def main(): # See all possible arguments in src/transformers/training_args.py # or by passing the --help flag to this script. # We now keep distinct sets of args, for a cleaner separation of concerns. parser = HfArgumentParser( (ModelArguments, DataTrainingArguments, TrainingArguments)) if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): # If we pass only one argument to the script and it's the path to a json file, # let's parse it to get our arguments. model_args, data_args, training_args = parser.parse_json_file( json_file=os.path.abspath(sys.argv[1])) else: model_args, data_args, training_args = parser.parse_args_into_dataclasses( ) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", handlers=[logging.StreamHandler(sys.stdout)], ) log_level = training_args.get_process_log_level() logger.setLevel(log_level) datasets.utils.logging.set_verbosity(log_level) transformers.utils.logging.set_verbosity(log_level) transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() # Log on each process the small summary: logger.warning( f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}" + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}" ) logger.info(f"Training/evaluation parameters {training_args}") # Detecting last checkpoint. last_checkpoint = None if os.path.isdir( training_args.output_dir ) and training_args.do_train and not training_args.overwrite_output_dir: last_checkpoint = get_last_checkpoint(training_args.output_dir) if last_checkpoint is None and len(os.listdir( training_args.output_dir)) > 0: raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty. " "Use --overwrite_output_dir to overcome.") elif last_checkpoint is not None and training_args.resume_from_checkpoint is None: logger.info( f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." ) # Set seed before initializing model. set_seed(training_args.seed) # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below) # or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub). # # For JSON files, this script will use the `question` column for the input question and `table` column for the corresponding table. # # If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this # single column. You can easily tweak this behavior (see below) # # In distributed training, the load_dataset function guarantee that only one local process can concurrently # download the dataset. if data_args.dataset_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir) else: # Loading a dataset from your local files. # CSV/JSON training and evaluation files are needed. data_files = { "train": data_args.train_file, "validation": data_args.validation_file } # Get the test dataset: you can provide your own CSV/JSON test file (see below) # when you use `do_predict` without specifying a GLUE benchmark task. if training_args.do_predict: if data_args.test_file is not None: train_extension = data_args.train_file.split(".")[-1] test_extension = data_args.test_file.split(".")[-1] assert ( test_extension == train_extension ), "`test_file` should have the same extension (csv or json) as `train_file`." data_files["test"] = data_args.test_file else: raise ValueError( "Need either a GLUE task or a test file for `do_predict`.") for key in data_files.keys(): logger.info(f"load a local file for {key}: {data_files[key]}") if data_args.train_file.endswith(".csv"): # Loading a dataset from local csv files raw_datasets = load_dataset("csv", data_files=data_files, cache_dir=model_args.cache_dir) else: # Loading a dataset from local json files raw_datasets = load_dataset("json", data_files=data_files, cache_dir=model_args.cache_dir) # See more about loading any type of standard or custom dataset at # https://huggingface.co/docs/datasets/loading_datasets.html. # Labels label_list = raw_datasets["train"].features["label"].names num_labels = len(label_list) # Load pretrained model and tokenizer # # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. config = AutoConfig.from_pretrained( model_args.config_name if model_args.config_name else model_args.model_name_or_path, num_labels=num_labels, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) # load tapex tokenizer tokenizer = TapexTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, add_prefix_space=True, ) model = BartForSequenceClassification.from_pretrained( model_args.model_name_or_path, from_tf=bool(".ckpt" in model_args.model_name_or_path), config=config, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) # Padding strategy if data_args.pad_to_max_length: padding = "max_length" else: # We will pad later, dynamically at batch creation, to the max sequence length in each batch padding = False # Some models have set the order of the labels to use, so let's make sure we do use it. model.config.label2id = {"Refused": 0, "Entailed": 1} model.config.id2label = {0: "Refused", 1: "Entailed"} if data_args.max_seq_length > tokenizer.model_max_length: logger.warning( f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the" f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}." ) max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length) def preprocess_tabfact_function(examples): # Tokenize the texts def _convert_table_text_to_pandas(_table_text): """Runs the structured pandas table object for _table_text. An example _table_text can be: round#clubs remaining\nfirst round#156\n """ _table_content = [ _table_row.split("#") for _table_row in _table_text.strip("\n").split("\n") ] _table_pd = pd.DataFrame.from_records(_table_content[1:], columns=_table_content[0]) return _table_pd questions = examples["statement"] tables = list( map(_convert_table_text_to_pandas, examples["table_text"])) result = tokenizer(tables, questions, padding=padding, max_length=max_seq_length, truncation=True) result["label"] = examples["label"] return result with training_args.main_process_first(desc="dataset map pre-processing"): raw_datasets = raw_datasets.map( preprocess_tabfact_function, batched=True, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on dataset", ) if training_args.do_train: if "train" not in raw_datasets: raise ValueError("--do_train requires a train dataset") train_dataset = raw_datasets["train"] if data_args.max_train_samples is not None: train_dataset = train_dataset.select( range(data_args.max_train_samples)) if training_args.do_eval: if "validation" not in raw_datasets and "validation_matched" not in raw_datasets: raise ValueError("--do_eval requires a validation dataset") eval_dataset = raw_datasets["validation"] if data_args.max_eval_samples is not None: eval_dataset = eval_dataset.select( range(data_args.max_eval_samples)) if training_args.do_predict or data_args.test_file is not None: if "test" not in raw_datasets and "test_matched" not in raw_datasets: raise ValueError("--do_predict requires a test dataset") predict_dataset = raw_datasets["test"] if data_args.max_predict_samples is not None: predict_dataset = predict_dataset.select( range(data_args.max_predict_samples)) # Log a few random samples from the training set: if training_args.do_train: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a # predictions and label_ids field) and has to return a dictionary string to float. def compute_metrics(p: EvalPrediction): preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions preds = np.argmax(preds, axis=1) return { "accuracy": (preds == p.label_ids).astype(np.float32).mean().item() } # Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding. if data_args.pad_to_max_length: data_collator = default_data_collator elif training_args.fp16: data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8) else: data_collator = None # Initialize our Trainer trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset if training_args.do_train else None, eval_dataset=eval_dataset if training_args.do_eval else None, compute_metrics=compute_metrics, tokenizer=tokenizer, data_collator=data_collator, ) # Training if training_args.do_train: checkpoint = None if training_args.resume_from_checkpoint is not None: checkpoint = training_args.resume_from_checkpoint elif last_checkpoint is not None: checkpoint = last_checkpoint train_result = trainer.train(resume_from_checkpoint=checkpoint) metrics = train_result.metrics max_train_samples = (data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)) metrics["train_samples"] = min(max_train_samples, len(train_dataset)) trainer.save_model() # Saves the tokenizer too for easy upload trainer.log_metrics("train", metrics) trainer.save_metrics("train", metrics) trainer.save_state() # Evaluation if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate(eval_dataset=eval_dataset) max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len( eval_dataset) metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset)) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics) if training_args.do_predict: logger.info("*** Predict ***") # Removing the `label` columns because it contains -1 and Trainer won't like that. predict_dataset = predict_dataset.remove_columns("label") predictions = trainer.predict(predict_dataset, metric_key_prefix="predict").predictions predictions = np.argmax(predictions, axis=1) output_predict_file = os.path.join(training_args.output_dir, "predict_results_tabfact.txt") if trainer.is_world_process_zero(): with open(output_predict_file, "w") as writer: logger.info("***** Predict Results *****") writer.write("index\tprediction\n") for index, item in enumerate(predictions): item = label_list[item] writer.write(f"{index}\t{item}\n") kwargs = { "finetuned_from": model_args.model_name_or_path, "tasks": "text-classification" } if training_args.push_to_hub: trainer.push_to_hub(**kwargs) else: trainer.create_model_card(**kwargs)
def main(): # See all possible arguments in src/transformers/training_args.py # or by passing the --help flag to this script. # We now keep distinct sets of args, for a cleaner separation of concerns. parser = HfArgumentParser( (ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments)) if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): # If we pass only one argument to the script and it's the path to a json file, # let's parse it to get our arguments. model_args, data_args, training_args = parser.parse_json_file( json_file=os.path.abspath(sys.argv[1])) else: model_args, data_args, training_args = parser.parse_args_into_dataclasses( ) # Detecting last checkpoint. last_checkpoint = None if os.path.isdir( training_args.output_dir ) and training_args.do_train and not training_args.overwrite_output_dir: last_checkpoint = get_last_checkpoint(training_args.output_dir) if last_checkpoint is None and len(os.listdir( training_args.output_dir)) > 0: raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty. " "Use --overwrite_output_dir to overcome.") elif last_checkpoint is not None and training_args.resume_from_checkpoint is None: logger.info( f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." ) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", handlers=[logging.StreamHandler(sys.stdout)], ) logger.setLevel(logging.INFO if is_main_process(training_args.local_rank ) else logging.WARN) # Log on each process the small summary: logger.warning( f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}" + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}" ) # Set the verbosity to info of the Transformers logger (on main process only): if is_main_process(training_args.local_rank): transformers.utils.logging.set_verbosity_info() logger.info(f"Training/evaluation parameters {training_args}") # Set seed before initializing model. set_seed(training_args.seed) # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below) # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/ # (the dataset will be downloaded automatically from the datasets Hub). # # For JSON files, this script will use the `question` column for the input question and `table` column for the corresponding table. # # In distributed training, the load_dataset function guarantee that only one local process can concurrently # download the dataset. if data_args.dataset_name is not None: # Downloading and loading a dataset from the hub. datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir) else: data_files = {} if data_args.train_file is not None: data_files["train"] = data_args.train_file extension = data_args.train_file.split(".")[-1] if data_args.validation_file is not None: data_files["validation"] = data_args.validation_file extension = data_args.validation_file.split(".")[-1] if data_args.test_file is not None: data_files["test"] = data_args.test_file extension = data_args.test_file.split(".")[-1] datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir) # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at # https://huggingface.co/docs/datasets/loading_datasets.html. # Load pretrained model and tokenizer # # Distributed training: # The .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. config = AutoConfig.from_pretrained( model_args.config_name if model_args.config_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) # IMPORTANT: the initial BART model's decoding is penalized by no_repeat_ngram_size, and thus # we should disable it here to avoid problematic generation config.no_repeat_ngram_size = 0 config.max_length = 1024 config.early_stopping = False # load tapex tokenizer tokenizer = TapexTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, add_prefix_space=True, ) # load Bart based Tapex model (default tapex-large) model = BartForConditionalGeneration.from_pretrained( model_args.model_name_or_path, from_tf=bool(".ckpt" in model_args.model_name_or_path), config=config, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) if model.config.decoder_start_token_id is None: raise ValueError( "Make sure that `config.decoder_start_token_id` is correctly defined" ) # Preprocessing the datasets. # We need to tokenize inputs and targets. if training_args.do_train: column_names = datasets["train"].column_names elif training_args.do_eval: column_names = datasets["validation"].column_names elif training_args.do_predict: column_names = datasets["test"].column_names else: logger.info( "There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`." ) return # Temporarily set max_target_length for training. max_target_length = data_args.max_target_length padding = "max_length" if data_args.pad_to_max_length else False if training_args.label_smoothing_factor > 0 and not hasattr( model, "prepare_decoder_input_ids_from_labels"): logger.warning( "label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for" f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory" ) def preprocess_tableqa_function(examples, is_training=False): """ The is_training FLAG is used to identify if we could use the supervision to truncate the table content if it is required. """ # this function is specific for WikiSQL since the util function need the data structure # to retrieve the WikiSQL answer for each question def _convert_table_types(_table): """Runs the type converter over the table cells.""" ret_table = deepcopy(_table) types = ret_table["types"] ret_table["real_rows"] = ret_table["rows"] typed_rows = [] for row in ret_table["rows"]: typed_row = [] for column, cell_value in enumerate(row): typed_row.append( _TYPE_CONVERTER[types[column]](cell_value)) typed_rows.append(typed_row) ret_table["rows"] = typed_rows return ret_table questions = [question.lower() for question in examples["question"]] example_tables = examples["table"] example_sqls = examples["sql"] tables = [ pd.DataFrame.from_records(example_table["rows"], columns=example_table["header"]) for example_table in example_tables ] # using tapas utils to obtain wikisql answer answers = [] for example_sql, example_table in zip(example_sqls, example_tables): tapas_table = _convert_table_types(example_table) answer_list: List[str] = retrieve_wikisql_query_answer_tapas( tapas_table, example_sql) # you can choose other delimiters to split each answer answers.append(answer_list) # IMPORTANT: we cannot pass by answers during evaluation, answers passed during training are used to # truncate large tables in the train set! if is_training: model_inputs = tokenizer( table=tables, query=questions, answer=answers, max_length=data_args.max_source_length, padding=padding, truncation=True, ) else: model_inputs = tokenizer(table=tables, query=questions, max_length=data_args.max_source_length, padding=padding, truncation=True) with tokenizer.as_target_tokenizer(): labels = tokenizer( answer=[", ".join(answer) for answer in answers], max_length=max_target_length, padding=padding, truncation=True, ) # If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore # padding in the loss. if padding == "max_length" and data_args.ignore_pad_token_for_loss: labels["input_ids"] = [[ (l if l != tokenizer.pad_token_id else -100) for l in label ] for label in labels["input_ids"]] model_inputs["labels"] = labels["input_ids"] return model_inputs # in training, we can use the answer as extra information to truncate large tables preprocess_tableqa_function_training = partial(preprocess_tableqa_function, is_training=True) if training_args.do_train: if "train" not in datasets: raise ValueError("--do_train requires a train dataset") train_dataset = datasets["train"] if data_args.max_train_samples is not None: train_dataset = train_dataset.select( range(data_args.max_train_samples)) train_dataset = train_dataset.map( preprocess_tableqa_function_training, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_eval: max_target_length = data_args.val_max_target_length if "validation" not in datasets: raise ValueError("--do_eval requires a validation dataset") eval_dataset = datasets["validation"] if data_args.max_eval_samples is not None: eval_dataset = eval_dataset.select( range(data_args.max_eval_samples)) eval_dataset = eval_dataset.map( preprocess_tableqa_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_predict: max_target_length = data_args.val_max_target_length if "test" not in datasets: raise ValueError("--do_predict requires a test dataset") predict_dataset = datasets["test"] if data_args.max_predict_samples is not None: predict_dataset = predict_dataset.select( range(data_args.max_predict_samples)) predict_dataset = predict_dataset.map( preprocess_tableqa_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # Data collator label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id data_collator = DataCollatorForSeq2Seq( tokenizer, model=model, label_pad_token_id=label_pad_token_id, pad_to_multiple_of=8 if training_args.fp16 else None, ) def postprocess_text(preds, labels): preds = [pred.strip() for pred in preds] labels = [label.strip() for label in labels] return preds, labels def compute_metrics(eval_preds): preds, labels = eval_preds if isinstance(preds, tuple): preds = preds[0] decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True) if data_args.ignore_pad_token_for_loss: # Replace -100 in the labels as we can't decode them. labels = np.where(labels != -100, labels, tokenizer.pad_token_id) decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) # Some simple post-processing decoded_preds, decoded_labels = postprocess_text( decoded_preds, decoded_labels) delimiter = ", " # define example evaluation def evaluate_example(predict_str: str, ground_str: str): predict_spans = predict_str.split(delimiter) ground_spans = ground_str.split(delimiter) predict_values = defaultdict(lambda: 0) ground_values = defaultdict(lambda: 0) for span in predict_spans: try: predict_values[float(span)] += 1 except ValueError: predict_values[span.strip()] += 1 for span in ground_spans: try: ground_values[float(span)] += 1 except ValueError: ground_values[span.strip()] += 1 is_correct = predict_values == ground_values return is_correct def get_denotation_accuracy(predictions: List[str], references: List[str]): assert len(predictions) == len(references) correct_num = 0 for predict_str, ground_str in zip(predictions, references): is_correct = evaluate_example(predict_str.lower(), ground_str.lower()) if is_correct: correct_num += 1 return correct_num / len(predictions) accuracy = get_denotation_accuracy(decoded_preds, decoded_labels) result = {"denotation_accuracy": accuracy} return result # Initialize our Trainer trainer = Seq2SeqTrainer( model=model, args=training_args, train_dataset=train_dataset if training_args.do_train else None, eval_dataset=eval_dataset if training_args.do_eval else None, tokenizer=tokenizer, data_collator=data_collator, compute_metrics=compute_metrics if training_args.predict_with_generate else None, ) if training_args.do_train: checkpoint = None if training_args.resume_from_checkpoint is not None: checkpoint = training_args.resume_from_checkpoint elif last_checkpoint is not None: checkpoint = last_checkpoint train_result = trainer.train(resume_from_checkpoint=checkpoint) trainer.save_model() # Saves the tokenizer too for easy upload metrics = train_result.metrics max_train_samples = (data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)) metrics["train_samples"] = min(max_train_samples, len(train_dataset)) trainer.log_metrics("train", metrics) trainer.save_metrics("train", metrics) trainer.save_state() # Evaluation results = {} if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate(max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, metric_key_prefix="eval") max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len( eval_dataset) metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset)) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics) if training_args.do_predict: logger.info("*** Predict ***") predict_results = trainer.predict( predict_dataset, metric_key_prefix="predict", max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, ) metrics = predict_results.metrics max_predict_samples = (data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)) metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset)) trainer.log_metrics("predict", metrics) trainer.save_metrics("predict", metrics) if trainer.is_world_process_zero(): if training_args.predict_with_generate: predictions = tokenizer.batch_decode( predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True) predictions = [pred.strip() for pred in predictions] output_prediction_file = os.path.join(training_args.output_dir, "tapex_predictions.txt") with open(output_prediction_file, "w") as writer: writer.write("\n".join(predictions)) return results