Upload run_roberta_train_dataset.py
Browse files- run_roberta_train_dataset.py +503 -0
run_roberta_train_dataset.py
ADDED
@@ -0,0 +1,503 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
# Copyright 2020 The HuggingFace 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 |
+
"""
|
17 |
+
Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) on a text file or a dataset.
|
18 |
+
|
19 |
+
Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
|
20 |
+
https://huggingface.co/models?filter=fill-mask
|
21 |
+
"""
|
22 |
+
# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
|
23 |
+
|
24 |
+
import logging
|
25 |
+
import math
|
26 |
+
import os
|
27 |
+
import sys
|
28 |
+
import warnings
|
29 |
+
from dataclasses import dataclass, field
|
30 |
+
from typing import Optional
|
31 |
+
|
32 |
+
import datasets
|
33 |
+
import evaluate
|
34 |
+
import torch
|
35 |
+
from datasets import DatasetDict
|
36 |
+
|
37 |
+
import transformers
|
38 |
+
from transformers import (
|
39 |
+
CONFIG_MAPPING,
|
40 |
+
MODEL_FOR_MASKED_LM_MAPPING,
|
41 |
+
AutoConfig,
|
42 |
+
AutoModelForMaskedLM,
|
43 |
+
DataCollatorForLanguageModeling,
|
44 |
+
HfArgumentParser,
|
45 |
+
Trainer,
|
46 |
+
TrainingArguments,
|
47 |
+
is_torch_xla_available,
|
48 |
+
set_seed,
|
49 |
+
)
|
50 |
+
from transformers.trainer_utils import get_last_checkpoint
|
51 |
+
from transformers.utils.versions import require_version
|
52 |
+
sys.path.append(os.path.abspath(os.path.dirname(__file__) + "/.."))
|
53 |
+
|
54 |
+
from dna_tokenizer_fast import DNATokenizerFast
|
55 |
+
sys.path.append(os.path.join(os.path.dirname(__file__), "accuracy"))
|
56 |
+
|
57 |
+
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
|
58 |
+
# check_min_version("4.40.0.dev0")
|
59 |
+
|
60 |
+
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
|
61 |
+
|
62 |
+
logger = logging.getLogger(__name__)
|
63 |
+
MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
|
64 |
+
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
|
65 |
+
|
66 |
+
|
67 |
+
@dataclass
|
68 |
+
class ModelArguments:
|
69 |
+
"""
|
70 |
+
Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
|
71 |
+
"""
|
72 |
+
|
73 |
+
model_name_or_path: Optional[str] = field(
|
74 |
+
default=None,
|
75 |
+
metadata={
|
76 |
+
"help": (
|
77 |
+
"The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
|
78 |
+
)
|
79 |
+
},
|
80 |
+
)
|
81 |
+
model_type: Optional[str] = field(
|
82 |
+
default=None,
|
83 |
+
metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
|
84 |
+
)
|
85 |
+
config_overrides: Optional[str] = field(
|
86 |
+
default=None,
|
87 |
+
metadata={
|
88 |
+
"help": (
|
89 |
+
"Override some existing default config settings when a model is trained from scratch. Example: "
|
90 |
+
"n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
|
91 |
+
)
|
92 |
+
},
|
93 |
+
)
|
94 |
+
config_name: Optional[str] = field(
|
95 |
+
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
|
96 |
+
)
|
97 |
+
tokenizer_name: Optional[str] = field(
|
98 |
+
default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
|
99 |
+
)
|
100 |
+
cache_dir: Optional[str] = field(
|
101 |
+
default=None,
|
102 |
+
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
|
103 |
+
)
|
104 |
+
use_fast_tokenizer: bool = field(
|
105 |
+
default=True,
|
106 |
+
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
|
107 |
+
)
|
108 |
+
model_revision: str = field(
|
109 |
+
default="main",
|
110 |
+
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
|
111 |
+
)
|
112 |
+
token: str = field(
|
113 |
+
default=None,
|
114 |
+
metadata={
|
115 |
+
"help": (
|
116 |
+
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
|
117 |
+
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
|
118 |
+
)
|
119 |
+
},
|
120 |
+
)
|
121 |
+
use_auth_token: bool = field(
|
122 |
+
default=None,
|
123 |
+
metadata={
|
124 |
+
"help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead."
|
125 |
+
},
|
126 |
+
)
|
127 |
+
trust_remote_code: bool = field(
|
128 |
+
default=False,
|
129 |
+
metadata={
|
130 |
+
"help": (
|
131 |
+
"Whether or not to allow for custom models defined on the Hub in their own modeling files. This option "
|
132 |
+
"should only be set to `True` for repositories you trust and in which you have read the code, as it will "
|
133 |
+
"execute code present on the Hub on your local machine."
|
134 |
+
)
|
135 |
+
},
|
136 |
+
)
|
137 |
+
torch_dtype: Optional[str] = field(
|
138 |
+
default=None,
|
139 |
+
metadata={
|
140 |
+
"help": (
|
141 |
+
"Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
|
142 |
+
"dtype will be automatically derived from the model's weights."
|
143 |
+
),
|
144 |
+
"choices": ["auto", "bfloat16", "float16", "float32"],
|
145 |
+
},
|
146 |
+
)
|
147 |
+
low_cpu_mem_usage: bool = field(
|
148 |
+
default=False,
|
149 |
+
metadata={
|
150 |
+
"help": (
|
151 |
+
"It is an option to create the model as an empty shell, then only materialize its parameters when the pretrained weights are loaded. "
|
152 |
+
"set True will benefit LLM loading time and RAM consumption."
|
153 |
+
)
|
154 |
+
},
|
155 |
+
)
|
156 |
+
|
157 |
+
def __post_init__(self):
|
158 |
+
if self.config_overrides is not None and (self.config_name is not None or self.model_name_or_path is not None):
|
159 |
+
raise ValueError(
|
160 |
+
"--config_overrides can't be used in combination with --config_name or --model_name_or_path"
|
161 |
+
)
|
162 |
+
|
163 |
+
|
164 |
+
@dataclass
|
165 |
+
class DataTrainingArguments:
|
166 |
+
"""
|
167 |
+
Arguments pertaining to what data we are going to input our model for training and eval.
|
168 |
+
"""
|
169 |
+
|
170 |
+
dataset_name: Optional[str] = field(
|
171 |
+
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
|
172 |
+
)
|
173 |
+
dataset_config_name: Optional[str] = field(
|
174 |
+
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
|
175 |
+
)
|
176 |
+
train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
|
177 |
+
validation_file: Optional[str] = field(
|
178 |
+
default=None,
|
179 |
+
metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
|
180 |
+
)
|
181 |
+
overwrite_cache: bool = field(
|
182 |
+
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
|
183 |
+
)
|
184 |
+
validation_split_percentage: Optional[int] = field(
|
185 |
+
default=5,
|
186 |
+
metadata={
|
187 |
+
"help": "The percentage of the train set used as validation set in case there's no validation split"
|
188 |
+
},
|
189 |
+
)
|
190 |
+
max_seq_length: Optional[int] = field(
|
191 |
+
default=None,
|
192 |
+
metadata={
|
193 |
+
"help": (
|
194 |
+
"The maximum total input sequence length after tokenization. Sequences longer "
|
195 |
+
"than this will be truncated."
|
196 |
+
)
|
197 |
+
},
|
198 |
+
)
|
199 |
+
preprocessing_num_workers: Optional[int] = field(
|
200 |
+
default=None,
|
201 |
+
metadata={"help": "The number of processes to use for the preprocessing."},
|
202 |
+
)
|
203 |
+
mlm_probability: float = field(
|
204 |
+
default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
|
205 |
+
)
|
206 |
+
line_by_line: bool = field(
|
207 |
+
default=False,
|
208 |
+
metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
|
209 |
+
)
|
210 |
+
pad_to_max_length: bool = field(
|
211 |
+
default=False,
|
212 |
+
metadata={
|
213 |
+
"help": (
|
214 |
+
"Whether to pad all samples to `max_seq_length`. "
|
215 |
+
"If False, will pad the samples dynamically when batching to the maximum length in the batch."
|
216 |
+
)
|
217 |
+
},
|
218 |
+
)
|
219 |
+
max_train_samples: Optional[int] = field(
|
220 |
+
default=None,
|
221 |
+
metadata={
|
222 |
+
"help": (
|
223 |
+
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
224 |
+
"value if set."
|
225 |
+
)
|
226 |
+
},
|
227 |
+
)
|
228 |
+
max_eval_samples: Optional[int] = field(
|
229 |
+
default=None,
|
230 |
+
metadata={
|
231 |
+
"help": (
|
232 |
+
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
|
233 |
+
"value if set."
|
234 |
+
)
|
235 |
+
},
|
236 |
+
)
|
237 |
+
streaming: bool = field(default=False, metadata={"help": "Enable streaming mode"})
|
238 |
+
|
239 |
+
def main():
|
240 |
+
# See all possible arguments in src/transformers/training_args.py
|
241 |
+
# or by passing the --help flag to this script.
|
242 |
+
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
243 |
+
|
244 |
+
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
|
245 |
+
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
246 |
+
# If we pass only one argument to the script and it's the path to a json file,
|
247 |
+
# let's parse it to get our arguments.
|
248 |
+
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
|
249 |
+
else:
|
250 |
+
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
251 |
+
|
252 |
+
if model_args.use_auth_token is not None:
|
253 |
+
warnings.warn(
|
254 |
+
"The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.",
|
255 |
+
FutureWarning,
|
256 |
+
)
|
257 |
+
if model_args.token is not None:
|
258 |
+
raise ValueError("`token` and `use_auth_token` are both specified. Please set only the argument `token`.")
|
259 |
+
model_args.token = model_args.use_auth_token
|
260 |
+
|
261 |
+
# Setup logging
|
262 |
+
logging.basicConfig(
|
263 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
264 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
265 |
+
handlers=[logging.StreamHandler(sys.stdout)],
|
266 |
+
)
|
267 |
+
|
268 |
+
if training_args.should_log:
|
269 |
+
# The default of training_args.log_level is passive, so we set log level at info here to have that default.
|
270 |
+
transformers.utils.logging.set_verbosity_info()
|
271 |
+
|
272 |
+
log_level = training_args.get_process_log_level()
|
273 |
+
logger.setLevel(log_level)
|
274 |
+
datasets.utils.logging.set_verbosity(log_level)
|
275 |
+
transformers.utils.logging.set_verbosity(log_level)
|
276 |
+
transformers.utils.logging.enable_default_handler()
|
277 |
+
transformers.utils.logging.enable_explicit_format()
|
278 |
+
|
279 |
+
# Log on each process the small summary:
|
280 |
+
logger.warning(
|
281 |
+
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
|
282 |
+
+ f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
|
283 |
+
)
|
284 |
+
# Set the verbosity to info of the Transformers logger (on main process only):
|
285 |
+
logger.info(f"Training/evaluation parameters {training_args}")
|
286 |
+
|
287 |
+
# logger.info(f"xiao_print change output_dir, {str(datetime.datetime.now())[:-3]}")
|
288 |
+
# import time
|
289 |
+
# training_args.output_dir = training_args.output_dir + str(time.time())[-4:]
|
290 |
+
|
291 |
+
# Detecting last checkpoint.
|
292 |
+
last_checkpoint = None
|
293 |
+
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
|
294 |
+
last_checkpoint = get_last_checkpoint(training_args.output_dir)
|
295 |
+
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
|
296 |
+
raise ValueError(
|
297 |
+
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
|
298 |
+
"Use --overwrite_output_dir to overcome."
|
299 |
+
)
|
300 |
+
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
|
301 |
+
logger.info(
|
302 |
+
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
|
303 |
+
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
|
304 |
+
)
|
305 |
+
|
306 |
+
# Set seed before initializing model.
|
307 |
+
set_seed(training_args.seed)
|
308 |
+
|
309 |
+
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
|
310 |
+
# https://huggingface.co/docs/datasets/loading_datasets.
|
311 |
+
|
312 |
+
# Load pretrained model and tokenizer
|
313 |
+
#
|
314 |
+
# Distributed training:
|
315 |
+
# The .from_pretrained methods guarantee that only one local process can concurrently
|
316 |
+
# download model & vocab.
|
317 |
+
config_kwargs = {
|
318 |
+
"cache_dir": model_args.cache_dir,
|
319 |
+
"revision": model_args.model_revision,
|
320 |
+
"token": model_args.token,
|
321 |
+
"trust_remote_code": model_args.trust_remote_code,
|
322 |
+
}
|
323 |
+
if model_args.config_name:
|
324 |
+
config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
|
325 |
+
elif model_args.model_name_or_path:
|
326 |
+
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
|
327 |
+
else:
|
328 |
+
config = CONFIG_MAPPING[model_args.model_type]()
|
329 |
+
logger.warning("You are instantiating a new config instance from scratch.")
|
330 |
+
if model_args.config_overrides is not None:
|
331 |
+
logger.info(f"Overriding config: {model_args.config_overrides}")
|
332 |
+
config.update_from_string(model_args.config_overrides)
|
333 |
+
logger.info(f"New config: {config}")
|
334 |
+
|
335 |
+
tokenizer_kwargs = {
|
336 |
+
"cache_dir": model_args.cache_dir,
|
337 |
+
"use_fast": model_args.use_fast_tokenizer,
|
338 |
+
"revision": model_args.model_revision,
|
339 |
+
"token": model_args.token,
|
340 |
+
"trust_remote_code": model_args.trust_remote_code,
|
341 |
+
}
|
342 |
+
if model_args.tokenizer_name:
|
343 |
+
tokenizer = DNATokenizerFast.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
|
344 |
+
elif model_args.model_name_or_path:
|
345 |
+
tokenizer = DNATokenizerFast.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
|
346 |
+
else:
|
347 |
+
raise ValueError(
|
348 |
+
"You are instantiating a new tokenizer from scratch. This is not supported by this script. "
|
349 |
+
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
|
350 |
+
)
|
351 |
+
|
352 |
+
if model_args.model_name_or_path:
|
353 |
+
torch_dtype = (
|
354 |
+
model_args.torch_dtype
|
355 |
+
if model_args.torch_dtype in ["auto", None]
|
356 |
+
else getattr(torch, model_args.torch_dtype)
|
357 |
+
)
|
358 |
+
model = AutoModelForMaskedLM.from_pretrained(
|
359 |
+
model_args.model_name_or_path,
|
360 |
+
from_tf=bool(".ckpt" in model_args.model_name_or_path),
|
361 |
+
config=config,
|
362 |
+
cache_dir=model_args.cache_dir,
|
363 |
+
revision=model_args.model_revision,
|
364 |
+
token=model_args.token,
|
365 |
+
trust_remote_code=model_args.trust_remote_code,
|
366 |
+
torch_dtype=torch_dtype,
|
367 |
+
low_cpu_mem_usage=model_args.low_cpu_mem_usage,
|
368 |
+
)
|
369 |
+
else:
|
370 |
+
logger.info("Training new model from scratch")
|
371 |
+
model = AutoModelForMaskedLM.from_config(config, trust_remote_code=model_args.trust_remote_code)
|
372 |
+
|
373 |
+
# We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
|
374 |
+
# on a small vocab and want a smaller embedding size, remove this test.
|
375 |
+
embedding_size = model.get_input_embeddings().weight.shape[0]
|
376 |
+
if len(tokenizer) > embedding_size:
|
377 |
+
model.resize_token_embeddings(len(tokenizer))
|
378 |
+
|
379 |
+
origin_datasets = datasets.load_from_disk(data_args.dataset_name)
|
380 |
+
|
381 |
+
data_args.validation_split_percentage = None \
|
382 |
+
if (isinstance(origin_datasets, DatasetDict) and "validation" in origin_datasets.keys()) \
|
383 |
+
else data_args.validation_split_percentage
|
384 |
+
if data_args.validation_split_percentage > 0.0:
|
385 |
+
split_ratio = data_args.validation_split_percentage / 100
|
386 |
+
tokenized_datasets = origin_datasets.train_test_split(split_ratio)
|
387 |
+
tokenized_datasets["validation"] = tokenized_datasets["test"]
|
388 |
+
else:
|
389 |
+
tokenized_datasets = origin_datasets
|
390 |
+
|
391 |
+
if training_args.do_train:
|
392 |
+
if "train" not in tokenized_datasets:
|
393 |
+
raise ValueError("--do_train requires a train dataset")
|
394 |
+
train_dataset = tokenized_datasets["train"]
|
395 |
+
if data_args.max_train_samples is not None:
|
396 |
+
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
|
397 |
+
train_dataset = train_dataset.select(range(max_train_samples))
|
398 |
+
|
399 |
+
if training_args.do_eval:
|
400 |
+
if "validation" not in tokenized_datasets:
|
401 |
+
raise ValueError("--do_eval requires a validation dataset")
|
402 |
+
eval_dataset = tokenized_datasets["validation"]
|
403 |
+
if data_args.max_eval_samples is not None:
|
404 |
+
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
|
405 |
+
eval_dataset = eval_dataset.select(range(max_eval_samples))
|
406 |
+
|
407 |
+
def preprocess_logits_for_metrics(logits, labels):
|
408 |
+
if isinstance(logits, tuple):
|
409 |
+
# Depending on the model and config, logits may contain extra tensors,
|
410 |
+
# like past_key_values, but logits always come first
|
411 |
+
logits = logits[0]
|
412 |
+
return logits.argmax(dim=-1)
|
413 |
+
|
414 |
+
metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
|
415 |
+
|
416 |
+
def compute_metrics(eval_preds):
|
417 |
+
preds, labels = eval_preds
|
418 |
+
# preds have the same shape as the labels, after the argmax(-1) has been calculated
|
419 |
+
# by preprocess_logits_for_metrics
|
420 |
+
labels = labels.reshape(-1)
|
421 |
+
preds = preds.reshape(-1)
|
422 |
+
mask = labels != -100
|
423 |
+
labels = labels[mask]
|
424 |
+
preds = preds[mask]
|
425 |
+
return metric.compute(predictions=preds, references=labels)
|
426 |
+
|
427 |
+
# Data collator
|
428 |
+
# This one will take care of randomly masking the tokens.
|
429 |
+
pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
|
430 |
+
data_collator = DataCollatorForLanguageModeling(
|
431 |
+
tokenizer=tokenizer,
|
432 |
+
mlm_probability=data_args.mlm_probability,
|
433 |
+
pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
|
434 |
+
)
|
435 |
+
|
436 |
+
# Initialize our Trainer
|
437 |
+
trainer = Trainer(
|
438 |
+
model=model,
|
439 |
+
args=training_args,
|
440 |
+
train_dataset=train_dataset if training_args.do_train else None,
|
441 |
+
eval_dataset=eval_dataset if training_args.do_eval else None,
|
442 |
+
tokenizer=tokenizer,
|
443 |
+
data_collator=data_collator,
|
444 |
+
compute_metrics=compute_metrics if training_args.do_eval and not is_torch_xla_available() else None,
|
445 |
+
preprocess_logits_for_metrics=preprocess_logits_for_metrics
|
446 |
+
if training_args.do_eval and not is_torch_xla_available()
|
447 |
+
else None,
|
448 |
+
)
|
449 |
+
|
450 |
+
# Training
|
451 |
+
if training_args.do_train:
|
452 |
+
checkpoint = None
|
453 |
+
if training_args.resume_from_checkpoint is not None:
|
454 |
+
checkpoint = training_args.resume_from_checkpoint
|
455 |
+
elif last_checkpoint is not None:
|
456 |
+
checkpoint = last_checkpoint
|
457 |
+
train_result = trainer.train(resume_from_checkpoint=checkpoint)
|
458 |
+
trainer.save_model() # Saves the tokenizer too for easy upload
|
459 |
+
metrics = train_result.metrics
|
460 |
+
|
461 |
+
max_train_samples = (
|
462 |
+
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
|
463 |
+
)
|
464 |
+
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
|
465 |
+
|
466 |
+
trainer.log_metrics("train", metrics)
|
467 |
+
trainer.save_metrics("train", metrics)
|
468 |
+
trainer.save_state()
|
469 |
+
|
470 |
+
# Evaluation
|
471 |
+
if training_args.do_eval:
|
472 |
+
logger.info("*** Evaluate ***")
|
473 |
+
|
474 |
+
metrics = trainer.evaluate()
|
475 |
+
|
476 |
+
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
|
477 |
+
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
|
478 |
+
try:
|
479 |
+
perplexity = math.exp(metrics["eval_loss"])
|
480 |
+
except OverflowError:
|
481 |
+
perplexity = float("inf")
|
482 |
+
metrics["perplexity"] = perplexity
|
483 |
+
|
484 |
+
trainer.log_metrics("eval", metrics)
|
485 |
+
trainer.save_metrics("eval", metrics)
|
486 |
+
|
487 |
+
# kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "fill-mask"}
|
488 |
+
# if data_args.dataset_name is not None:
|
489 |
+
# kwargs["dataset_tags"] = data_args.dataset_name
|
490 |
+
# if data_args.dataset_config_name is not None:
|
491 |
+
# kwargs["dataset_args"] = data_args.dataset_config_name
|
492 |
+
# kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
|
493 |
+
# else:
|
494 |
+
# kwargs["dataset"] = data_args.dataset_name
|
495 |
+
|
496 |
+
# if training_args.push_to_hub:
|
497 |
+
# trainer.push_to_hub(**kwargs)
|
498 |
+
# else:
|
499 |
+
# trainer.create_model_card(**kwargs)
|
500 |
+
|
501 |
+
|
502 |
+
if __name__ == "__main__":
|
503 |
+
main()
|