+
# Unless required by applicable law or agreed to in writing, software
|
12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
14 |
+
# See the License for the specific language governing permissions and
|
15 |
+
# limitations under the License.
|
16 |
+
""" Finetuning the library models for text classification."""
|
17 |
+
# You can also adapt this script on your own text classification task. Pointers for this are left as comments.
|
18 |
+
|
19 |
+
import logging
|
20 |
+
import os
|
21 |
+
import random
|
22 |
+
import sys
|
23 |
+
import warnings
|
24 |
+
from dataclasses import dataclass, field
|
25 |
+
from typing import List, Optional
|
26 |
+
|
27 |
+
import datasets
|
28 |
+
import evaluate
|
29 |
+
import numpy as np
|
30 |
+
from datasets import Value, load_dataset
|
31 |
+
|
32 |
+
import transformers
|
33 |
+
from transformers import (
|
34 |
+
AutoConfig,
|
35 |
+
AutoModelForSequenceClassification,
|
36 |
+
AutoTokenizer,
|
37 |
+
DataCollatorWithPadding,
|
38 |
+
EvalPrediction,
|
39 |
+
HfArgumentParser,
|
40 |
+
Trainer,
|
41 |
+
TrainingArguments,
|
42 |
+
default_data_collator,
|
43 |
+
set_seed,
|
44 |
+
)
|
45 |
+
from transformers.trainer_utils import get_last_checkpoint
|
46 |
+
from transformers.utils import check_min_version, send_example_telemetry
|
47 |
+
from transformers.utils.versions import require_version
|
48 |
+
|
49 |
+
|
50 |
+
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
|
51 |
+
# check_min_version("4.38.0.dev0")
|
52 |
+
|
53 |
+
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
|
54 |
+
|
55 |
+
|
56 |
+
logger = logging.getLogger(__name__)
|
57 |
+
|
58 |
+
|
59 |
+
@dataclass
|
60 |
+
class DataTrainingArguments:
|
61 |
+
"""
|
62 |
+
Arguments pertaining to what data we are going to input our model for training and eval.
|
63 |
+
|
64 |
+
Using `HfArgumentParser` we can turn this class
|
65 |
+
into argparse arguments to be able to specify them on
|
66 |
+
the command line.
|
67 |
+
"""
|
68 |
+
|
69 |
+
dataset_name: Optional[str] = field(
|
70 |
+
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
|
71 |
+
)
|
72 |
+
dataset_config_name: Optional[str] = field(
|
73 |
+
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
74 |
+
)
|
75 |
+
do_regression: bool = field(
|
76 |
+
default=None,
|
77 |
+
metadata={
|
78 |
+
"help": "Whether to do regression instead of classification. If None, will be inferred from the dataset."
|
79 |
+
},
|
80 |
+
)
|
81 |
+
text_column_names: Optional[str] = field(
|
82 |
+
default=None,
|
83 |
+
metadata={
|
84 |
+
"help": (
|
85 |
+
"The name of the text column in the input dataset or a CSV/JSON file. "
|
86 |
+
'If not specified, will use the "sentence" column for single/multi-label classification task.'
|
87 |
+
)
|
88 |
+
},
|
89 |
+
)
|
90 |
+
text_column_delimiter: Optional[str] = field(
|
91 |
+
default=" ", metadata={"help": "THe delimiter to use to join text columns into a single sentence."}
|
92 |
+
)
|
93 |
+
train_split_name: Optional[str] = field(
|
94 |
+
default=None,
|
95 |
+
metadata={
|
96 |
+
"help": 'The name of the train split in the input dataset. If not specified, will use the "train" split when do_train is enabled'
|
97 |
+
},
|
98 |
+
)
|
99 |
+
validation_split_name: Optional[str] = field(
|
100 |
+
default=None,
|
101 |
+
metadata={
|
102 |
+
"help": 'The name of the validation split in the input dataset. If not specified, will use the "validation" split when do_eval is enabled'
|
103 |
+
},
|
104 |
+
)
|
105 |
+
test_split_name: Optional[str] = field(
|
106 |
+
default=None,
|
107 |
+
metadata={
|
108 |
+
"help": 'The name of the test split in the input dataset. If not specified, will use the "test" split when do_predict is enabled'
|
109 |
+
},
|
110 |
+
)
|
111 |
+
remove_splits: Optional[str] = field(
|
112 |
+
default=None,
|
113 |
+
metadata={"help": "The splits to remove from the dataset. Multiple splits should be separated by commas."},
|
114 |
+
)
|
115 |
+
remove_columns: Optional[str] = field(
|
116 |
+
default=None,
|
117 |
+
metadata={"help": "The columns to remove from the dataset. Multiple columns should be separated by commas."},
|
118 |
+
)
|
119 |
+
label_column_name: Optional[str] = field(
|
120 |
+
default=None,
|
121 |
+
metadata={
|
122 |
+
"help": (
|
123 |
+
"The name of the label column in the input dataset or a CSV/JSON file. "
|
124 |
+
'If not specified, will use the "label" column for single/multi-label classification task'
|
125 |
+
)
|
126 |
+
},
|
127 |
+
)
|
128 |
+
max_seq_length: int = field(
|
129 |
+
default=128,
|
130 |
+
metadata={
|
131 |
+
"help": (
|
132 |
+
"The maximum total input sequence length after tokenization. Sequences longer "
|
133 |
+
"than this will be truncated, sequences shorter will be padded."
|
134 |
+
)
|
135 |
+
},
|
136 |
+
)
|
137 |
+
overwrite_cache: bool = field(
|
138 |
+
default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
|
139 |
+
)
|
140 |
+
pad_to_max_length: bool = field(
|
141 |
+
default=True,
|
142 |
+
metadata={
|
143 |
+
"help": (
|
144 |
+
"Whether to pad all samples to `max_seq_length`. "
|
145 |
+
"If False, will pad the samples dynamically when batching to the maximum length in the batch."
|
146 |
+
)
|
147 |
+
},
|
148 |
+
)
|
149 |
+
shuffle_train_dataset: bool = field(
|
150 |
+
default=False, metadata={"help": "Whether to shuffle the train dataset or not."}
|
151 |
+
)
|
152 |
+
shuffle_seed: int = field(
|
153 |
+
default=42, metadata={"help": "Random seed that will be used to shuffle the train dataset."}
|
154 |
+
)
|
155 |
+
max_train_samples: Optional[int] = field(
|
156 |
+
default=None,
|
157 |
+
metadata={
|
158 |
+
"help": (
|
159 |
+
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
160 |
+
"value if set."
|
161 |
+
)
|
162 |
+
},
|
163 |
+
)
|
164 |
+
max_eval_samples: Optional[int] = field(
|
165 |
+
default=None,
|
166 |
+
metadata={
|
167 |
+
"help": (
|
168 |
+
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
|
169 |
+
"value if set."
|
170 |
+
)
|
171 |
+
},
|
172 |
+
)
|
173 |
+
max_predict_samples: Optional[int] = field(
|
174 |
+
default=None,
|
175 |
+
metadata={
|
176 |
+
"help": (
|
177 |
+
"For debugging purposes or quicker training, truncate the number of prediction examples to this "
|
178 |
+
"value if set."
|
179 |
+
)
|
180 |
+
},
|
181 |
+
)
|
182 |
+
metric_name: Optional[str] = field(default=None, metadata={"help": "The metric to use for evaluation."})
|
183 |
+
train_file: Optional[str] = field(
|
184 |
+
default=None, metadata={"help": "A csv or a json file containing the training data."}
|
185 |
+
)
|
186 |
+
validation_file: Optional[str] = field(
|
187 |
+
default=None, metadata={"help": "A csv or a json file containing the validation data."}
|
188 |
+
)
|
189 |
+
test_file: Optional[str] = field(default=None, metadata={"help": "A csv or a json file containing the test data."})
|
190 |
+
|
191 |
+
def __post_init__(self):
|
192 |
+
if self.dataset_name is None:
|
193 |
+
if self.train_file is None or self.validation_file is None:
|
194 |
+
raise ValueError(" training/validation file or a dataset name.")
|
195 |
+
|
196 |
+
train_extension = self.train_file.split(".")[-1]
|
197 |
+
assert train_extension in ["csv", "json"], "`train_file` should be a csv or a json file."
|
198 |
+
validation_extension = self.validation_file.split(".")[-1]
|
199 |
+
assert (
|
200 |
+
validation_extension == train_extension
|
201 |
+
), "`validation_file` should have the same extension (csv or json) as `train_file`."
|
202 |
+
|
203 |
+
|
204 |
+
@dataclass
|
205 |
+
class ModelArguments:
|
206 |
+
"""
|
207 |
+
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
|
208 |
+
"""
|
209 |
+
|
210 |
+
model_name_or_path: str = field(
|
211 |
+
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
|
212 |
+
)
|
213 |
+
config_name: Optional[str] = field(
|
214 |
+
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
|
215 |
+
)
|
216 |
+
tokenizer_name: Optional[str] = field(
|
217 |
+
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
|
218 |
+
)
|
219 |
+
cache_dir: Optional[str] = field(
|
220 |
+
default=None,
|
221 |
+
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
|
222 |
+
)
|
223 |
+
use_fast_tokenizer: bool = field(
|
224 |
+
default=True,
|
225 |
+
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
|
226 |
+
)
|
227 |
+
model_revision: str = field(
|
228 |
+
default="main",
|
229 |
+
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
|
230 |
+
)
|
231 |
+
token: str = field(
|
232 |
+
default=None,
|
233 |
+
metadata={
|
234 |
+
"help": (
|
235 |
+
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
|
236 |
+
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
|
237 |
+
)
|
238 |
+
},
|
239 |
+
)
|
240 |
+
use_auth_token: bool = field(
|
241 |
+
default=None,
|
242 |
+
metadata={
|
243 |
+
"help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead."
|
244 |
+
},
|
245 |
+
)
|
246 |
+
trust_remote_code: bool = field(
|
247 |
+
default=False,
|
248 |
+
metadata={
|
249 |
+
"help": (
|
250 |
+
"Whether or not to allow for custom models defined on the Hub in their own modeling files. This option "
|
251 |
+
"should only be set to `True` for repositories you trust and in which you have read the code, as it will "
|
252 |
+
"execute code present on the Hub on your local machine."
|
253 |
+
)
|
254 |
+
},
|
255 |
+
)
|
256 |
+
ignore_mismatched_sizes: bool = field(
|
257 |
+
default=False,
|
258 |
+
metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."},
|
259 |
+
)
|
260 |
+
|
261 |
+
|
262 |
+
def get_label_list(raw_dataset, split="train") -> List[str]:
|
263 |
+
"""Get the list of labels from a multi-label dataset"""
|
264 |
+
|
265 |
+
if isinstance(raw_dataset[split]["label"][0], list):
|
266 |
+
label_list = [label for sample in raw_dataset[split]["label"] for label in sample]
|
267 |
+
label_list = list(set(label_list))
|
268 |
+
else:
|
269 |
+
label_list = raw_dataset[split].unique("label")
|
270 |
+
# we will treat the label list as a list of string instead of int, consistent with model.config.label2id
|
271 |
+
label_list = [str(label) for label in label_list]
|
272 |
+
return label_list
|
273 |
+
|
274 |
+
|
275 |
+
def main():
|
276 |
+
# See all possible arguments in src/transformers/training_args.py
|
277 |
+
# or by passing the --help flag to this script.
|
278 |
+
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
279 |
+
|
280 |
+
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
|
281 |
+
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
282 |
+
# If we pass only one argument to the script and it's the path to a json file,
|
283 |
+
# let's parse it to get our arguments.
|
284 |
+
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
|
285 |
+
else:
|
286 |
+
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
287 |
+
|
288 |
+
if model_args.use_auth_token is not None:
|
289 |
+
warnings.warn(
|
290 |
+
"The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.",
|
291 |
+
FutureWarning,
|
292 |
+
)
|
293 |
+
if model_args.token is not None:
|
294 |
+
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
|
295 |
+
model_args.token = model_args.use_auth_token
|
296 |
+
|
297 |
+
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
|
298 |
+
# information sent is the one passed as arguments along with your Python/PyTorch versions.
|
299 |
+
# send_example_telemetry("run_classification", model_args, data_args)
|
300 |
+
|
301 |
+
# Setup logging
|
302 |
+
logging.basicConfig(
|
303 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
304 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
305 |
+
handlers=[logging.StreamHandler(sys.stdout)],
|
306 |
+
)
|
307 |
+
|
308 |
+
if training_args.should_log:
|
309 |
+
# The default of training_args.log_level is passive, so we set log level at info here to have that default.
|
310 |
+
transformers.utils.logging.set_verbosity_info()
|
311 |
+
|
312 |
+
log_level = training_args.get_process_log_level()
|
313 |
+
logger.setLevel(log_level)
|
314 |
+
datasets.utils.logging.set_verbosity(log_level)
|
315 |
+
transformers.utils.logging.set_verbosity(log_level)
|
316 |
+
transformers.utils.logging.enable_default_handler()
|
317 |
+
transformers.utils.logging.enable_explicit_format()
|
318 |
+
|
319 |
+
# Log on each process the small summary:
|
320 |
+
logger.warning(
|
321 |
+
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
|
322 |
+
+ f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
|
323 |
+
)
|
324 |
+
logger.info(f"Training/evaluation parameters {training_args}")
|
325 |
+
|
326 |
+
# Detecting last checkpoint.
|
327 |
+
last_checkpoint = None
|
328 |
+
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
|
329 |
+
last_checkpoint = get_last_checkpoint(training_args.output_dir)
|
330 |
+
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
|
331 |
+
raise ValueError(
|
332 |
+
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
|
333 |
+
"Use --overwrite_output_dir to overcome."
|
334 |
+
)
|
335 |
+
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
|
336 |
+
logger.info(
|
337 |
+
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
|
338 |
+
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
|
339 |
+
)
|
340 |
+
|
341 |
+
# Set seed before initializing model.
|
342 |
+
set_seed(training_args.seed)
|
343 |
+
|
344 |
+
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files, or specify a dataset name
|
345 |
+
# to load from huggingface/datasets. In ether case, you can specify a the key of the column(s) containing the text and
|
346 |
+
# the key of the column containing the label. If multiple columns are specified for the text, they will be joined together
|
347 |
+
# for the actual text value.
|
348 |
+
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
|
349 |
+
# download the dataset.
|
350 |
+
if data_args.dataset_name is not None:
|
351 |
+
# Downloading and loading a dataset from the hub.
|
352 |
+
raw_datasets = load_dataset(
|
353 |
+
data_args.dataset_name,
|
354 |
+
data_args.dataset_config_name,
|
355 |
+
cache_dir=model_args.cache_dir,
|
356 |
+
token=model_args.token,
|
357 |
+
)
|
358 |
+
# Try print some info about the dataset
|
359 |
+
logger.info(f"Dataset loaded: {raw_datasets}")
|
360 |
+
logger.info(raw_datasets)
|
361 |
+
else:
|
362 |
+
# Loading a dataset from your local files.
|
363 |
+
# CSV/JSON training and evaluation files are needed.
|
364 |
+
data_files = {"train": data_args.train_file, "validation": data_args.validation_file}
|
365 |
+
|
366 |
+
# Get the test dataset: you can provide your own CSV/JSON test file
|
367 |
+
if training_args.do_predict:
|
368 |
+
if data_args.test_file is not None:
|
369 |
+
train_extension = data_args.train_file.split(".")[-1]
|
370 |
+
test_extension = data_args.test_file.split(".")[-1]
|
371 |
+
assert (
|
372 |
+
test_extension == train_extension
|
373 |
+
), "`test_file` should have the same extension (csv or json) as `train_file`."
|
374 |
+
data_files["test"] = data_args.test_file
|
375 |
+
else:
|
376 |
+
raise ValueError("Need either a dataset name or a test file for `do_predict`.")
|
377 |
+
|
378 |
+
for key in data_files.keys():
|
379 |
+
logger.info(f"load a local file for {key}: {data_files[key]}")
|
380 |
+
|
381 |
+
if data_args.train_file.endswith(".csv"):
|
382 |
+
# Loading a dataset from local csv files
|
383 |
+
raw_datasets = load_dataset(
|
384 |
+
"csv",
|
385 |
+
data_files=data_files,
|
386 |
+
cache_dir=model_args.cache_dir,
|
387 |
+
token=model_args.token,
|
388 |
+
)
|
389 |
+
else:
|
390 |
+
# Loading a dataset from local json files
|
391 |
+
raw_datasets = load_dataset(
|
392 |
+
"json",
|
393 |
+
data_files=data_files,
|
394 |
+
cache_dir=model_args.cache_dir,
|
395 |
+
token=model_args.token,
|
396 |
+
)
|
397 |
+
|
398 |
+
# See more about loading any type of standard or custom dataset at
|
399 |
+
# https://huggingface.co/docs/datasets/loading_datasets.
|
400 |
+
|
401 |
+
if data_args.remove_splits is not None:
|
402 |
+
for split in data_args.remove_splits.split(","):
|
403 |
+
logger.info(f"removing split {split}")
|
404 |
+
raw_datasets.pop(split)
|
405 |
+
|
406 |
+
if data_args.train_split_name is not None:
|
407 |
+
logger.info(f"using {data_args.train_split_name} as train set")
|
408 |
+
raw_datasets["train"] = raw_datasets[data_args.train_split_name]
|
409 |
+
raw_datasets.pop(data_args.train_split_name)
|
410 |
+
|
411 |
+
if data_args.validation_split_name is not None:
|
412 |
+
logger.info(f"using {data_args.validation_split_name} as validation set")
|
413 |
+
raw_datasets["validation"] = raw_datasets[data_args.validation_split_name]
|
414 |
+
raw_datasets.pop(data_args.validation_split_name)
|
415 |
+
|
416 |
+
if data_args.test_split_name is not None:
|
417 |
+
logger.info(f"using {data_args.test_split_name} as test set")
|
418 |
+
raw_datasets["test"] = raw_datasets[data_args.test_split_name]
|
419 |
+
raw_datasets.pop(data_args.test_split_name)
|
420 |
+
|
421 |
+
if data_args.remove_columns is not None:
|
422 |
+
for split in raw_datasets.keys():
|
423 |
+
for column in data_args.remove_columns.split(","):
|
424 |
+
logger.info(f"removing column {column} from split {split}")
|
425 |
+
raw_datasets[split].remove_columns(column)
|
426 |
+
|
427 |
+
if data_args.label_column_name is not None and data_args.label_column_name != "label":
|
428 |
+
for key in raw_datasets.keys():
|
429 |
+
raw_datasets[key] = raw_datasets[key].rename_column(data_args.label_column_name, "label")
|
430 |
+
|
431 |
+
# Trying to have good defaults here, don't hesitate to tweak to your needs.
|
432 |
+
|
433 |
+
is_regression = (
|
434 |
+
raw_datasets["train"].features["label"].dtype in ["float32", "float64"]
|
435 |
+
if data_args.do_regression is None
|
436 |
+
else data_args.do_regression
|
437 |
+
)
|
438 |
+
|
439 |
+
is_multi_label = False
|
440 |
+
if is_regression:
|
441 |
+
label_list = None
|
442 |
+
num_labels = 1
|
443 |
+
# regession requires float as label type, let's cast it if needed
|
444 |
+
for split in raw_datasets.keys():
|
445 |
+
if raw_datasets[split].features["label"].dtype not in ["float32", "float64"]:
|
446 |
+
logger.warning(
|
447 |
+
f"Label type for {split} set to float32, was {raw_datasets[split].features['label'].dtype}"
|
448 |
+
)
|
449 |
+
features = raw_datasets[split].features
|
450 |
+
features.update({"label": Value("float32")})
|
451 |
+
try:
|
452 |
+
raw_datasets[split] = raw_datasets[split].cast(features)
|
453 |
+
except TypeError as error:
|
454 |
+
logger.error(
|
455 |
+
f"Unable to cast {split} set to float32, please check the labels are correct, or maybe try with --do_regression=False"
|
456 |
+
)
|
457 |
+
raise error
|
458 |
+
|
459 |
+
else: # classification
|
460 |
+
if raw_datasets["train"].features["label"].dtype == "list": # multi-label classification
|
461 |
+
is_multi_label = True
|
462 |
+
logger.info("Label type is list, doing multi-label classification")
|
463 |
+
# Trying to find the number of labels in a multi-label classification task
|
464 |
+
# We have to deal with common cases that labels appear in the training set but not in the validation/test set.
|
465 |
+
# So we build the label list from the union of labels in train/val/test.
|
466 |
+
label_list = get_label_list(raw_datasets, split="train")
|
467 |
+
for split in ["validation", "test"]:
|
468 |
+
if split in raw_datasets:
|
469 |
+
val_or_test_labels = get_label_list(raw_datasets, split=split)
|
470 |
+
diff = set(val_or_test_labels).difference(set(label_list))
|
471 |
+
if len(diff) > 0:
|
472 |
+
# add the labels that appear in val/test but not in train, throw a warning
|
473 |
+
logger.warning(
|
474 |
+
f"Labels {diff} in {split} set but not in training set, adding them to the label list"
|
475 |
+
)
|
476 |
+
label_list += list(diff)
|
477 |
+
# if label is -1, we throw a warning and remove it from the label list
|
478 |
+
for label in label_list:
|
479 |
+
if label == -1:
|
480 |
+
logger.warning("Label -1 found in label list, removing it.")
|
481 |
+
label_list.remove(label)
|
482 |
+
|
483 |
+
label_list.sort()
|
484 |
+
num_labels = len(label_list)
|
485 |
+
if num_labels <= 1:
|
486 |
+
raise ValueError("You need more than one label to do classification.")
|
487 |
+
|
488 |
+
# Load pretrained model and tokenizer
|
489 |
+
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
|
490 |
+
# download model & vocab.
|
491 |
+
config = AutoConfig.from_pretrained(
|
492 |
+
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
|
493 |
+
num_labels=num_labels,
|
494 |
+
finetuning_task="text-classification",
|
495 |
+
cache_dir=model_args.cache_dir,
|
496 |
+
revision=model_args.model_revision,
|
497 |
+
token=model_args.token,
|
498 |
+
trust_remote_code=model_args.trust_remote_code,
|
499 |
+
)
|
500 |
+
|
501 |
+
if is_regression:
|
502 |
+
config.problem_type = "regression"
|
503 |
+
logger.info("setting problem type to regression")
|
504 |
+
elif is_multi_label:
|
505 |
+
config.problem_type = "multi_label_classification"
|
506 |
+
logger.info("setting problem type to multi label classification")
|
507 |
+
else:
|
508 |
+
config.problem_type = "single_label_classification"
|
509 |
+
logger.info("setting problem type to single label classification")
|
510 |
+
|
511 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
512 |
+
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
|
513 |
+
cache_dir=model_args.cache_dir,
|
514 |
+
use_fast=model_args.use_fast_tokenizer,
|
515 |
+
revision=model_args.model_revision,
|
516 |
+
token=model_args.token,
|
517 |
+
trust_remote_code=model_args.trust_remote_code,
|
518 |
+
)
|
519 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
520 |
+
model_args.model_name_or_path,
|
521 |
+
from_tf=bool(".ckpt" in model_args.model_name_or_path),
|
522 |
+
config=config,
|
523 |
+
cache_dir=model_args.cache_dir,
|
524 |
+
revision=model_args.model_revision,
|
525 |
+
token=model_args.token,
|
526 |
+
trust_remote_code=model_args.trust_remote_code,
|
527 |
+
ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
|
528 |
+
)
|
529 |
+
|
530 |
+
# Padding strategy
|
531 |
+
if data_args.pad_to_max_length:
|
532 |
+
padding = "max_length"
|
533 |
+
else:
|
534 |
+
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
|
535 |
+
padding = False
|
536 |
+
|
537 |
+
# for training ,we will update the config with label infos,
|
538 |
+
# if do_train is not set, we will use the label infos in the config
|
539 |
+
if training_args.do_train and not is_regression: # classification, training
|
540 |
+
label_to_id = {v: i for i, v in enumerate(label_list)}
|
541 |
+
# update config with label infos
|
542 |
+
if model.config.label2id != label_to_id:
|
543 |
+
logger.warning(
|
544 |
+
"The label2id key in the model config.json is not equal to the label2id key of this "
|
545 |
+
"run. You can ignore this if you are doing finetuning."
|
546 |
+
)
|
547 |
+
model.config.label2id = label_to_id
|
548 |
+
model.config.id2label = {id: label for label, id in label_to_id.items()}
|
549 |
+
elif not is_regression: # classification, but not training
|
550 |
+
logger.info("using label infos in the model config")
|
551 |
+
logger.info("label2id: {}".format(model.config.label2id))
|
552 |
+
label_to_id = model.config.label2id
|
553 |
+
else: # regression
|
554 |
+
label_to_id = None
|
555 |
+
|
556 |
+
if data_args.max_seq_length > tokenizer.model_max_length:
|
557 |
+
logger.warning(
|
558 |
+
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the "
|
559 |
+
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
|
560 |
+
)
|
561 |
+
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
|
562 |
+
|
563 |
+
def multi_labels_to_ids(labels: List[str]) -> List[float]:
|
564 |
+
ids = [0.0] * len(label_to_id) # BCELoss requires float as target type
|
565 |
+
for label in labels:
|
566 |
+
ids[label_to_id[label]] = 1.0
|
567 |
+
return ids
|
568 |
+
|
569 |
+
def preprocess_function(examples):
|
570 |
+
if data_args.text_column_names is not None:
|
571 |
+
text_column_names = data_args.text_column_names.split(",")
|
572 |
+
# join together text columns into "sentence" column
|
573 |
+
examples["sentence"] = examples[text_column_names[0]]
|
574 |
+
for column in text_column_names[1:]:
|
575 |
+
for i in range(len(examples[column])):
|
576 |
+
examples["sentence"][i] += data_args.text_column_delimiter + examples[column][i]
|
577 |
+
# Tokenize the texts
|
578 |
+
result = tokenizer(examples["sentence"], padding=padding, max_length=max_seq_length, truncation=True)
|
579 |
+
if label_to_id is not None and "label" in examples:
|
580 |
+
if is_multi_label:
|
581 |
+
result["label"] = [multi_labels_to_ids(l) for l in examples["label"]]
|
582 |
+
else:
|
583 |
+
result["label"] = [(label_to_id[str(l)] if l != -1 else -1) for l in examples["label"]]
|
584 |
+
return result
|
585 |
+
|
586 |
+
# Running the preprocessing pipeline on all the datasets
|
587 |
+
with training_args.main_process_first(desc="dataset map pre-processing"):
|
588 |
+
raw_datasets = raw_datasets.map(
|
589 |
+
preprocess_function,
|
590 |
+
batched=True,
|
591 |
+
load_from_cache_file=not data_args.overwrite_cache,
|
592 |
+
desc="Running tokenizer on dataset",
|
593 |
+
)
|
594 |
+
|
595 |
+
if training_args.do_train:
|
596 |
+
if "train" not in raw_datasets:
|
597 |
+
raise ValueError("--do_train requires a train dataset.")
|
598 |
+
train_dataset = raw_datasets["train"]
|
599 |
+
if data_args.shuffle_train_dataset:
|
600 |
+
logger.info("Shuffling the training dataset")
|
601 |
+
train_dataset = train_dataset.shuffle(seed=data_args.shuffle_seed)
|
602 |
+
if data_args.max_train_samples is not None:
|
603 |
+
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
|
604 |
+
train_dataset = train_dataset.select(range(max_train_samples))
|
605 |
+
|
606 |
+
if training_args.do_eval:
|
607 |
+
if "validation" not in raw_datasets and "validation_matched" not in raw_datasets:
|
608 |
+
if "test" not in raw_datasets and "test_matched" not in raw_datasets:
|
609 |
+
raise ValueError("--do_eval requires a validation or test dataset if validation is not defined.")
|
610 |
+
else:
|
611 |
+
logger.warning("Validation dataset not found. Falling back to test dataset for validation.")
|
612 |
+
eval_dataset = raw_datasets["test"]
|
613 |
+
else:
|
614 |
+
eval_dataset = raw_datasets["validation"]
|
615 |
+
|
616 |
+
if data_args.max_eval_samples is not None:
|
617 |
+
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
|
618 |
+
eval_dataset = eval_dataset.select(range(max_eval_samples))
|
619 |
+
|
620 |
+
if training_args.do_predict or data_args.test_file is not None:
|
621 |
+
if "test" not in raw_datasets:
|
622 |
+
raise ValueError("--do_predict requires a test dataset")
|
623 |
+
predict_dataset = raw_datasets["test"]
|
624 |
+
# remove label column if it exists
|
625 |
+
if data_args.max_predict_samples is not None:
|
626 |
+
max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
|
627 |
+
predict_dataset = predict_dataset.select(range(max_predict_samples))
|
628 |
+
|
629 |
+
# Log a few random samples from the training set:
|
630 |
+
if training_args.do_train:
|
631 |
+
for index in random.sample(range(len(train_dataset)), 3):
|
632 |
+
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
|
633 |
+
|
634 |
+
if data_args.metric_name is not None:
|
635 |
+
metric = (
|
636 |
+
evaluate.load(data_args.metric_name, config_name="multilabel", cache_dir=model_args.cache_dir)
|
637 |
+
if is_multi_label
|
638 |
+
else evaluate.load(data_args.metric_name, cache_dir=model_args.cache_dir)
|
639 |
+
)
|
640 |
+
logger.info(f"Using metric {data_args.metric_name} for evaluation.")
|
641 |
+
else:
|
642 |
+
if is_regression:
|
643 |
+
metric = evaluate.load("mse", cache_dir=model_args.cache_dir)
|
644 |
+
logger.info("Using mean squared error (mse) as regression score, you can use --metric_name to overwrite.")
|
645 |
+
else:
|
646 |
+
if is_multi_label:
|
647 |
+
metric = evaluate.load("f1", config_name="multilabel", cache_dir=model_args.cache_dir)
|
648 |
+
logger.info(
|
649 |
+
"Using multilabel F1 for multi-label classification task, you can use --metric_name to overwrite."
|
650 |
+
)
|
651 |
+
else:
|
652 |
+
metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
|
653 |
+
logger.info("Using accuracy as classification score, you can use --metric_name to overwrite.")
|
654 |
+
|
655 |
+
def compute_metrics(p: EvalPrediction):
|
656 |
+
preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
|
657 |
+
if is_regression:
|
658 |
+
preds = np.squeeze(preds)
|
659 |
+
result = metric.compute(predictions=preds, references=p.label_ids)
|
660 |
+
elif is_multi_label:
|
661 |
+
preds = np.array([np.where(p > 0, 1, 0) for p in preds]) # convert logits to multi-hot encoding
|
662 |
+
# Micro F1 is commonly used in multi-label classification
|
663 |
+
result = metric.compute(predictions=preds, references=p.label_ids, average="micro")
|
664 |
+
else:
|
665 |
+
preds = np.argmax(preds, axis=1)
|
666 |
+
result = metric.compute(predictions=preds, references=p.label_ids)
|
667 |
+
if len(result) > 1:
|
668 |
+
result["combined_score"] = np.mean(list(result.values())).item()
|
669 |
+
return result
|
670 |
+
|
671 |
+
# Data collator will default to DataCollatorWithPadding when the tokenizer is passed to Trainer, so we change it if
|
672 |
+
# we already did the padding.
|
673 |
+
if data_args.pad_to_max_length:
|
674 |
+
data_collator = default_data_collator
|
675 |
+
elif training_args.fp16:
|
676 |
+
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
|
677 |
+
else:
|
678 |
+
data_collator = None
|
679 |
+
|
680 |
+
# Initialize our Trainer
|
681 |
+
trainer = Trainer(
|
682 |
+
model=model,
|
683 |
+
args=training_args,
|
684 |
+
train_dataset=train_dataset if training_args.do_train else None,
|
685 |
+
eval_dataset=eval_dataset if training_args.do_eval else None,
|
686 |
+
compute_metrics=compute_metrics,
|
687 |
+
tokenizer=tokenizer,
|
688 |
+
data_collator=data_collator,
|
689 |
+
)
|
690 |
+
|
691 |
+
# Training
|
692 |
+
if training_args.do_train:
|
693 |
+
checkpoint = None
|
694 |
+
if training_args.resume_from_checkpoint is not None:
|
695 |
+
checkpoint = training_args.resume_from_checkpoint
|
696 |
+
elif last_checkpoint is not None:
|
697 |
+
checkpoint = last_checkpoint
|
698 |
+
train_result = trainer.train(resume_from_checkpoint=checkpoint)
|
699 |
+
metrics = train_result.metrics
|
700 |
+
max_train_samples = (
|
701 |
+
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
|
702 |
+
)
|
703 |
+
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
|
704 |
+
trainer.save_model() # Saves the tokenizer too for easy upload
|
705 |
+
trainer.log_metrics("train", metrics)
|
706 |
+
trainer.save_metrics("train", metrics)
|
707 |
+
trainer.save_state()
|
708 |
+
|
709 |
+
# Evaluation
|
710 |
+
if training_args.do_eval:
|
711 |
+
logger.info("*** Evaluate ***")
|
712 |
+
metrics = trainer.evaluate(eval_dataset=eval_dataset)
|
713 |
+
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
|
714 |
+
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
|
715 |
+
trainer.log_metrics("eval", metrics)
|
716 |
+
trainer.save_metrics("eval", metrics)
|
717 |
+
|
718 |
+
if training_args.do_predict:
|
719 |
+
logger.info("*** Predict ***")
|
720 |
+
# Removing the `label` columns if exists because it might contains -1 and Trainer won't like that.
|
721 |
+
if "label" in predict_dataset.features:
|
722 |
+
predict_dataset = predict_dataset.remove_columns("label")
|
723 |
+
predictions = trainer.predict(predict_dataset, metric_key_prefix="predict").predictions
|
724 |
+
if is_regression:
|
725 |
+
predictions = np.squeeze(predictions)
|
726 |
+
elif is_multi_label:
|
727 |
+
# Convert logits to multi-hot encoding. We compare the logits to 0 instead of 0.5, because the sigmoid is not applied.
|
728 |
+
# You can also pass `preprocess_logits_for_metrics=lambda logits, labels: nn.functional.sigmoid(logits)` to the Trainer
|
729 |
+
# and set p > 0.5 below (less efficient in this case)
|
730 |
+
predictions = np.array([np.where(p > 0, 1, 0) for p in predictions])
|
731 |
+
else:
|
732 |
+
predictions = np.argmax(predictions, axis=1)
|
733 |
+
output_predict_file = os.path.join(training_args.output_dir, "predict_results.txt")
|
734 |
+
if trainer.is_world_process_zero():
|
735 |
+
with open(output_predict_file, "w") as writer:
|
736 |
+
logger.info("***** Predict results *****")
|
737 |
+
writer.write("index\tprediction\n")
|
738 |
+
for index, item in enumerate(predictions):
|
739 |
+
if is_regression:
|
740 |
+
writer.write(f"{index}\t{item:3.3f}\n")
|
741 |
+
elif is_multi_label:
|
742 |
+
# recover from multi-hot encoding
|
743 |
+
item = [label_list[i] for i in range(len(item)) if item[i] == 1]
|
744 |
+
writer.write(f"{index}\t{item}\n")
|
745 |
+
else:
|
746 |
+
item = label_list[item]
|
747 |
+
writer.write(f"{index}\t{item}\n")
|
748 |
+
logger.info("Predict results saved at {}".format(output_predict_file))
|
749 |
+
kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "text-classification"}
|
750 |
+
|
751 |
+
if training_args.push_to_hub:
|
752 |
+
trainer.push_to_hub(**kwargs)
|
753 |
+
else:
|
754 |
+
trainer.create_model_card(**kwargs)
|
755 |
+
|
756 |
+
|
757 |
+
def _mp_fn(index):
|
758 |
+
# For xla_spawn (TPUs)
|
759 |
+
main()
|
760 |
+
|
761 |
+
|
762 |
+
if __name__ == "__main__":
|
763 |
+
main()
|