()
| 174 | |
| 175 | |
| 176 | def main(): |
| 177 | # See all possible arguments in src/transformers/training_args.py |
| 178 | # or by passing the --help flag to this script. |
| 179 | # We now keep distinct sets of args, for a cleaner separation of concerns. |
| 180 | |
| 181 | parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments)) |
| 182 | if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): |
| 183 | # If we pass only one argument to the script and it's the path to a json file, |
| 184 | # let's parse it to get our arguments. |
| 185 | model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1])) |
| 186 | else: |
| 187 | model_args, data_args, training_args = parser.parse_args_into_dataclasses() |
| 188 | |
| 189 | # Setup logging |
| 190 | logging.basicConfig( |
| 191 | format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", |
| 192 | datefmt="%m/%d/%Y %H:%M:%S", |
| 193 | handlers=[logging.StreamHandler(sys.stdout)], |
| 194 | ) |
| 195 | |
| 196 | if training_args.should_log: |
| 197 | # The default of training_args.log_level is passive, so we set log level at info here to have that default. |
| 198 | transformers.utils.logging.set_verbosity_info() |
| 199 | |
| 200 | log_level = training_args.get_process_log_level() |
| 201 | logger.setLevel(log_level) |
| 202 | datasets.utils.logging.set_verbosity(log_level) |
| 203 | transformers.utils.logging.set_verbosity(log_level) |
| 204 | transformers.utils.logging.enable_default_handler() |
| 205 | transformers.utils.logging.enable_explicit_format() |
| 206 | |
| 207 | # Log on each process the small summary: |
| 208 | logger.warning( |
| 209 | f"Process rank: {training_args.local_process_index}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, " |
| 210 | + f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}" |
| 211 | ) |
| 212 | logger.info(f"Training/evaluation parameters {training_args}") |
| 213 | |
| 214 | # Set seed before initializing model. |
| 215 | set_seed(training_args.seed) |
| 216 | |
| 217 | # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below) |
| 218 | # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/ |
| 219 | # (the dataset will be downloaded automatically from the datasets Hub). |
| 220 | |
| 221 | # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called |
| 222 | # 'text' is found. You can easily tweak this behavior (see below). |
| 223 | |
| 224 | # In distributed training, the load_dataset function guarantee that only one local process can concurrently |
| 225 | # download the dataset. |
| 226 | if data_args.train_file is not None or data_args.validation_file is not None: |
| 227 | data_files = {} |
| 228 | if data_args.train_file is not None: |
| 229 | data_files["train"] = data_args.train_file |
| 230 | extension = data_args.train_file.split(".")[-1] |
| 231 | if data_args.validation_file is not None: |
| 232 | data_files["validation"] = data_args.validation_file |
| 233 | extension = data_args.validation_file.split(".")[-1] |
no test coverage detected