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( ) check_output_dir(training_args) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN, ) logger.warning( "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", training_args.local_rank, training_args.device, training_args.n_gpu, bool(training_args.parallel_mode == ParallelMode.DISTRIBUTED), training_args.fp16, ) transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() # 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("Training/evaluation parameters %s", training_args) # Set seed set_seed(training_args.seed) # 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, ) extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout") for p in extra_model_params: if getattr(training_args, p, None): assert hasattr( config, p ), f"({config.__class__.__name__}) doesn't have a `{p}` attribute" setattr(config, p, getattr(training_args, p)) tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, ) model = AutoModelForSeq2SeqLM.from_pretrained( model_args.model_name_or_path, from_tf=".ckpt" in model_args.model_name_or_path, config=config, cache_dir=model_args.cache_dir, ) # use task specific params use_task_specific_params(model, data_args.task) # set num_beams for evaluation if data_args.eval_beams is None: data_args.eval_beams = model.config.num_beams # set decoder_start_token_id for MBart if model.config.decoder_start_token_id is None and isinstance( tokenizer, (MBartTokenizer, MBartTokenizerFast)): assert (data_args.tgt_lang is not None and data_args.src_lang is not None), "mBart requires --tgt_lang and --src_lang" if isinstance(tokenizer, MBartTokenizer): model.config.decoder_start_token_id = tokenizer.lang_code_to_id[ data_args.tgt_lang] else: model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids( data_args.tgt_lang) if model_args.freeze_embeds: freeze_embeds(model) if model_args.freeze_encoder: freeze_params(model.get_encoder()) assert_all_frozen(model.get_encoder()) dataset_class = Seq2SeqDataset # Get datasets train_dataset = (dataset_class( tokenizer, type_path="train", data_dir=data_args.data_dir, n_obs=data_args.n_train, max_target_length=data_args.max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_train else None) eval_dataset = (dataset_class( tokenizer, type_path="val", data_dir=data_args.data_dir, n_obs=data_args.n_val, max_target_length=data_args.val_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_eval or training_args.evaluation_strategy != EvaluationStrategy.NO else None) test_dataset = (dataset_class( tokenizer, type_path="test", data_dir=data_args.data_dir, n_obs=data_args.n_test, max_target_length=data_args.test_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_predict else None) # Initialize our Trainer compute_metrics_fn = (build_compute_metrics_fn(data_args.task, tokenizer) if training_args.predict_with_generate else None) trainer = Seq2SeqTrainer( model=model, args=training_args, data_args=data_args, train_dataset=train_dataset, eval_dataset=eval_dataset, data_collator=Seq2SeqDataCollator(tokenizer, data_args, model.config.decoder_start_token_id, training_args.tpu_num_cores), compute_metrics=compute_metrics_fn, tokenizer=tokenizer, ) all_metrics = {} # Training if training_args.do_train: logger.info("*** Train ***") train_result = trainer.train( model_path=model_args.model_name_or_path if os.path. isdir(model_args.model_name_or_path) else None) metrics = train_result.metrics metrics["train_n_objs"] = data_args.n_train trainer.save_model() # this also saves the tokenizer if trainer.is_world_process_zero(): handle_metrics("train", metrics, training_args.output_dir) all_metrics.update(metrics) # Need to save the state, since Trainer.save_model saves only the tokenizer with the model trainer.state.save_to_json( os.path.join(training_args.output_dir, "trainer_state.json")) # For convenience, we also re-save the tokenizer to the same directory, # so that you can share your model easily on huggingface.co/models =) tokenizer.save_pretrained(training_args.output_dir) # Evaluation if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate(metric_key_prefix="val") metrics["val_n_objs"] = data_args.n_val metrics["val_loss"] = round(metrics["val_loss"], 4) if trainer.is_world_process_zero(): handle_metrics("val", metrics, training_args.output_dir) all_metrics.update(metrics) if training_args.do_predict: logger.info("*** Predict ***") test_output = trainer.predict(test_dataset=test_dataset, metric_key_prefix="test") metrics = test_output.metrics metrics["test_n_objs"] = data_args.n_test if trainer.is_world_process_zero(): metrics["test_loss"] = round(metrics["test_loss"], 4) handle_metrics("test", metrics, training_args.output_dir) all_metrics.update(metrics) if training_args.predict_with_generate: test_preds = tokenizer.batch_decode( test_output.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True) test_preds = lmap(str.strip, test_preds) write_txt_file( test_preds, os.path.join(training_args.output_dir, "test_generations.txt")) if trainer.is_world_process_zero(): save_json(all_metrics, os.path.join(training_args.output_dir, "all_results.json")) return all_metrics
def __init__(self, hparams, **kwargs): if hparams.sortish_sampler and hparams.gpus > 1: hparams.replace_sampler_ddp = False elif hparams.max_tokens_per_batch is not None: if hparams.gpus > 1: raise NotImplementedError( "Dynamic Batch size does not work for multi-gpu training") if hparams.sortish_sampler: raise ValueError( "--sortish_sampler and --max_tokens_per_batch may not be used simultaneously" ) super().__init__(hparams, num_labels=None, mode=self.mode, **kwargs) use_task_specific_params(self.model, "summarization") save_git_info(self.hparams.output_dir) self.metrics_save_path = Path(self.output_dir) / "metrics.json" self.hparams_save_path = Path(self.output_dir) / "hparams.pkl" pickle_save(self.hparams, self.hparams_save_path) self.step_count = 0 self.metrics = defaultdict(list) self.model_type = self.config.model_type self.vocab_size = self.config.tgt_vocab_size if self.model_type == "fsmt" else self.config.vocab_size self.dataset_kwargs: dict = dict( data_dir=self.hparams.data_dir, max_source_length=self.hparams.max_source_length, prefix=self.model.config.prefix or "", ) n_observations_per_split = { "train": self.hparams.n_train, "val": self.hparams.n_val, "test": self.hparams.n_test, } self.n_obs = { k: v if v >= 0 else None for k, v in n_observations_per_split.items() } self.target_lens = { "train": self.hparams.max_target_length, "val": self.hparams.val_max_target_length, "test": self.hparams.test_max_target_length, } assert self.target_lens["train"] <= self.target_lens[ "val"], f"target_lens: {self.target_lens}" assert self.target_lens["train"] <= self.target_lens[ "test"], f"target_lens: {self.target_lens}" if self.hparams.freeze_embeds: freeze_embeds(self.model) if self.hparams.freeze_encoder: freeze_params(self.model.get_encoder()) assert_all_frozen(self.model.get_encoder()) self.hparams.git_sha = get_git_info()["repo_sha"] self.num_workers = hparams.num_workers self.decoder_start_token_id = None # default to config if self.model.config.decoder_start_token_id is None and isinstance( self.tokenizer, MBartTokenizer): self.decoder_start_token_id = self.tokenizer.lang_code_to_id[ hparams.tgt_lang] self.model.config.decoder_start_token_id = self.decoder_start_token_id self.dataset_class = (Seq2SeqDataset if hasattr( self.tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset) self.already_saved_batch = False self.eval_beams = self.model.config.num_beams if self.hparams.eval_beams is None else self.hparams.eval_beams if self.hparams.eval_max_gen_length is not None: self.eval_max_length = self.hparams.eval_max_gen_length else: self.eval_max_length = self.model.config.max_length self.val_metric = self.default_val_metric if self.hparams.val_metric is None else self.hparams.val_metric
def __init__(self, hparams, **kwargs): if hparams.sortish_sampler and hparams.gpus > 1: hparams.replace_sampler_ddp = False elif hparams.max_tokens_per_batch is not None: if hparams.gpus > 1: raise NotImplementedError( "Dynamic Batch size does not work for multi-gpu training") if hparams.sortish_sampler: raise ValueError( "--sortish_sampler and --max_tokens_per_batch may not be used simultaneously" ) super().__init__(hparams, num_labels=None, mode=self.mode, **kwargs) # print(self.tokenizer.model_max_length) self.tokenizer.add_special_tokens({ 'additional_special_tokens': [ '<|HOME|>', '<|AWAY|>', '<|PLAYER-START_POSITION|>', '<|PLAYER-MIN|>', '<|PLAYER-PTS|>', '<|PLAYER-FGM|>', '<|PLAYER-FGA|>', '<|PLAYER-FG_PCT|>', '<|PLAYER-FG3M|>', '<|PLAYER-FG3A|>', '<|PLAYER-FG3_PCT|>', '<|PLAYER-FTM|>', '<|PLAYER-FTA|>', '<|PLAYER-FT_PCT|>', '<|PLAYER-OREB|>', '<|PLAYER-DREB|>', '<|PLAYER-REB|>', '<|PLAYER-AST|>', '<|PLAYER-TO|>', '<|PLAYER-STL|>', '<|PLAYER-BLK|>', '<|PLAYER-PF|>', '<|TEAM-PTS_QTR1|>', '<|TEAM-PTS_QTR2|>', '<|TEAM-PTS_QTR3|>', '<|TEAM-PTS_QTR4|>', '<|TEAM-PTS|>', '<|TEAM-FG_PCT|>', '<|TEAM-FG3_PCT|>', '<|TEAM-FT_PCT|>', '<|TEAM-REB|>', '<|TEAM-AST|>', '<|TEAM-TOV|>', '<|TEAM-WINS|>', '<|TEAM-LOSSES|>', '<|TEAM-CITY|>', '<|TEAM-NAME|>', ] }) # self.tokenizer.model_max_length = 1300 # self.tokenizer.max_length = 1300 self.model.resize_token_embeddings(len(self.tokenizer)) print(len(self.tokenizer)) use_task_specific_params(self.model, "summarization") # save_git_info(self.hparams.output_dir) self.metrics_save_path = Path(self.output_dir) / "metrics.json" self.hparams_save_path = Path(self.output_dir) / "hparams.pkl" pickle_save(self.hparams, self.hparams_save_path) self.step_count = 0 self.metrics = defaultdict(list) self.model_type = self.config.model_type self.vocab_size = self.config.tgt_vocab_size if self.model_type == "fsmt" else self.config.vocab_size self.dataset_kwargs: dict = dict( data_dir=self.hparams.data_dir, max_source_length=self.hparams.max_source_length, prefix=self.model.config.prefix or "", ) n_observations_per_split = { "train": self.hparams.n_train, "val": self.hparams.n_val, "test": self.hparams.n_test, } self.n_obs = { k: v if v >= 0 else None for k, v in n_observations_per_split.items() } self.target_lens = { "train": self.hparams.max_target_length, "val": self.hparams.val_max_target_length, "test": self.hparams.test_max_target_length, } assert self.target_lens["train"] <= self.target_lens[ "val"], f"target_lens: {self.target_lens}" assert self.target_lens["train"] <= self.target_lens[ "test"], f"target_lens: {self.target_lens}" if self.hparams.freeze_embeds: self.freeze_embeds() if self.hparams.freeze_encoder: freeze_params(self.model.get_encoder()) assert_all_frozen(self.model.get_encoder()) # self.hparams.git_sha = get_git_info()["repo_sha"] self.num_workers = hparams.num_workers self.decoder_start_token_id = None # default to config if self.model.config.decoder_start_token_id is None and isinstance( self.tokenizer, MBartTokenizer): self.decoder_start_token_id = self.tokenizer.lang_code_to_id[ hparams.tgt_lang] self.model.config.decoder_start_token_id = self.decoder_start_token_id self.dataset_class = (Seq2SeqDataset if hasattr( self.tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset) self.eval_beams = self.model.config.num_beams if self.hparams.eval_beams is None else self.hparams.eval_beams assert self.eval_beams >= 1, f"got self.eval_beams={self.eval_beams}. Need an integer > 1" if self.hparams.eval_max_gen_length is not None: self.eval_max_length = self.hparams.eval_max_gen_length else: self.eval_max_length = self.model.config.max_length self.val_metric = self.default_val_metric if self.hparams.val_metric is None else self.hparams.val_metric self.get_freq_sequences(self.hparams.data_dir) self.seq_loss_weight = 2
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() if ( os.path.exists(training_args.output_dir) and os.listdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir ): raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome." ) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN, ) logger.warning( "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", training_args.local_rank, training_args.device, training_args.n_gpu, bool(training_args.local_rank != -1), training_args.fp16, ) logger.info("Training/evaluation parameters %s", training_args) # Set seed set_seed(training_args.seed) # 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, ) tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, ) model = AutoModelForSeq2SeqLM.from_pretrained( model_args.model_name_or_path, from_tf=".ckpt" in model_args.model_name_or_path, config=config, cache_dir=model_args.cache_dir, ) # use task specific params use_task_specific_params(model, data_args.task) # set num_beams for evaluation if data_args.eval_beams is None: data_args.eval_beams = model.config.num_beams # set decoder_start_token_id for MBart if model.config.decoder_start_token_id is None and isinstance(tokenizer, MBartTokenizer): assert ( data_args.tgt_lang is not None and data_args.src_lang is not None ), "mBart requires --tgt_lang and --src_lang" model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang] if model_args.freeze_embeds: freeze_embeds(model) if model_args.freeze_encoder: freeze_params(model.get_encoder()) assert_all_frozen(model.get_encoder()) dataset_class = Seq2SeqDataset if hasattr(tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset # Get datasets train_dataset = ( dataset_class( tokenizer, type_path="train", data_dir=data_args.data_dir, n_obs=data_args.n_train, max_target_length=data_args.max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_train else None ) eval_dataset = ( dataset_class( tokenizer, type_path="val", data_dir=data_args.data_dir, n_obs=data_args.n_val, max_target_length=data_args.val_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_eval or training_args.evaluation_strategy != EvaluationStrategy.NO else None ) test_dataset = ( dataset_class( tokenizer, type_path="test", data_dir=data_args.data_dir, n_obs=data_args.n_test, max_target_length=data_args.test_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_predict else None ) # Initialize our Trainer compute_metrics_fn = ( build_compute_metrics_fn(data_args.task, tokenizer) if training_args.predict_with_generate else None ) trainer = Seq2SeqTrainer( model=model, config=config, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset, data_collator=Seq2SeqDataCollator(tokenizer, data_args, training_args.tpu_num_cores), compute_metrics=compute_metrics_fn, data_args=data_args, ) # Training if training_args.do_train: trainer.train( model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None ) trainer.save_model() # For convenience, we also re-save the tokenizer to the same directory, # so that you can share your model easily on huggingface.co/models =) if trainer.is_world_process_zero(): trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json")) tokenizer.save_pretrained(training_args.output_dir) # Evaluation eval_results = {} if training_args.do_eval: logger.info("*** Evaluate ***") result = trainer.evaluate() if trainer.is_world_process_zero(): logger.info("***** Eval results *****") for key, value in result.items(): logger.info(" %s = %s", key, value) save_json(result, os.path.join(training_args.output_dir, "eval_results.json")) eval_results.update(result) if training_args.do_predict: logging.info("*** Test ***") test_output = trainer.predict(test_dataset=test_dataset) test_metrics = {k.replace("eval", "test"): v for k, v in test_output.metrics.items()} if trainer.is_world_process_zero(): logger.info("***** Test results *****") for key, value in test_metrics.items(): logger.info(" %s = %s", key, value) save_json(test_metrics, os.path.join(training_args.output_dir, "test_results.json")) eval_results.update(test_metrics) if training_args.predict_with_generate: test_preds = tokenizer.batch_decode( test_output.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True ) test_preds = lmap(str.strip, test_preds) write_txt_file(test_preds, os.path.join(training_args.output_dir, "test_generations.txt")) if trainer.is_world_process_zero(): save_json(eval_results, "all_results.json") return eval_results
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( ) if (os.path.exists(training_args.output_dir) and os.listdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir): raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome." ) # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN, ) logger.warning( "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", training_args.local_rank, training_args.device, training_args.n_gpu, bool(training_args.local_rank != -1), training_args.fp16, ) logger.info("Training/evaluation parameters %s", training_args) # Set seed set_seed(training_args.seed) # 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, ) tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, cache_dir=model_args.cache_dir, ) model = AutoModelForSeq2SeqLM.from_pretrained( model_args.model_name_or_path, from_tf=".ckpt" in model_args.model_name_or_path, config=config, cache_dir=model_args.cache_dir, ) # use task specific params use_task_specific_params(model, data_args.task) # set num_beams for evaluation if data_args.eval_beams is not None: model.config.num_beams = data_args.eval_beams assert model.config.num_beams >= 1, f"got eval_beams={model.config.num_beams}. Need an integer >= 1" # set max length for generation model.config.max_generate_length = data_args.val_max_target_length # set decoder_start_token_id for MBart if model.config.decoder_start_token_id is None and isinstance( tokenizer, MBartTokenizer): decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang] model.config.decoder_start_token_id = decoder_start_token_id def build_compute_metrics_fn( task_name: str) -> Callable[[EvalPrediction], Dict]: def non_pad_len(tokens: np.ndarray) -> int: return np.count_nonzero(tokens != tokenizer.pad_token_id) def decode_pred(pred: EvalPrediction) -> Tuple[List[str], List[str]]: pred_str = tokenizer.batch_decode(pred.predictions, skip_special_tokens=True) label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True) pred_str = lmap(str.strip, pred_str) label_str = lmap(str.strip, label_str) return pred_str, label_str def summarization_metrics(pred: EvalPrediction) -> Dict: pred_str, label_str = decode_pred(pred) rouge: Dict = calculate_rouge(pred_str, label_str) summ_len = np.mean(lmap(non_pad_len, pred.predictions)) rouge.update({"gen_len": summ_len}) return rouge def translation_metrics(pred: EvalPrediction) -> Dict: pred_str, label_str = decode_pred(pred) bleu: Dict = calculate_bleu(pred_str, label_str) gen_len = np.mean(lmap(non_pad_len, pred.predictions)) bleu.update({"gen_len": gen_len}) return bleu compute_metrics_fn = summarization_metrics if "summarization" in task_name else translation_metrics return compute_metrics_fn def freeze_embeds(model: torch.nn.Module): """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5.""" try: freeze_params(model.model.shared) for d in [model.model.encoder, model.model.decoder]: freeze_params(d.embed_positions) freeze_params(d.embed_tokens) except AttributeError: freeze_params(model.shared) for d in [model.encoder, model.decoder]: freeze_params(d.embed_tokens) if model_args.freeze_embeds: freeze_embeds(model) if model_args.freeze_encoder: freeze_params(model.get_encoder()) assert_all_frozen(model.get_encoder()) dataset_class = Seq2SeqDataset if hasattr( tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset # Get datasets train_dataset = (dataset_class( tokenizer, type_path="train", data_dir=data_args.data_dir, n_obs=data_args.n_train, max_target_length=data_args.max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_train else None) eval_dataset = (dataset_class( tokenizer, type_path="val", data_dir=data_args.data_dir, n_obs=data_args.n_val, max_target_length=data_args.val_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_eval else None) test_dataset = (dataset_class( tokenizer, type_path="test", data_dir=data_args.data_dir, n_obs=data_args.n_test, max_target_length=data_args.test_max_target_length, max_source_length=data_args.max_source_length, prefix=model.config.prefix or "", ) if training_args.do_predict else None) # Initialize our Trainer trainer = Seq2SeqTrainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset, data_collator=Seq2SeqDataCollator(tokenizer, data_args, training_args.tpu_num_cores), compute_metrics=build_compute_metrics_fn(data_args.task) if training_args.predict_with_generate else None, ) # Training if training_args.do_train: trainer.train(model_path=model_args.model_name_or_path if os.path. isdir(model_args.model_name_or_path) else None) trainer.save_model() # For convenience, we also re-save the tokenizer to the same directory, # so that you can share your model easily on huggingface.co/models =) if trainer.is_world_process_zero(): tokenizer.save_pretrained(training_args.output_dir) # Evaluation eval_results = {} if training_args.do_eval: logger.info("*** Evaluate ***") result = trainer.evaluate() output_eval_file = os.path.join(training_args.output_dir, "eval_results.json") if trainer.is_world_process_zero(): logger.info("***** Eval results *****") for key, value in result.items(): logger.info(" %s = %s", key, value) with open(output_eval_file, "w") as f: json.dump(result, f) eval_results.update(result) if training_args.do_predict: logging.info("*** Test ***") test_output = trainer.predict(test_dataset=test_dataset) test_metrics = test_output.metrics test_metrics = { k.replace("eval", "test"): v for k, v in test_metrics.items() } output_test_file = os.path.join(training_args.output_dir, "test_results.json") if trainer.is_world_process_zero(): logger.info("***** Test results *****") for key, value in test_metrics.items(): logger.info(" %s = %s", key, value) with open(output_test_file, "w") as f: json.dump(test_metrics, f) if training_args.predict_with_generate: test_preds = tokenizer.batch_decode(test_output.predictions, skip_special_tokens=True) test_preds = lmap(str.strip, test_preds) output_test_pred_file = os.path.join(training_args.output_dir, "test_generations.txt") with open(output_test_pred_file, "w") as f: f.write("\n".join(test_preds)) return eval_results