def evaluate(self, test_df: pd.DataFrame, metrics: str = "rouge"): """ :param test_df: :param metrics: :return: Output metrics for keytotext """ metric = load_metric(metrics) input_text = test_df["input_text"] references = test_df["output_text"] predictions = [self.predict(x) for x in input_text] results = metric.compute(predictions=predictions, references=references) output = { "Rouge 1": { "Rouge_1 Low Precision": results["rouge1"].low.precision, "Rouge_1 Low recall": results["rouge1"].low.recall, "Rouge_1 Low F1": results["rouge1"].low.fmeasure, "Rouge_1 Mid Precision": results["rouge1"].mid.precision, "Rouge_1 Mid recall": results["rouge1"].mid.recall, "Rouge_1 Mid F1": results["rouge1"].mid.fmeasure, "Rouge_1 High Precision": results["rouge1"].high.precision, "Rouge_1 High recall": results["rouge1"].high.recall, "Rouge_1 High F1": results["rouge1"].high.fmeasure, }, "Rouge 2": { "Rouge_2 Low Precision": results["rouge2"].low.precision, "Rouge_2 Low recall": results["rouge2"].low.recall, "Rouge_2 Low F1": results["rouge2"].low.fmeasure, "Rouge_2 Mid Precision": results["rouge2"].mid.precision, "Rouge_2 Mid recall": results["rouge2"].mid.recall, "Rouge_2 Mid F1": results["rouge2"].mid.fmeasure, "Rouge_2 High Precision": results["rouge2"].high.precision, "Rouge_2 High recall": results["rouge2"].high.recall, "Rouge_2 High F1": results["rouge2"].high.fmeasure, }, "Rouge L": { "Rouge_L Low Precision": results["rougeL"].low.precision, "Rouge_L Low recall": results["rougeL"].low.recall, "Rouge_L Low F1": results["rougeL"].low.fmeasure, "Rouge_L Mid Precision": results["rougeL"].mid.precision, "Rouge_L Mid recall": results["rougeL"].mid.recall, "Rouge_L Mid F1": results["rougeL"].mid.fmeasure, "Rouge_L High Precision": results["rougeL"].high.precision, "Rouge_L High recall": results["rougeL"].high.recall, "Rouge_L High F1": results["rougeL"].high.fmeasure, }, "rougeLsum": { "rougeLsum Low Precision": results["rougeLsum"].low.precision, "rougeLsum Low recall": results["rougeLsum"].low.recall, "rougeLsum Low F1": results["rougeLsum"].low.fmeasure, "rougeLsum Mid Precision": results["rougeLsum"].mid.precision, "rougeLsum Mid recall": results["rougeLsum"].mid.recall, "rougeLsum Mid F1": results["rougeLsum"].mid.fmeasure, "rougeLsum High Precision": results["rougeLsum"].high.precision, "rougeLsum High recall": results["rougeLsum"].high.recall, "rougeLsum High F1": results["rougeLsum"].high.fmeasure, }, } return output
def main(): args = parse_args() # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. # If we're using tracking, we also need to initialize it here and it will pick up all supported trackers in the environment accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator() logger.info(accelerator.state) # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state, main_process_only=False) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Handle the repository creation if accelerator.is_main_process: if args.push_to_hub: if args.hub_model_id is None: repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token) else: repo_name = args.hub_model_id repo = Repository(args.output_dir, clone_from=repo_name) with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore: if "step_*" not in gitignore: gitignore.write("step_*\n") if "epoch_*" not in gitignore: gitignore.write("epoch_*\n") elif args.output_dir is not None: os.makedirs(args.output_dir, exist_ok=True) accelerator.wait_for_everyone() # Get the datasets: you can either provide your own training and evaluation files (see below) # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub). # In distributed training, the load_dataset function guarantees that only one local process can concurrently # download the dataset. if args.dataset_name is not None: # Downloading and loading a dataset from the hub. dataset = load_dataset(args.dataset_name, task="image-classification") else: data_files = {} if args.train_dir is not None: data_files["train"] = os.path.join(args.train_dir, "**") if args.validation_dir is not None: data_files["validation"] = os.path.join(args.validation_dir, "**") dataset = load_dataset( "imagefolder", data_files=data_files, cache_dir=args.cache_dir, task="image-classification", ) # See more about loading custom images at # https://huggingface.co/docs/datasets/v2.0.0/en/image_process#imagefolder. # If we don't have a validation split, split off a percentage of train as validation. args.train_val_split = None if "validation" in dataset.keys() else args.train_val_split if isinstance(args.train_val_split, float) and args.train_val_split > 0.0: split = dataset["train"].train_test_split(args.train_val_split) dataset["train"] = split["train"] dataset["validation"] = split["test"] # Prepare label mappings. # We'll include these in the model's config to get human readable labels in the Inference API. labels = dataset["train"].features["labels"].names label2id = {label: str(i) for i, label in enumerate(labels)} id2label = {str(i): label for i, label in enumerate(labels)} # Load pretrained model and feature extractor # # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. config = AutoConfig.from_pretrained( args.model_name_or_path, num_labels=len(labels), i2label=id2label, label2id=label2id, finetuning_task="image-classification", ) feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_name_or_path) model = AutoModelForImageClassification.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ignore_mismatched_sizes=args.ignore_mismatched_sizes, ) # Preprocessing the datasets # Define torchvision transforms to be applied to each image. normalize = Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std) train_transforms = Compose( [ RandomResizedCrop(feature_extractor.size), RandomHorizontalFlip(), ToTensor(), normalize, ] ) val_transforms = Compose( [ Resize(feature_extractor.size), CenterCrop(feature_extractor.size), ToTensor(), normalize, ] ) def preprocess_train(example_batch): """Apply _train_transforms across a batch.""" example_batch["pixel_values"] = [train_transforms(image.convert("RGB")) for image in example_batch["image"]] return example_batch def preprocess_val(example_batch): """Apply _val_transforms across a batch.""" example_batch["pixel_values"] = [val_transforms(image.convert("RGB")) for image in example_batch["image"]] return example_batch with accelerator.main_process_first(): if args.max_train_samples is not None: dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) # Set the training transforms train_dataset = dataset["train"].with_transform(preprocess_train) if args.max_eval_samples is not None: dataset["validation"] = dataset["validation"].shuffle(seed=args.seed).select(range(args.max_eval_samples)) # Set the validation transforms eval_dataset = dataset["validation"].with_transform(preprocess_val) # DataLoaders creation: def collate_fn(examples): pixel_values = torch.stack([example["pixel_values"] for example in examples]) labels = torch.tensor([example["labels"] for example in examples]) return {"pixel_values": pixel_values, "labels": labels} train_dataloader = DataLoader( train_dataset, shuffle=True, collate_fn=collate_fn, batch_size=args.per_device_train_batch_size ) eval_dataloader = DataLoader(eval_dataset, collate_fn=collate_fn, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], "weight_decay": args.weight_decay, }, { "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0, }, ] optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Scheduler and math around the number of training steps. num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch else: args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader, lr_scheduler ) # We need to recalculate our total training steps as the size of the training dataloader may have changed. num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch # Figure out how many steps we should save the Accelerator states if hasattr(args.checkpointing_steps, "isdigit"): checkpointing_steps = args.checkpointing_steps if args.checkpointing_steps.isdigit(): checkpointing_steps = int(args.checkpointing_steps) else: checkpointing_steps = None # We need to initialize the trackers we use, and also store our configuration if args.with_tracking: experiment_config = vars(args) # TensorBoard cannot log Enums, need the raw value experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value accelerator.init_trackers("image_classification_no_trainer", experiment_config) # Get the metric function metric = load_metric("accuracy") # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}") logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 starting_epoch = 0 # Potentially load in the weights and states from a previous save if args.resume_from_checkpoint: if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "": accelerator.print(f"Resumed from checkpoint: {args.resume_from_checkpoint}") accelerator.load_state(args.resume_from_checkpoint) path = os.path.basename(args.resume_from_checkpoint) else: # Get the most recent checkpoint dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()] dirs.sort(key=os.path.getctime) path = dirs[-1] # Sorts folders by date modified, most recent checkpoint is the last # Extract `epoch_{i}` or `step_{i}` training_difference = os.path.splitext(path)[0] if "epoch" in training_difference: starting_epoch = int(training_difference.replace("epoch_", "")) + 1 resume_step = None else: resume_step = int(training_difference.replace("step_", "")) starting_epoch = resume_step // len(train_dataloader) resume_step -= starting_epoch * len(train_dataloader) for epoch in range(starting_epoch, args.num_train_epochs): model.train() if args.with_tracking: total_loss = 0 for step, batch in enumerate(train_dataloader): # We need to skip steps until we reach the resumed step if args.resume_from_checkpoint and epoch == starting_epoch: if resume_step is not None and step < resume_step: completed_steps += 1 continue outputs = model(**batch) loss = outputs.loss # We keep track of the loss at each epoch if args.with_tracking: total_loss += loss.detach().float() loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if isinstance(checkpointing_steps, int): if completed_steps % checkpointing_steps == 0: output_dir = f"step_{completed_steps }" if args.output_dir is not None: output_dir = os.path.join(args.output_dir, output_dir) accelerator.save_state(output_dir) if args.push_to_hub and epoch < args.num_train_epochs - 1: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained( args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save, ) if accelerator.is_main_process: feature_extractor.save_pretrained(args.output_dir) repo.push_to_hub( commit_message=f"Training in progress {completed_steps} steps", blocking=False, auto_lfs_prune=True, ) if completed_steps >= args.max_train_steps: break model.eval() samples_seen = 0 for step, batch in enumerate(eval_dataloader): with torch.no_grad(): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) predictions, references = accelerator.gather((predictions, batch["labels"])) # If we are in a multiprocess environment, the last batch has duplicates if accelerator.num_processes > 1: if step == len(eval_dataloader) - 1: predictions = predictions[: len(eval_dataloader.dataset) - samples_seen] references = references[: len(eval_dataloader.dataset) - samples_seen] else: samples_seen += references.shape[0] metric.add_batch( predictions=predictions, references=references, ) eval_metric = metric.compute() logger.info(f"epoch {epoch}: {eval_metric}") if args.with_tracking: accelerator.log( { "accuracy": eval_metric, "train_loss": total_loss, "epoch": epoch, "step": completed_steps, }, ) if args.push_to_hub and epoch < args.num_train_epochs - 1: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained( args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save ) if accelerator.is_main_process: feature_extractor.save_pretrained(args.output_dir) repo.push_to_hub( commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True ) if args.checkpointing_steps == "epoch": output_dir = f"epoch_{epoch}" if args.output_dir is not None: output_dir = os.path.join(args.output_dir, output_dir) accelerator.save_state(output_dir) if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained( args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save ) if accelerator.is_main_process: feature_extractor.save_pretrained(args.output_dir) if args.push_to_hub: repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True) if args.output_dir is not None: with open(os.path.join(args.output_dir, "all_results.json"), "w") as f: json.dump({"eval_accuracy": eval_metric["accuracy"]}, f)
def main(): # region Argument parsing # 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, TFTrainingArguments)) 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() output_dir = Path(training_args.output_dir) output_dir.mkdir(parents=True, exist_ok=True) # endregion # region Checkpoints checkpoint = None if len(os.listdir(training_args.output_dir)) > 0 and not training_args.overwrite_output_dir: if (output_dir / CONFIG_NAME).is_file() and (output_dir / TF2_WEIGHTS_NAME).is_file(): checkpoint = output_dir logger.info( f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this" " behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch." ) else: raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty. " "Use --overwrite_output_dir to continue regardless." ) # endregion # region 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 training_args.should_log else logging.WARN) # Set the verbosity to info of the Transformers logger (on main process only): if training_args.should_log: transformers.utils.logging.set_verbosity_info() transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() logger.info(f"Training/evaluation parameters {training_args}") # endregion # Set seed before initializing model. set_seed(training_args.seed) # region Load Data # Get the datasets: you can either provide your own CSV/JSON/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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. 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, field="data", 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. # endregion # region 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, ) 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, use_fast=True, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) # endregion # region Tokenizer check: this script requires a fast tokenizer. if not isinstance(tokenizer, PreTrainedTokenizerFast): raise ValueError( "This example script only works for models that have a fast tokenizer. Checkout the big table of models " "at https://huggingface.co/transformers/index.html#supported-frameworks to find the model types that meet this " "requirement" ) # endregion # region Preprocessing the datasets # Preprocessing is slightly different for training and evaluation. if training_args.do_train: column_names = datasets["train"].column_names elif training_args.do_eval: column_names = datasets["validation"].column_names else: column_names = datasets["test"].column_names question_column_name = "question" if "question" in column_names else column_names[0] context_column_name = "context" if "context" in column_names else column_names[1] answer_column_name = "answers" if "answers" in column_names else column_names[2] # Padding side determines if we do (question|context) or (context|question). pad_on_right = tokenizer.padding_side == "right" 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) if data_args.pad_to_max_length or isinstance(training_args.strategy, tf.distribute.TPUStrategy): logger.info("Padding all batches to max length because argument was set or we're on TPU.") padding = "max_length" else: padding = False # Training preprocessing def prepare_train_features(examples): # Some of the questions have lots of whitespace on the left, which is not useful and will make the # truncation of the context fail (the tokenized question will take a lots of space). So we remove that # left whitespace examples[question_column_name] = [q.lstrip() for q in examples[question_column_name]] # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[question_column_name if pad_on_right else context_column_name], examples[context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding=padding, ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # The offset mappings will give us a map from token to character position in the original context. This will # help us compute the start_positions and end_positions. offset_mapping = tokenized_examples.pop("offset_mapping") # Let's label those examples! tokenized_examples["start_positions"] = [] tokenized_examples["end_positions"] = [] for i, offsets in enumerate(offset_mapping): # We will label impossible answers with the index of the CLS token. input_ids = tokenized_examples["input_ids"][i] cls_index = input_ids.index(tokenizer.cls_token_id) # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] answers = examples[answer_column_name][sample_index] # If no answers are given, set the cls_index as answer. if len(answers["answer_start"]) == 0: tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Start/end character index of the answer in the text. start_char = answers["answer_start"][0] end_char = start_char + len(answers["text"][0]) # Start token index of the current span in the text. token_start_index = 0 while sequence_ids[token_start_index] != (1 if pad_on_right else 0): token_start_index += 1 # End token index of the current span in the text. token_end_index = len(input_ids) - 1 while sequence_ids[token_end_index] != (1 if pad_on_right else 0): token_end_index -= 1 # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index). if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char): tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Otherwise move the token_start_index and token_end_index to the two ends of the answer. # Note: we could go after the last offset if the answer is the last word (edge case). while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char: token_start_index += 1 tokenized_examples["start_positions"].append(token_start_index - 1) while offsets[token_end_index][1] >= end_char: token_end_index -= 1 tokenized_examples["end_positions"].append(token_end_index + 1) return tokenized_examples processed_datasets = dict() 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: # We will select sample from whole data if agument is specified train_dataset = train_dataset.select(range(data_args.max_train_samples)) # Create train feature from dataset train_dataset = train_dataset.map( prepare_train_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if data_args.max_train_samples is not None: # Number of samples might increase during Feature Creation, We select only specified max samples train_dataset = train_dataset.select(range(data_args.max_train_samples)) processed_datasets["train"] = train_dataset # Validation preprocessing def prepare_validation_features(examples): # Some of the questions have lots of whitespace on the left, which is not useful and will make the # truncation of the context fail (the tokenized question will take a lots of space). So we remove that # left whitespace examples[question_column_name] = [q.lstrip() for q in examples[question_column_name]] # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[question_column_name if pad_on_right else context_column_name], examples[context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding=padding, ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the # corresponding example_id and we will store the offset mappings. tokenized_examples["example_id"] = [] for i in range(len(tokenized_examples["input_ids"])): # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) context_index = 1 if pad_on_right else 0 # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] tokenized_examples["example_id"].append(examples["id"][sample_index]) # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token # position is part of the context or not. tokenized_examples["offset_mapping"][i] = [ (o if sequence_ids[k] == context_index else None) for k, o in enumerate(tokenized_examples["offset_mapping"][i]) ] return tokenized_examples if training_args.do_eval: if "validation" not in datasets: raise ValueError("--do_eval requires a validation dataset") eval_examples = datasets["validation"] if data_args.max_eval_samples is not None: # We will select sample from whole data eval_examples = eval_examples.select(range(data_args.max_eval_samples)) # Validation Feature Creation eval_dataset = eval_examples.map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if data_args.max_eval_samples is not None: # During Feature creation dataset samples might increase, we will select required samples again eval_dataset = eval_dataset.select(range(data_args.max_eval_samples)) processed_datasets["validation"] = eval_dataset if training_args.do_predict: if "test" not in datasets: raise ValueError("--do_predict requires a test dataset") predict_examples = datasets["test"] if data_args.max_predict_samples is not None: # We will select sample from whole data predict_examples = predict_examples.select(range(data_args.max_predict_samples)) # Predict Feature Creation predict_dataset = predict_examples.map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if data_args.max_predict_samples is not None: # During Feature creation dataset samples might increase, we will select required samples again predict_dataset = predict_dataset.select(range(data_args.max_predict_samples)) processed_datasets["test"] = predict_dataset # endregion # region Metrics and Post-processing: def post_processing_function(examples, features, predictions, stage="eval"): # Post-processing: we match the start logits and end logits to answers in the original context. predictions = postprocess_qa_predictions( examples=examples, features=features, predictions=predictions, version_2_with_negative=data_args.version_2_with_negative, n_best_size=data_args.n_best_size, max_answer_length=data_args.max_answer_length, null_score_diff_threshold=data_args.null_score_diff_threshold, output_dir=training_args.output_dir, prefix=stage, ) # Format the result to the format the metric expects. if data_args.version_2_with_negative: formatted_predictions = [ {"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in predictions.items() ] else: formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()] references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples] return EvalPrediction(predictions=formatted_predictions, label_ids=references) metric = load_metric("squad_v2" if data_args.version_2_with_negative else "squad") def compute_metrics(p: EvalPrediction): return metric.compute(predictions=p.predictions, references=p.label_ids) # endregion with training_args.strategy.scope(): # region Load model if checkpoint is None: model_path = model_args.model_name_or_path else: model_path = checkpoint model = TFAutoModelForQuestionAnswering.from_pretrained( model_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, ) optimizer = tf.keras.optimizers.Adam( learning_rate=training_args.learning_rate, beta_1=training_args.adam_beta1, beta_2=training_args.adam_beta2, epsilon=training_args.adam_epsilon, clipnorm=training_args.max_grad_norm, ) # no user-specified loss = will use the model internal loss model.compile(optimizer=optimizer) # endregion # region Training if padding: data_collator = DefaultDataCollator(return_tensors="tf") else: data_collator = DataCollatorWithPadding(tokenizer, return_tensors="tf") tensor_keys = ["attention_mask", "input_ids"] label_keys = ["start_positions", "end_positions"] if training_args.do_train: # Make a tf.data.Dataset for this training_dataset = processed_datasets["train"].to_tf_dataset( # labels are passed as input, as we will use the model's internal loss columns=tensor_keys + label_keys, shuffle=True, batch_size=training_args.per_device_train_batch_size, collate_fn=data_collator, drop_remainder=True, ) model.fit(training_dataset, epochs=int(training_args.num_train_epochs)) # endregion # region Evaluation if training_args.do_eval: logger.info("*** Evaluation ***") eval_inputs = { "input_ids": tf.ragged.constant(processed_datasets["validation"]["input_ids"]).to_tensor(), "attention_mask": tf.ragged.constant(processed_datasets["validation"]["attention_mask"]).to_tensor(), } eval_predictions = model.predict(eval_inputs) post_processed_eval = post_processing_function( datasets["validation"], processed_datasets["validation"], (eval_predictions.start_logits, eval_predictions.end_logits), ) metrics = compute_metrics(post_processed_eval) logging.info("Evaluation metrics:") for metric, value in metrics.items(): logging.info(f"{metric}: {value:.3f}") # endregion # region Prediction if training_args.do_predict: logger.info("*** Predict ***") predict_inputs = { "input_ids": tf.ragged.constant(processed_datasets["test"]["input_ids"]).to_tensor(), "attention_mask": tf.ragged.constant(processed_datasets["test"]["attention_mask"]).to_tensor(), } test_predictions = model.predict(predict_inputs) post_processed_test = post_processing_function( datasets["test"], processed_datasets["test"], (test_predictions.start_logits, test_predictions.end_logits), ) metrics = compute_metrics(post_processed_test) logging.info("Test metrics:") for metric, value in metrics.items(): logging.info(f"{metric}: {value:.3f}") # endregion if training_args.push_to_hub: model.push_to_hub()
def main(): args = parse_args() # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. accelerator = Accelerator() # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state) # Setup logging, we only want one process per machine to log things on the screen. # accelerator.is_local_main_process is only True for one process per machine. logger.setLevel( logging.INFO if accelerator.is_local_main_process else logging.ERROR) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Get the datasets: you can either provide your own CSV/JSON/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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 args.dataset_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name) else: data_files = {} if args.train_file is not None: data_files["train"] = args.train_file if args.validation_file is not None: data_files["validation"] = args.validation_file extension = args.train_file.split(".")[-1] raw_datasets = load_dataset(extension, data_files=data_files) # Trim a number of training examples if args.debug: for split in raw_datasets.keys(): raw_datasets[split] = raw_datasets[split].select(range(100)) # 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. if raw_datasets["train"] is not None: column_names = raw_datasets["train"].column_names else: column_names = raw_datasets["validation"].column_names # When using your own dataset or a different dataset from swag, you will probably need to change this. ending_names = [f"ending{i}" for i in range(4)] context_name = "sent1" question_header_name = "sent2" label_column_name = "label" if "label" in column_names else "labels" # Load pretrained model and tokenizer # # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. if args.config_name: config = AutoConfig.from_pretrained(args.model_name_or_path) elif args.model_name_or_path: config = AutoConfig.from_pretrained(args.model_name_or_path) else: config = CONFIG_MAPPING[args.model_type]() logger.warning( "You are instantiating a new config instance from scratch.") if args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained( args.tokenizer_name, use_fast=not args.use_slow_tokenizer) elif args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained( args.model_name_or_path, use_fast=not args.use_slow_tokenizer) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if args.model_name_or_path: model = AutoModelForMultipleChoice.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ) else: logger.info("Training new model from scratch") model = AutoModelForMultipleChoice.from_config(config) model.resize_token_embeddings(len(tokenizer)) # Preprocessing the datasets. # First we tokenize all the texts. padding = "max_length" if args.pad_to_max_length else False def preprocess_function(examples): first_sentences = [[context] * 4 for context in examples[context_name]] question_headers = examples[question_header_name] second_sentences = [[ f"{header} {examples[end][i]}" for end in ending_names ] for i, header in enumerate(question_headers)] labels = examples[label_column_name] # Flatten out first_sentences = sum(first_sentences, []) second_sentences = sum(second_sentences, []) # Tokenize tokenized_examples = tokenizer( first_sentences, second_sentences, max_length=args.max_length, padding=padding, truncation=True, ) # Un-flatten tokenized_inputs = { k: [v[i:i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items() } tokenized_inputs["labels"] = labels return tokenized_inputs processed_datasets = raw_datasets.map( preprocess_function, batched=True, remove_columns=raw_datasets["train"].column_names) train_dataset = processed_datasets["train"] eval_dataset = processed_datasets["validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # DataLoaders creation: if args.pad_to_max_length: # If padding was already done ot max length, we use the default data collator that will just convert everything # to tensors. data_collator = default_data_collator else: # Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of # the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple # of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta). data_collator = DataCollatorForMultipleChoice( tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)) train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size) eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [ p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) ], "weight_decay": args.weight_decay, }, { "params": [ p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) ], "weight_decay": 0.0, }, ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Use the device given by the `accelerator` object. device = accelerator.device model.to(device) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader) # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be # shorter in multiprocess) # Scheduler and math around the number of training steps. num_update_steps_per_epoch = math.ceil( len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch else: args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Metrics metric = load_metric("accuracy") # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info( f" Instantaneous batch size per device = {args.per_device_train_batch_size}" ) logger.info( f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}" ) logger.info( f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 for epoch in range(args.num_train_epochs): model.train() for step, batch in enumerate(train_dataloader): outputs = model(**batch) loss = outputs.loss loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len( train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if completed_steps >= args.max_train_steps: break model.eval() for step, batch in enumerate(eval_dataloader): with torch.no_grad(): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) metric.add_batch( predictions=accelerator.gather(predictions), references=accelerator.gather(batch["labels"]), ) eval_metric = metric.compute() accelerator.print(f"epoch {epoch}: {eval_metric}") if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
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)) 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: 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 distant debugging if needed if data_args.server_ip and data_args.server_port: # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd print("Waiting for debugger attach") ptvsd.enable_attach(address=(data_args.server_ip, data_args.server_port), redirect_output=True) ptvsd.wait_for_attach() # 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() transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() logger.info(f"Training/evaluation parameters {training_args}") # Set seed before initializing model. set_seed(training_args.seed) # In distributed training, the load_dataset function guarantees that only one local process can concurrently # download the dataset. # Downloading and loading xnli dataset from the hub. if model_args.train_language is None: train_dataset = load_dataset("xnli", model_args.language, split="train") else: train_dataset = load_dataset("xnli", model_args.train_language, split="train") eval_dataset = load_dataset("xnli", model_args.language, split="validation") # Labels label_list = train_dataset.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, finetuning_task="xnli", cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path, do_lower_case=model_args.do_lower_case, 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, ) model = AutoModelForSequenceClassification.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, ) # Preprocessing the datasets # 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 def preprocess_function(examples): # Tokenize the texts return tokenizer( examples["premise"], examples["hypothesis"], padding=padding, max_length=data_args.max_seq_length, truncation=True, ) if training_args.do_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_function, batched=True, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_eval: if data_args.max_val_samples is not None: eval_dataset = eval_dataset.select(range(data_args.max_val_samples)) eval_dataset = eval_dataset.map( preprocess_function, batched=True, load_from_cache_file=not data_args.overwrite_cache, ) # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info(f"Sample {index} of the training set: {train_dataset[index]}.") # Get the metric function metric = load_metric("xnli") # 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 metric.compute(predictions=preds, references=p.label_ids) # 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, 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: if last_checkpoint is not None: model_path = last_checkpoint elif os.path.isdir(model_args.model_name_or_path): model_path = model_args.model_name_or_path else: model_path = None train_result = trainer.train(model_path=model_path) 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_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset) metrics["eval_samples"] = min(max_val_samples, len(eval_dataset)) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics)
def main(): args = parse_args() # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The # information sent is the one passed as arguments along with your Python/PyTorch versions. send_example_telemetry("run_glue_no_trainer", args) # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. # If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers # in the environment accelerator = ( Accelerator(log_with=args.report_to, logging_dir=args.output_dir) if args.with_tracking else Accelerator() ) # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state, main_process_only=False) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Handle the repository creation if accelerator.is_main_process: if args.push_to_hub: if args.hub_model_id is None: repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token) else: repo_name = args.hub_model_id repo = Repository(args.output_dir, clone_from=repo_name) with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore: if "step_*" not in gitignore: gitignore.write("step_*\n") if "epoch_*" not in gitignore: gitignore.write("epoch_*\n") elif args.output_dir is not None: os.makedirs(args.output_dir, exist_ok=True) accelerator.wait_for_everyone() # 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 CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named # label if at least two columns are provided. # 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 args.task_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset("glue", args.task_name) else: # Loading the dataset from local csv or json file. data_files = {} if args.train_file is not None: data_files["train"] = args.train_file if args.validation_file is not None: data_files["validation"] = args.validation_file extension = (args.train_file if args.train_file is not None else args.validation_file).split(".")[-1] raw_datasets = load_dataset(extension, data_files=data_files) # See more about loading any type of standard or custom dataset at # https://huggingface.co/docs/datasets/loading_datasets.html. # Labels if args.task_name is not None: is_regression = args.task_name == "stsb" if not is_regression: label_list = raw_datasets["train"].features["label"].names num_labels = len(label_list) else: num_labels = 1 else: # Trying to have good defaults here, don't hesitate to tweak to your needs. is_regression = raw_datasets["train"].features["label"].dtype in ["float32", "float64"] if is_regression: num_labels = 1 else: # A useful fast method: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique label_list = raw_datasets["train"].unique("label") label_list.sort() # Let's sort it for determinism 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(args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer) model = AutoModelForSequenceClassification.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ignore_mismatched_sizes=args.ignore_mismatched_sizes, ) # Preprocessing the datasets if args.task_name is not None: sentence1_key, sentence2_key = task_to_keys[args.task_name] else: # Again, we try to have some nice defaults but don't hesitate to tweak to your use case. non_label_column_names = [name for name in raw_datasets["train"].column_names if name != "label"] if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names: sentence1_key, sentence2_key = "sentence1", "sentence2" else: if len(non_label_column_names) >= 2: sentence1_key, sentence2_key = non_label_column_names[:2] else: sentence1_key, sentence2_key = non_label_column_names[0], None # Some models have set the order of the labels to use, so let's make sure we do use it. label_to_id = None if ( model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id and args.task_name is not None and not is_regression ): # Some have all caps in their config, some don't. label_name_to_id = {k.lower(): v for k, v in model.config.label2id.items()} if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)): logger.info( f"The configuration of the model provided the following label correspondence: {label_name_to_id}. " "Using it!" ) label_to_id = {i: label_name_to_id[label_list[i]] for i in range(num_labels)} else: logger.warning( "Your model seems to have been trained with labels, but they don't match the dataset: ", f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}." "\nIgnoring the model labels as a result.", ) elif args.task_name is None and not is_regression: label_to_id = {v: i for i, v in enumerate(label_list)} if label_to_id is not None: model.config.label2id = label_to_id model.config.id2label = {id: label for label, id in config.label2id.items()} elif args.task_name is not None and not is_regression: model.config.label2id = {l: i for i, l in enumerate(label_list)} model.config.id2label = {id: label for label, id in config.label2id.items()} padding = "max_length" if args.pad_to_max_length else False def preprocess_function(examples): # Tokenize the texts texts = ( (examples[sentence1_key],) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key]) ) result = tokenizer(*texts, padding=padding, max_length=args.max_length, truncation=True) if "label" in examples: if label_to_id is not None: # Map labels to IDs (not necessary for GLUE tasks) result["labels"] = [label_to_id[l] for l in examples["label"]] else: # In all cases, rename the column to labels because the model will expect that. result["labels"] = examples["label"] return result with accelerator.main_process_first(): processed_datasets = raw_datasets.map( preprocess_function, batched=True, remove_columns=raw_datasets["train"].column_names, desc="Running tokenizer on dataset", ) train_dataset = processed_datasets["train"] eval_dataset = processed_datasets["validation_matched" if args.task_name == "mnli" else "validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info(f"Sample {index} of the training set: {train_dataset[index]}.") # DataLoaders creation: if args.pad_to_max_length: # If padding was already done ot max length, we use the default data collator that will just convert everything # to tensors. data_collator = default_data_collator else: # Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of # the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple # of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta). data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)) train_dataloader = DataLoader( train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size ) eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], "weight_decay": args.weight_decay, }, { "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0, }, ] optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Scheduler and math around the number of training steps. overrode_max_train_steps = False num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch overrode_max_train_steps = True lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader, lr_scheduler ) # We need to recalculate our total training steps as the size of the training dataloader may have changed num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if overrode_max_train_steps: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch # Afterwards we recalculate our number of training epochs args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) # Figure out how many steps we should save the Accelerator states if hasattr(args.checkpointing_steps, "isdigit"): checkpointing_steps = args.checkpointing_steps if args.checkpointing_steps.isdigit(): checkpointing_steps = int(args.checkpointing_steps) else: checkpointing_steps = None # We need to initialize the trackers we use, and also store our configuration. # We initialize the trackers only on main process because `accelerator.log` # only logs on main process and we don't want empty logs/runs on other processes. if args.with_tracking: if accelerator.is_main_process: experiment_config = vars(args) # TensorBoard cannot log Enums, need the raw value experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value accelerator.init_trackers("glue_no_trainer", experiment_config) # Get the metric function if args.task_name is not None: metric = load_metric("glue", args.task_name) else: metric = load_metric("accuracy") # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}") logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 starting_epoch = 0 # Potentially load in the weights and states from a previous save if args.resume_from_checkpoint: if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "": accelerator.print(f"Resumed from checkpoint: {args.resume_from_checkpoint}") accelerator.load_state(args.resume_from_checkpoint) path = os.path.basename(args.resume_from_checkpoint) else: # Get the most recent checkpoint dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()] dirs.sort(key=os.path.getctime) path = dirs[-1] # Sorts folders by date modified, most recent checkpoint is the last # Extract `epoch_{i}` or `step_{i}` training_difference = os.path.splitext(path)[0] if "epoch" in training_difference: starting_epoch = int(training_difference.replace("epoch_", "")) + 1 resume_step = None else: resume_step = int(training_difference.replace("step_", "")) starting_epoch = resume_step // len(train_dataloader) resume_step -= starting_epoch * len(train_dataloader) for epoch in range(starting_epoch, args.num_train_epochs): model.train() if args.with_tracking: total_loss = 0 for step, batch in enumerate(train_dataloader): # We need to skip steps until we reach the resumed step if args.resume_from_checkpoint and epoch == starting_epoch: if resume_step is not None and step < resume_step: completed_steps += 1 continue outputs = model(**batch) loss = outputs.loss # We keep track of the loss at each epoch if args.with_tracking: total_loss += loss.detach().float() loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if isinstance(checkpointing_steps, int): if completed_steps % checkpointing_steps == 0: output_dir = f"step_{completed_steps }" if args.output_dir is not None: output_dir = os.path.join(args.output_dir, output_dir) accelerator.save_state(output_dir) if completed_steps >= args.max_train_steps: break model.eval() samples_seen = 0 for step, batch in enumerate(eval_dataloader): with torch.no_grad(): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) if not is_regression else outputs.logits.squeeze() predictions, references = accelerator.gather((predictions, batch["labels"])) # If we are in a multiprocess environment, the last batch has duplicates if accelerator.num_processes > 1: if step == len(eval_dataloader) - 1: predictions = predictions[: len(eval_dataloader.dataset) - samples_seen] references = references[: len(eval_dataloader.dataset) - samples_seen] else: samples_seen += references.shape[0] metric.add_batch( predictions=predictions, references=references, ) eval_metric = metric.compute() logger.info(f"epoch {epoch}: {eval_metric}") if args.with_tracking: accelerator.log( { "accuracy" if args.task_name is not None else "glue": eval_metric, "train_loss": total_loss.item() / len(train_dataloader), "epoch": epoch, "step": completed_steps, }, step=completed_steps, ) if args.push_to_hub and epoch < args.num_train_epochs - 1: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained( args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save ) if accelerator.is_main_process: tokenizer.save_pretrained(args.output_dir) repo.push_to_hub( commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True ) if args.checkpointing_steps == "epoch": output_dir = f"epoch_{epoch}" if args.output_dir is not None: output_dir = os.path.join(args.output_dir, output_dir) accelerator.save_state(output_dir) if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained( args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save ) if accelerator.is_main_process: tokenizer.save_pretrained(args.output_dir) if args.push_to_hub: repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True) if args.task_name == "mnli": # Final evaluation on mismatched validation set eval_dataset = processed_datasets["validation_mismatched"] eval_dataloader = DataLoader( eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size ) eval_dataloader = accelerator.prepare(eval_dataloader) model.eval() for step, batch in enumerate(eval_dataloader): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) metric.add_batch( predictions=accelerator.gather(predictions), references=accelerator.gather(batch["labels"]), ) eval_metric = metric.compute() logger.info(f"mnli-mm: {eval_metric}") if args.output_dir is not None: with open(os.path.join(args.output_dir, "all_results.json"), "w") as f: json.dump({"eval_accuracy": eval_metric["accuracy"]}, f)
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( ) # 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}") if data_args.source_prefix is None and model_args.model_name_or_path in [ "t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b", ]: logger.warning( "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with " "`--source_prefix 'summarize: ' `") # 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 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 CSV/JSON files this script will use the first column for the full texts and the second column for the # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments). # # 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, use_auth_token=True if model_args.use_auth_token else None, ) 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] raw_datasets = load_dataset( extension, data_files=data_files, cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) # 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, ) 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, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForSeq2SeqLM.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, ) model.resize_token_embeddings(len(tokenizer)) if model.config.decoder_start_token_id is None and isinstance( tokenizer, (MBartTokenizer, MBartTokenizerFast)): if isinstance(tokenizer, MBartTokenizer): model.config.decoder_start_token_id = tokenizer.lang_code_to_id[ data_args.lang] else: model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids( data_args.lang) if model.config.decoder_start_token_id is None: raise ValueError( "Make sure that `config.decoder_start_token_id` is correctly defined" ) if (hasattr(model.config, "max_position_embeddings") and model.config.max_position_embeddings < data_args.max_source_length): if model_args.resize_position_embeddings is None: logger.warning( "Increasing the model's number of position embedding vectors from" f" {model.config.max_position_embeddings} to {data_args.max_source_length}." ) model.resize_position_embeddings(data_args.max_source_length) elif model_args.resize_position_embeddings: model.resize_position_embeddings(data_args.max_source_length) else: raise ValueError( f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has" f" {model.config.max_position_embeddings} position encodings. Consider either reducing" f" `--max_source_length` to {model.config.max_position_embeddings} or to automatically resize the" " model's position encodings by passing `--resize_position_embeddings`." ) prefix = data_args.source_prefix if data_args.source_prefix is not None else "" # Preprocessing the datasets. # We need to tokenize inputs and targets. if training_args.do_train: column_names = raw_datasets["train"].column_names elif training_args.do_eval: column_names = raw_datasets["validation"].column_names elif training_args.do_predict: column_names = raw_datasets["test"].column_names else: logger.info( "There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`." ) return if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)): assert ( data_args.lang is not None ), f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --lang argument" tokenizer.src_lang = data_args.lang tokenizer.tgt_lang = data_args.lang # For multilingual translation models like mBART-50 and M2M100 we need to force the target language token # as the first generated token. We ask the user to explicitly provide this as --forced_bos_token argument. forced_bos_token_id = ( tokenizer.lang_code_to_id[data_args.forced_bos_token] if data_args.forced_bos_token is not None else None) model.config.forced_bos_token_id = forced_bos_token_id # Get the column names for input/target. dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None) if data_args.text_column is None: text_column = dataset_columns[ 0] if dataset_columns is not None else column_names[0] else: text_column = data_args.text_column if text_column not in column_names: raise ValueError( f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}" ) if data_args.summary_column is None: summary_column = dataset_columns[ 1] if dataset_columns is not None else column_names[1] else: summary_column = data_args.summary_column if summary_column not in column_names: raise ValueError( f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}" ) # 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_function(examples): # remove pairs where at least one record is None inputs, targets = [], [] for i in range(len(examples[text_column])): if examples[text_column][i] is not None and examples[ summary_column][i] is not None: inputs.append(examples[text_column][i]) targets.append(examples[summary_column][i]) inputs = [prefix + inp for inp in inputs] model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True) # Setup the tokenizer for targets with tokenizer.as_target_tokenizer(): labels = tokenizer(targets, 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 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: max_train_samples = min(len(train_dataset), data_args.max_train_samples) train_dataset = train_dataset.select(range(max_train_samples)) with training_args.main_process_first( desc="train dataset map pre-processing"): train_dataset = train_dataset.map( preprocess_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on train dataset", ) if training_args.do_eval: max_target_length = data_args.val_max_target_length if "validation" 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: max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples) eval_dataset = eval_dataset.select(range(max_eval_samples)) with training_args.main_process_first( desc="validation dataset map pre-processing"): eval_dataset = eval_dataset.map( preprocess_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on validation dataset", ) if training_args.do_predict: max_target_length = data_args.val_max_target_length if "test" 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: max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples) predict_dataset = predict_dataset.select( range(max_predict_samples)) with training_args.main_process_first( desc="prediction dataset map pre-processing"): predict_dataset = predict_dataset.map( preprocess_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on prediction dataset", ) # 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, ) # Metric metric = load_metric("rouge") def postprocess_text(preds, labels): preds = [pred.strip() for pred in preds] labels = [label.strip() for label in labels] # rougeLSum expects newline after each sentence preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds] labels = ["\n".join(nltk.sent_tokenize(label)) 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) result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True) # Extract a few results from ROUGE result = { key: value.mid.fmeasure * 100 for key, value in result.items() } prediction_lens = [ np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds ] result["gen_len"] = np.mean(prediction_lens) result = {k: round(v, 4) for k, v in result.items()} 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, ) # 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) 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 = {} max_length = (training_args.generation_max_length if training_args.generation_max_length is not None else data_args.val_max_target_length) num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate(max_length=max_length, num_beams=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=max_length, num_beams=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, "generated_predictions.txt") with open(output_prediction_file, "w") as writer: writer.write("\n".join(predictions)) kwargs = { "finetuned_from": model_args.model_name_or_path, "tasks": "summarization" } if data_args.dataset_name is not None: kwargs["dataset_tags"] = data_args.dataset_name if data_args.dataset_config_name is not None: kwargs["dataset_args"] = data_args.dataset_config_name kwargs[ "dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}" else: kwargs["dataset"] = data_args.dataset_name if data_args.lang is not None: kwargs["language"] = data_args.lang if training_args.push_to_hub: trainer.push_to_hub(**kwargs) else: trainer.create_model_card(**kwargs) return 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, 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( ) # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) # Setup logging, we only want one process per machine to log things on the screen. logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR) if jax.process_index() == 0: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # Handle the repository creation if training_args.push_to_hub: if training_args.hub_model_id is None: repo_name = get_full_repo_name(Path( training_args.output_dir).absolute().name, token=training_args.hub_token) else: repo_name = training_args.hub_model_id repo = Repository(training_args.output_dir, clone_from=repo_name) # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below) # or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/ # (the dataset will be downloaded automatically from the datasets Hub). # # For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called # 'tokens' is found. 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 the dataset from local csv or json file. data_files = {} if data_args.train_file is not None: data_files["train"] = data_args.train_file if data_args.validation_file is not None: data_files["validation"] = data_args.validation_file extension = (data_args.train_file if data_args.train_file is not None else data_args.valid_file).split(".")[-1] raw_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 at # https://huggingface.co/docs/datasets/loading_datasets.html. if raw_datasets["train"] is not None: column_names = raw_datasets["train"].column_names features = raw_datasets["train"].features else: column_names = raw_datasets["validation"].column_names features = raw_datasets["validation"].features if data_args.text_column_name is not None: text_column_name = data_args.text_column_name elif "tokens" in column_names: text_column_name = "tokens" else: text_column_name = column_names[0] if data_args.label_column_name is not None: label_column_name = data_args.label_column_name elif f"{data_args.task_name}_tags" in column_names: label_column_name = f"{data_args.task_name}_tags" else: label_column_name = column_names[1] # In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the # unique labels. def get_label_list(labels): unique_labels = set() for label in labels: unique_labels = unique_labels | set(label) label_list = list(unique_labels) label_list.sort() return label_list if isinstance(features[label_column_name].feature, ClassLabel): label_list = features[label_column_name].feature.names # No need to convert the labels since they are already ints. label_to_id = {i: i for i in range(len(label_list))} else: label_list = get_label_list(raw_datasets["train"][label_column_name]) label_to_id = {l: i for i, l in enumerate(label_list)} num_labels = len(label_list) # Load pretrained model and tokenizer config = AutoConfig.from_pretrained( model_args.config_name if model_args.config_name else model_args.model_name_or_path, num_labels=num_labels, label2id=label_to_id, id2label={i: l for l, i in label_to_id.items()}, finetuning_task=data_args.task_name, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path if config.model_type in {"gpt2", "roberta"}: tokenizer = AutoTokenizer.from_pretrained( tokenizer_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, add_prefix_space=True, ) else: tokenizer = AutoTokenizer.from_pretrained( tokenizer_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, ) model = FlaxAutoModelForTokenClassification.from_pretrained( 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, ) # Preprocessing the datasets # Tokenize all texts and align the labels with them. def tokenize_and_align_labels(examples): tokenized_inputs = tokenizer( examples[text_column_name], max_length=data_args.max_seq_length, padding="max_length", truncation=True, # We use this argument because the texts in our dataset are lists of words (with a label for each word). is_split_into_words=True, ) labels = [] for i, label in enumerate(examples[label_column_name]): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label_to_id[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label_to_id[label[word_idx]] if data_args. label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs processed_raw_datasets = raw_datasets.map( tokenize_and_align_labels, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, remove_columns=raw_datasets["train"].column_names, desc="Running tokenizer on dataset", ) train_dataset = processed_raw_datasets["train"] eval_dataset = processed_raw_datasets["validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # Define a summary writer has_tensorboard = is_tensorboard_available() if has_tensorboard and jax.process_index() == 0: try: from flax.metrics.tensorboard import SummaryWriter summary_writer = SummaryWriter(training_args.output_dir) summary_writer.hparams({ **training_args.to_dict(), **vars(model_args), **vars(data_args) }) except ImportError as ie: has_tensorboard = False logger.warning( f"Unable to display metrics through TensorBoard because some package are not installed: {ie}" ) else: logger.warning( "Unable to display metrics through TensorBoard because the package is not installed: " "Please run pip install tensorboard to enable.") def write_train_metric(summary_writer, train_metrics, train_time, step): summary_writer.scalar("train_time", train_time, step) train_metrics = get_metrics(train_metrics) for key, vals in train_metrics.items(): tag = f"train_{key}" for i, val in enumerate(vals): summary_writer.scalar(tag, val, step - len(vals) + i + 1) def write_eval_metric(summary_writer, eval_metrics, step): for metric_name, value in eval_metrics.items(): summary_writer.scalar(f"eval_{metric_name}", value, step) num_epochs = int(training_args.num_train_epochs) rng = jax.random.PRNGKey(training_args.seed) dropout_rngs = jax.random.split(rng, jax.local_device_count()) train_batch_size = training_args.per_device_train_batch_size * jax.local_device_count( ) eval_batch_size = training_args.per_device_eval_batch_size * jax.local_device_count( ) learning_rate_fn = create_learning_rate_fn( len(train_dataset), train_batch_size, training_args.num_train_epochs, training_args.warmup_steps, training_args.learning_rate, ) state = create_train_state(model, learning_rate_fn, num_labels=num_labels, training_args=training_args) # define step functions def train_step( state: train_state.TrainState, batch: Dict[str, Array], dropout_rng: PRNGKey) -> Tuple[train_state.TrainState, float]: """Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`.""" dropout_rng, new_dropout_rng = jax.random.split(dropout_rng) targets = batch.pop("labels") def loss_fn(params): logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0] loss = state.loss_fn(logits, targets) return loss grad_fn = jax.value_and_grad(loss_fn) loss, grad = grad_fn(state.params) grad = jax.lax.pmean(grad, "batch") new_state = state.apply_gradients(grads=grad) metrics = jax.lax.pmean( { "loss": loss, "learning_rate": learning_rate_fn(state.step) }, axis_name="batch") return new_state, metrics, new_dropout_rng p_train_step = jax.pmap(train_step, axis_name="batch", donate_argnums=(0, )) def eval_step(state, batch): logits = state.apply_fn(**batch, params=state.params, train=False)[0] return state.logits_fn(logits) p_eval_step = jax.pmap(eval_step, axis_name="batch") metric = load_metric("seqeval") def get_labels(y_pred, y_true): # Transform predictions and references tensos to numpy arrays # Remove ignored index (special tokens) true_predictions = [[ label_list[p] for (p, l) in zip(pred, gold_label) if l != -100 ] for pred, gold_label in zip(y_pred, y_true)] true_labels = [[ label_list[l] for (p, l) in zip(pred, gold_label) if l != -100 ] for pred, gold_label in zip(y_pred, y_true)] return true_predictions, true_labels def compute_metrics(): results = metric.compute() if data_args.return_entity_level_metrics: # Unpack nested dictionaries final_results = {} for key, value in results.items(): if isinstance(value, dict): for n, v in value.items(): final_results[f"{key}_{n}"] = v else: final_results[key] = value return final_results else: return { "precision": results["overall_precision"], "recall": results["overall_recall"], "f1": results["overall_f1"], "accuracy": results["overall_accuracy"], } logger.info(f"===== Starting training ({num_epochs} epochs) =====") train_time = 0 # make sure weights are replicated on each device state = replicate(state) train_time = 0 step_per_epoch = len(train_dataset) // train_batch_size total_steps = step_per_epoch * num_epochs epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0) for epoch in epochs: train_start = time.time() train_metrics = [] # Create sampling rng rng, input_rng = jax.random.split(rng) # train for step, batch in enumerate( tqdm( train_data_collator(input_rng, train_dataset, train_batch_size), total=step_per_epoch, desc="Training...", position=1, )): state, train_metric, dropout_rngs = p_train_step( state, batch, dropout_rngs) train_metrics.append(train_metric) cur_step = (epoch * step_per_epoch) + (step + 1) if cur_step % training_args.logging_steps == 0 and cur_step > 0: # Save metrics train_metric = unreplicate(train_metric) train_time += time.time() - train_start if has_tensorboard and jax.process_index() == 0: write_train_metric(summary_writer, train_metrics, train_time, cur_step) epochs.write( f"Step... ({cur_step}/{total_steps} | Training Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})" ) train_metrics = [] if cur_step % training_args.eval_steps == 0 and cur_step > 0: eval_metrics = {} # evaluate for batch in tqdm( eval_data_collator(eval_dataset, eval_batch_size), total=len(eval_dataset) // eval_batch_size, desc="Evaluating ...", position=2, ): labels = batch.pop("labels") predictions = p_eval_step(state, batch) predictions = np.array( [pred for pred in chain(*predictions)]) labels = np.array([label for label in chain(*labels)]) labels[np.array(chain( *batch["attention_mask"])) == 0] = -100 preds, refs = get_labels(predictions, labels) metric.add_batch( predictions=preds, references=refs, ) # evaluate also on leftover examples (not divisible by batch_size) num_leftover_samples = len(eval_dataset) % eval_batch_size # make sure leftover batch is evaluated on one device if num_leftover_samples > 0 and jax.process_index() == 0: # take leftover samples batch = eval_dataset[-num_leftover_samples:] batch = {k: np.array(v) for k, v in batch.items()} labels = batch.pop("labels") predictions = eval_step(unreplicate(state), batch) labels = np.array(labels) labels[np.array(batch["attention_mask"]) == 0] = -100 preds, refs = get_labels(predictions, labels) metric.add_batch( predictions=preds, references=refs, ) eval_metrics = compute_metrics() if data_args.return_entity_level_metrics: logger.info( f"Step... ({cur_step}/{total_steps} | Validation metrics: {eval_metrics}" ) else: logger.info( f"Step... ({cur_step}/{total_steps} | Validation f1: {eval_metrics['f1']}, Validation Acc: {eval_metrics['accuracy']})" ) if has_tensorboard and jax.process_index() == 0: write_eval_metric(summary_writer, eval_metrics, cur_step) if (cur_step % training_args.save_steps == 0 and cur_step > 0) or (cur_step == total_steps): # save checkpoint after each epoch and push checkpoint to the hub if jax.process_index() == 0: params = jax.device_get(unreplicate(state.params)) model.save_pretrained(training_args.output_dir, params=params) tokenizer.save_pretrained(training_args.output_dir) if training_args.push_to_hub: repo.push_to_hub( commit_message= f"Saving weights and logs of step {cur_step}", blocking=False) epochs.desc = f"Epoch ... {epoch + 1}/{num_epochs}" # Eval after training if training_args.do_eval: eval_metrics = {} eval_loader = eval_data_collator(eval_dataset, eval_batch_size) for batch in tqdm(eval_loader, total=len(eval_dataset) // eval_batch_size, desc="Evaluating ...", position=2): labels = batch.pop("labels") predictions = p_eval_step(state, batch) predictions = np.array([pred for pred in chain(*predictions)]) labels = np.array([label for label in chain(*labels)]) labels[np.array(chain(*batch["attention_mask"])) == 0] = -100 preds, refs = get_labels(predictions, labels) metric.add_batch(predictions=preds, references=refs) # evaluate also on leftover examples (not divisible by batch_size) num_leftover_samples = len(eval_dataset) % eval_batch_size # make sure leftover batch is evaluated on one device if num_leftover_samples > 0 and jax.process_index() == 0: # take leftover samples batch = eval_dataset[-num_leftover_samples:] batch = {k: np.array(v) for k, v in batch.items()} labels = np.array(batch.pop("labels")) predictions = eval_step(unreplicate(state), batch) labels[np.array(batch["attention_mask"]) == 0] = -100 preds, refs = get_labels(predictions, labels) metric.add_batch(predictions=preds, references=refs) eval_metrics = compute_metrics() if jax.process_index() == 0: eval_metrics = { f"eval_{metric_name}": value for metric_name, value in eval_metrics.items() } path = os.path.join(training_args.output_dir, "eval_results.json") with open(path, "w") as f: json.dump(eval_metrics, f, indent=4, sort_keys=True)
training_args = Seq2SeqTrainingArguments( predict_with_generate=True, evaluation_strategy="steps", per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, fp16=True, output_dir="./trainedgigaword/", logging_steps=1000, save_steps=500, eval_steps=7500, warmup_steps=2000, save_total_limit=3, ) rouge = load_metric("rouge") def compute_metrics(pred): labels_ids = pred.label_ids pred_ids = pred.predictions pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True) labels_ids[labels_ids == -100] = tokenizer.pad_token_id label_str = tokenizer.batch_decode(labels_ids, skip_special_tokens=True) rouge_output = rouge.compute(predictions=pred_str, references=label_str, rouge_types=["rouge2"])["rouge2"].mid return {
def load_local_metric(metric_name, *args, **kwargs): return load_metric(os.path.join("metrics", metric_name), *args, **kwargs)
def test_seqeval_raises_when_incorrect_scheme(): metric = load_metric(os.path.join("metrics", "seqeval")) wrong_scheme = "ERROR" error_message = f"Scheme should be one of [IOB1, IOB2, IOE1, IOE2, IOBES, BILOU], got {wrong_scheme}" with pytest.raises(ValueError, match=re.escape(error_message)): metric.compute(predictions=[], references=[], scheme=wrong_scheme)
import transformers from transformers import EncoderDecoderModel, AutoTokenizer from tokenizers import Tokenizer import torch import sys from datasets import load_metric metric = load_metric('sacrebleu') chk_dir = sys.argv[1] chk_num = sys.argv[2] num_beams = int(sys.argv[3]) code_tok = False if sys.argv[4] == 'false' else True print_bool = False if sys.argv[5] == 'false' else True model = EncoderDecoderModel.from_pretrained('./{}/checkpoint-{}/'.format( chk_dir, chk_num)) code_tokenizer = Tokenizer.from_file( 'code_tokenizer.json') if code_tok else AutoTokenizer.from_pretrained( 'bert-base-uncased') text_tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') #print(text_tokenizer.convert_tokens_to_ids()) pad_token_id = 1 if code_tok else 0 bos_token_id = 2 if code_tok else 101 eos_token_id = 3 if code_tok else 102 f = open('tok-eval.tsv', 'r') for i, line in enumerate(f): if i == 0: continue if i > 1000:
def run_mrc(data_args, training_args, model_args, datasets, tokenizer, model): # Preprocessing the datasets. # Preprocessing is slighlty different for training and evaluation. if training_args.do_train: column_names = datasets["train"].column_names else: column_names = datasets["validation"].column_names question_column_name = "question" if "question" in column_names else column_names[ 0] context_column_name = "context" if "context" in column_names else column_names[ 1] answer_column_name = "answers" if "answers" in column_names else column_names[ 2] # Padding side determines if we do (question|context) or (context|question). pad_on_right = tokenizer.padding_side == "right" # check if there is an error last_checkpoint, max_seq_length = check_no_error(training_args, data_args, tokenizer, datasets) # Training preprocessing def prepare_train_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[ question_column_name if pad_on_right else context_column_name], examples[ context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding="max_length" if data_args.pad_to_max_length else False, ) # Since one example might give us several features if it has a long context, # we need a map from a feature to its corresponding example. # This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # The offset mappings will give us a map from token to character position in the original context. This will # help us compute the start_positions and end_positions. offset_mapping = tokenized_examples.pop("offset_mapping") # Let's label those examples! tokenized_examples["start_positions"] = [] tokenized_examples["end_positions"] = [] for i, offsets in enumerate(offset_mapping): # We will label impossible answers with the index of the CLS token. input_ids = tokenized_examples["input_ids"][i] cls_index = input_ids.index(tokenizer.cls_token_id) # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] answers = examples[answer_column_name][sample_index] # print(answers) # If no answers are given, set the cls_index as answer. if len(answers["answer_start"]) == 0: tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Start/end character index of the answer in the text. start_char = answers["answer_start"][0] end_char = start_char + len(answers["text"][0]) # Start token index of the current span in the text. token_start_index = 0 while sequence_ids[token_start_index] != (1 if pad_on_right else 0): token_start_index += 1 # End token index of the current span in the text. token_end_index = len(input_ids) - 1 while sequence_ids[token_end_index] != (1 if pad_on_right else 0): token_end_index -= 1 # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index). if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char): tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Otherwise move the token_start_index and token_end_index to the two ends of the answer. # Note: we could go after the last offset if the answer is the last word (edge case). while (token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char): token_start_index += 1 tokenized_examples["start_positions"].append( token_start_index - 1) while offsets[token_end_index][1] >= end_char: token_end_index -= 1 tokenized_examples["end_positions"].append( token_end_index + 1) return tokenized_examples if training_args.do_train: if "train" not in datasets: raise ValueError("--do_train requires a train dataset") train_dataset = datasets["train"] # Create train feature from dataset train_dataset = train_dataset.map( prepare_train_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # Validation preprocessing def prepare_validation_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[ question_column_name if pad_on_right else context_column_name], examples[ context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding="max_length" if data_args.pad_to_max_length else False, ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the # corresponding example_id and we will store the offset mappings. tokenized_examples["example_id"] = [] for i in range(len(tokenized_examples["input_ids"])): # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) context_index = 1 if pad_on_right else 0 # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] tokenized_examples["example_id"].append( examples["id"][sample_index]) # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token # position is part of the context or not. tokenized_examples["offset_mapping"][i] = [ (o if sequence_ids[k] == context_index else None) for k, o in enumerate(tokenized_examples["offset_mapping"][i]) ] return tokenized_examples if training_args.do_eval: eval_dataset = datasets["validation"] # Validation Feature Creation eval_dataset = eval_dataset.map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # Data collator # We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data collator. data_collator = (DataCollatorWithPadding( tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)) # Post-processing: def post_processing_function(examples, features, predictions, training_args): # Post-processing: we match the start logits and end logits to answers in the original context. predictions = postprocess_qa_predictions( examples=examples, features=features, predictions=predictions, max_answer_length=data_args.max_answer_length, output_dir=training_args.output_dir, ) # Format the result to the format the metric expects. formatted_predictions = [{ "id": k, "prediction_text": v } for k, v in predictions.items()] if training_args.do_predict: return formatted_predictions elif training_args.do_eval: references = [{ "id": ex["id"], "answers": ex[answer_column_name] } for ex in datasets["validation"]] return EvalPrediction(predictions=formatted_predictions, label_ids=references) metric = load_metric("squad") def compute_metrics(p: EvalPrediction): return metric.compute(predictions=p.predictions, references=p.label_ids) # Initialize our Trainer trainer = QuestionAnsweringTrainer( 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, eval_examples=datasets["validation"] if training_args.do_eval else None, tokenizer=tokenizer, data_collator=data_collator, post_process_function=post_processing_function, compute_metrics=compute_metrics, ) # Training if training_args.do_train: if last_checkpoint is not None: checkpoint = last_checkpoint elif os.path.isdir(model_args.model_name_or_path): checkpoint = model_args.model_name_or_path else: checkpoint = None train_result = trainer.train(resume_from_checkpoint=checkpoint) trainer.save_model() # Saves the tokenizer too for easy upload metrics = train_result.metrics metrics["train_samples"] = len(train_dataset) trainer.log_metrics("train", metrics) trainer.save_metrics("train", metrics) trainer.save_state() output_train_file = os.path.join(training_args.output_dir, "train_results.txt") with open(output_train_file, "w") as writer: logger.info("***** Train results *****") for key, value in sorted(train_result.metrics.items()): logger.info(f" {key} = {value}") writer.write(f"{key} = {value}\n") # 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")) # Evaluation if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate() metrics["eval_samples"] = len(eval_dataset) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics)
def main(): # 查看SRC/Transformers/Training_args.py中的所有可能参数 # or by passing the --help flag to this script. # 我们现在保留不同的args集,以便更干净地分离关注的参数。 parser = HfArgumentParser( (ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments)) if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): # 如果我们只向脚本传递一个参数,而且是一个json文件的路径,那么我们就来解析它以获得我们的参数。 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 data_args.source_prefix is None and model_args.model_name_or_path in [ "t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b", ]: logger.warning( "你正在运行一个T5模型,但没有提供一个源前缀,这是必须的, e.g. with `--source_prefix 'summarize: ' `" ) # 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: 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) # 获取数据集:你可以提供你自己的JSON训练和评估文件(见下文)。 # 或只是提供中心上的一个公共数据集的名称 at https://huggingface.co/datasets/ # (数据集将自动从数据集hub下载)。 # # 对于翻译,只支持JSON文件,其中有一个名为 "translation"的字段,包含源语言和目标语言的两个key(除非你调整下面的内容)。 # 在分布式训练中,load_dataset函数保证只有一个本地进程可以同时下载数据集。 if data_args.dataset_name == 'csl': datasets = load_dataset(path='data_csl/csl.py', name='csl', data_files={ 'train': 'data_csl/train.json', 'validation': 'data_csl/dev.json', 'test': 'data_csl/test.json' }) elif data_args.dataset_name is not None: # 从hub下载并加载数据集。 if os.path.exists('data_script/' + data_args.dataset_name): dataset_path = 'data_script/' + data_args.dataset_name + '.py' else: dataset_path = data_args.dataset_name #如果手动把数据下载到本地,那么请设置忽略校验md5 datasets = load_dataset(dataset_path, data_args.dataset_config_name, cache_dir=model_args.cache_dir, ignore_verifications=True) 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) # 查看更多关于加载任何类型的标准或自定义数据集(从文件、python dict、pandas DataFrame等)的信息,请访问 # https://huggingface.co/docs/datasets/loading_datasets.html. # 2. 加载预训练的模型和tokenizer # # 分布式训练: # .from_pretrained方法保证只有一个本地进程可以同时下载模型和单词表。 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, ) 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, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForSeq2SeqLM.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" ) prefix = data_args.source_prefix if data_args.source_prefix is not None else "" # 对数据集进行预处理。 # 我们需要对输入和目标进行tokenize。, column_names: ['article', 'highlights', 'id'] 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("没有什么可以做的。 请传入 `do_train`, `do_eval` and/or `do_predict`.") return # 获取输入/目标的列名称。, 我们只需要 dataset_columns: ('article', 'highlights') dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None) if data_args.text_column is None: text_column = dataset_columns[ 0] if dataset_columns is not None else column_names[0] else: text_column = data_args.text_column if text_column not in column_names: raise ValueError( f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}" ) if data_args.summary_column is None: summary_column = dataset_columns[ 1] if dataset_columns is not None else column_names[1] else: summary_column = data_args.summary_column if summary_column not in column_names: raise ValueError( f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}" ) # 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_function(examples): """ 数据预处理的函数 Args: examples (): Returns: """ # inputs 是一个批次的数据,是全文, targets是摘要数据,是一个批次的, prefix代表每个输入前都加前缀,代表要执行的任务: 'summarize: ' inputs = examples[text_column] targets = examples[summary_column] inputs = [prefix + inp for inp in inputs] model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True) # Setup the tokenizer for targets with tokenizer.as_target_tokenizer(): labels = tokenizer(targets, 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 if training_args.do_train: train_dataset = datasets["train"] if "train" not in datasets: raise ValueError("--do_train requires a train dataset") 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_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_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_val_samples is not None: eval_dataset = eval_dataset.select(range( data_args.max_val_samples)) eval_dataset = eval_dataset.map( preprocess_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") test_dataset = datasets["test"] if data_args.max_test_samples is not None: test_dataset = test_dataset.select( range(data_args.max_test_samples)) test_dataset = test_dataset.map( preprocess_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, ) # Metric if os.path.exists('data_script/rouge.py'): metric = load_metric("data_script/rouge.py") else: metric = load_metric("rouge") def postprocess_text(preds, labels): preds = [pred.strip() for pred in preds] labels = [label.strip() for label in labels] # rougeLSum expects newline after each sentence preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds] labels = ["\n".join(nltk.sent_tokenize(label)) 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) result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True) # Extract a few results from ROUGE result = { key: value.mid.fmeasure * 100 for key, value in result.items() } prediction_lens = [ np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds ] result["gen_len"] = np.mean(prediction_lens) result = {k: round(v, 4) for k, v in result.items()} 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, ) # Training if training_args.do_train: logger.info("*** 训练 ***") if last_checkpoint is not None: checkpoint = last_checkpoint elif os.path.isdir(model_args.model_name_or_path): checkpoint = model_args.model_name_or_path else: checkpoint = None 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("*** 评估 ***") metrics = trainer.evaluate(max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, metric_key_prefix="eval") max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len( eval_dataset) metrics["eval_samples"] = min(max_val_samples, len(eval_dataset)) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics) if training_args.do_predict: logger.info("*** 测试 ***") test_results = trainer.predict( test_dataset, metric_key_prefix="test", max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, ) metrics = test_results.metrics max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len( test_dataset) metrics["test_samples"] = min(max_test_samples, len(test_dataset)) trainer.log_metrics("test", metrics) trainer.save_metrics("test", metrics) if trainer.is_world_process_zero(): if training_args.predict_with_generate: test_preds = tokenizer.batch_decode( test_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True) test_preds = [pred.strip() for pred in test_preds] output_test_preds_file = os.path.join(training_args.output_dir, "test_generations.txt") with open(output_test_preds_file, "w") as writer: writer.write("\n".join(test_preds)) return 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, 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) 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." ) # Initialize our dataset and prepare it for the 'image-classification' task. ds = load_dataset( data_args.dataset_name, data_args.dataset_config_name, data_files=data_args.data_files, cache_dir=model_args.cache_dir, task="image-classification", ) # Define torchvision transforms to be applied to each image. normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) _train_transforms = Compose([ RandomResizedCrop(data_args.image_size), RandomHorizontalFlip(), ToTensor(), normalize, ]) _val_transforms = Compose([ Resize(data_args.image_size), CenterCrop(data_args.image_size), ToTensor(), normalize, ]) def train_transforms(example_batch): """Apply _train_transforms across a batch.""" example_batch["pixel_values"] = [ _train_transforms(pil_loader(f)) for f in example_batch["image_file_path"] ] return example_batch def val_transforms(example_batch): """Apply _val_transforms across a batch.""" example_batch["pixel_values"] = [ _val_transforms(pil_loader(f)) for f in example_batch["image_file_path"] ] return example_batch # If we don't have a validation split, split off a percentage of train as validation. data_args.train_val_split = None if "validation" in ds.keys( ) else data_args.train_val_split if isinstance(data_args.train_val_split, float) and data_args.train_val_split > 0.0: split = ds["train"].train_test_split(data_args.train_val_split) ds["train"] = split["train"] ds["validation"] = split["test"] # Prepare label mappings. # We'll include these in the model's config to get human readable labels in the Inference API. labels = ds["train"].features["labels"].names label2id, id2label = dict(), dict() for i, label in enumerate(labels): label2id[label] = str(i) id2label[str(i)] = label # Load the accuracy metric from the datasets package metric = datasets.load_metric("accuracy") # Define our 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): """Computes accuracy on a batch of predictions""" return metric.compute(predictions=np.argmax(p.predictions, axis=1), references=p.label_ids) config = AutoConfig.from_pretrained( model_args.config_name or model_args.model_name_or_path, num_labels=len(labels), label2id=label2id, id2label=id2label, finetuning_task="image-classification", cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForImageClassification.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, ) # NOTE - We aren't directly using this feature extractor since we defined custom transforms above. # We initialize this instance below and pass it to Trainer to ensure that the feature extraction # config, preprocessor_config.json, is included in output directories. # This way if we push a model to the hub, the inference widget will work. feature_extractor = AutoFeatureExtractor.from_pretrained( model_args.feature_extractor_name or 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, size=data_args.image_size, image_mean=normalize.mean, image_std=normalize.std, ) if training_args.do_train: if "train" not in ds: raise ValueError("--do_train requires a train dataset") if data_args.max_train_samples is not None: ds["train"] = ds["train"].shuffle(seed=training_args.seed).select( range(data_args.max_train_samples)) # Set the training transforms ds["train"].set_transform(train_transforms) if training_args.do_eval: if "validation" not in ds: raise ValueError("--do_eval requires a validation dataset") if data_args.max_eval_samples is not None: ds["validation"] = (ds["validation"].shuffle( seed=training_args.seed).select( range(data_args.max_eval_samples))) # Set the validation transforms ds["validation"].set_transform(val_transforms) # Initalize our trainer trainer = Trainer( model=model, args=training_args, train_dataset=ds["train"] if training_args.do_train else None, eval_dataset=ds["validation"] if training_args.do_eval else None, compute_metrics=compute_metrics, tokenizer=feature_extractor, data_collator=collate_fn, ) # 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) trainer.save_model() trainer.log_metrics("train", train_result.metrics) trainer.save_metrics("train", train_result.metrics) trainer.save_state() # Evaluation if training_args.do_eval: metrics = trainer.evaluate() trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics) # Write model card and (optionally) push to hub kwargs = { "finetuned_from": model_args.model_name_or_path, "tasks": "image-classification", "dataset": data_args.dataset_name, "tags": ["image-classification"], } if training_args.push_to_hub: trainer.push_to_hub(**kwargs) else: trainer.create_model_card(**kwargs)
def train(): processor = Wav2Vec2Processor.from_pretrained('./asr_output/pretrained_processor') # torch.cuda.empty_cache() data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True) wer_metric = load_metric("wer") # print(processor.tokenizer) def compute_metrics(pred): pred_logits = pred.predictions pred_ids = np.argmax(pred_logits, axis=-1) pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id pred_str = processor.batch_decode(pred_ids) # we do not want to group tokens when computing the metrics label_str = processor.batch_decode(pred.label_ids, group_tokens=False) wer = wer_metric.compute(predictions=pred_str, references=label_str) return {"wer": wer} model = Wav2Vec2ForCTC.from_pretrained( "asr_output/checkpoint-906", attention_dropout=0.1, hidden_dropout=0.1, feat_proj_dropout=0.0, mask_time_prob=0.05, layerdrop=0.1, gradient_checkpointing=True, ctc_loss_reduction="mean", pad_token_id=processor.tokenizer.pad_token_id, vocab_size=len(processor.tokenizer) ) # print(model) # exit() model.freeze_feature_extractor() training_args = TrainingArguments( output_dir="./asr_output/", # output_dir="./wav2vec2-large-xlsr-turkish-demo", overwrite_output_dir = True, group_by_length=True, per_device_train_batch_size=16, gradient_accumulation_steps=2, evaluation_strategy="steps", num_train_epochs=250, fp16=True, save_steps=1, eval_steps=100, logging_steps=500, learning_rate=1e-4, weight_decay=0.005, # learning_rate=3e-4, warmup_steps=500, save_total_limit=2, ) print("loading data...") with open('./data/speech-sme-asr/train_asr.pkl', 'rb') as f: train = pickle.load(f) with open('./data/speech-sme-asr/test_asr.pkl', 'rb') as f: test = pickle.load(f) print('loaded...') trainer = Trainer( model=model, data_collator=data_collator, args=training_args, compute_metrics=compute_metrics, train_dataset=train, eval_dataset=test, tokenizer=processor.feature_extractor, ) print('Starting the trainer...') trainer.train()
def __init__(self, *args): _all_metrics = list_metrics() self._metrics = {m: load_metric(m) for m in args if m in _all_metrics and isinstance(m, str)}
def metric(self): if self._metric is None: self._metric = load_metric("accuracy") return self._metric
def main(): args = parse_args() if args.source_prefix is None and args.model_name_or_path in [ "t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b", ]: logger.warning( "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with " "`--source_prefix 'summarize: ' `" ) # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. accelerator = Accelerator() # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state) # Setup logging, we only want one process per machine to log things on the screen. # accelerator.is_local_main_process is only True for one process per machine. logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Get the datasets: you can either provide your own CSV/JSON/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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 args.dataset_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name) else: data_files = {} if args.train_file is not None: data_files["train"] = args.train_file if args.validation_file is not None: data_files["validation"] = args.validation_file extension = args.train_file.split(".")[-1] raw_datasets = load_dataset(extension, data_files=data_files) # 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 # # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. if args.config_name: config = AutoConfig.from_pretrained(args.config_name) elif args.model_name_or_path: config = AutoConfig.from_pretrained(args.model_name_or_path) else: config = CONFIG_MAPPING[args.model_type]() logger.warning("You are instantiating a new config instance from scratch.") if args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer) elif args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if args.model_name_or_path: model = AutoModelForSeq2SeqLM.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ) else: logger.info("Training new model from scratch") model = AutoModelForSeq2SeqLM.from_config(config) model.resize_token_embeddings(len(tokenizer)) if model.config.decoder_start_token_id is None: raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined") prefix = args.source_prefix if args.source_prefix is not None else "" # Preprocessing the datasets. # First we tokenize all the texts. column_names = raw_datasets["train"].column_names # Get the column names for input/target. dataset_columns = summarization_name_mapping.get(args.dataset_name, None) if args.text_column is None: text_column = dataset_columns[0] if dataset_columns is not None else column_names[0] else: text_column = args.text_column if text_column not in column_names: raise ValueError( f"--text_column' value '{args.text_column}' needs to be one of: {', '.join(column_names)}" ) if args.summary_column is None: summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1] else: summary_column = args.summary_column if summary_column not in column_names: raise ValueError( f"--summary_column' value '{args.summary_column}' needs to be one of: {', '.join(column_names)}" ) # Temporarily set max_target_length for training. max_target_length = args.max_target_length padding = "max_length" if args.pad_to_max_length else False def preprocess_function(examples): inputs = examples[text_column] targets = examples[summary_column] inputs = [prefix + inp for inp in inputs] model_inputs = tokenizer(inputs, max_length=args.max_source_length, padding=padding, truncation=True) # Setup the tokenizer for targets with tokenizer.as_target_tokenizer(): labels = tokenizer(targets, 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 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 processed_datasets = raw_datasets.map( preprocess_function, batched=True, remove_columns=column_names, load_from_cache_file=not args.overwrite_cache, desc="Running tokenizer on dataset", ) train_dataset = processed_datasets["train"] eval_dataset = processed_datasets["validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 1): logger.info(f"Sample {index} of the training set: {train_dataset[index]}.") label_pad_token_id = -100 if 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 accelerator.use_fp16 else None, ) def postprocess_text(preds, labels): preds = [pred.strip() for pred in preds] labels = [label.strip() for label in labels] # rougeLSum expects newline after each sentence preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds] labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels] return preds, labels train_dataloader = DataLoader( train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size ) eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], "weight_decay": args.weight_decay, }, { "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0, }, ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader ) # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be # shorter in multiprocess) # Scheduler and math around the number of training steps. num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch else: args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Metric metric = load_metric("rouge") # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}") logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 for epoch in range(args.num_train_epochs): model.train() for step, batch in enumerate(train_dataloader): outputs = model(**batch) loss = outputs.loss loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if completed_steps >= args.max_train_steps: break model.eval() if args.val_max_target_length is None: args.val_max_target_length = args.max_target_length gen_kwargs = { "max_length": args.val_max_target_length if args is not None else config.max_length, "num_beams": args.num_beams, } for step, batch in enumerate(eval_dataloader): with torch.no_grad(): generated_tokens = accelerator.unwrap_model(model).generate( batch["input_ids"], attention_mask=batch["attention_mask"], **gen_kwargs, ) generated_tokens = accelerator.pad_across_processes( generated_tokens, dim=1, pad_index=tokenizer.pad_token_id ) labels = batch["labels"] if not args.pad_to_max_length: # If we did not pad to max length, we need to pad the labels too labels = accelerator.pad_across_processes(batch["labels"], dim=1, pad_index=tokenizer.pad_token_id) generated_tokens = accelerator.gather(generated_tokens).cpu().numpy() labels = accelerator.gather(labels).cpu().numpy() if 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) if isinstance(generated_tokens, tuple): generated_tokens = generated_tokens[0] decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True) decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels) metric.add_batch(predictions=decoded_preds, references=decoded_labels) result = metric.compute(use_stemmer=True) # Extract a few results from ROUGE result = {key: value.mid.fmeasure * 100 for key, value in result.items()} result = {k: round(v, 4) for k, v in result.items()} logger.info(result) if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
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( ) # 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: 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() transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() 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/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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. datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name) else: data_files = {} if data_args.train_file is not None: data_files["train"] = data_args.train_file if data_args.validation_file is not None: data_files["validation"] = data_args.validation_file if data_args.test_file is not None: data_files["test"] = data_args.test_file extension = data_args.train_file.split(".")[-1] datasets = load_dataset(extension, data_files=data_files) # 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. if training_args.do_train: column_names = datasets["train"].column_names features = datasets["train"].features else: column_names = datasets["validation"].column_names features = datasets["validation"].features text_column_name = "tokens" if "tokens" in column_names else column_names[0] label_column_name = (f"{data_args.task_name}_tags" if f"{data_args.task_name}_tags" in column_names else column_names[1]) # In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the # unique labels. def get_label_list(labels): unique_labels = set() for label in labels: unique_labels = unique_labels | set(label) label_list = list(unique_labels) label_list.sort() return label_list if isinstance(features[label_column_name].feature, ClassLabel): label_list = features[label_column_name].feature.names # No need to convert the labels since they are already ints. label_to_id = {i: i for i in range(len(label_list))} else: label_list = get_label_list(datasets["train"][label_column_name]) label_to_id = {l: i for i, l in enumerate(label_list)} num_labels = len(label_list) # 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, num_labels=num_labels, finetuning_task=data_args.task_name, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) 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, use_fast=True, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForTokenClassification.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, ) # Tokenizer check: this script requires a fast tokenizer. if not isinstance(tokenizer, PreTrainedTokenizerFast): raise ValueError( "This example script only works for models that have a fast tokenizer. Checkout the big table of models " "at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this " "requirement") # Preprocessing the dataset # Padding strategy padding = "max_length" if data_args.pad_to_max_length else False # Tokenize all texts and align the labels with them. def tokenize_and_align_labels(examples): tokenized_inputs = tokenizer( examples[text_column_name], padding=padding, truncation=True, # We use this argument because the texts in our dataset are lists of words (with a label for each word). is_split_into_words=True, ) labels = [] for i, label in enumerate(examples[label_column_name]): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label_to_id[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label_to_id[label[word_idx]] if data_args. label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs 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( tokenize_and_align_labels, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_eval: if "validation" not in datasets: raise ValueError("--do_eval requires a validation dataset") eval_dataset = datasets["validation"] if data_args.max_val_samples is not None: eval_dataset = eval_dataset.select(range( data_args.max_val_samples)) eval_dataset = eval_dataset.map( tokenize_and_align_labels, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_predict: if "test" not in datasets: raise ValueError("--do_predict requires a test dataset") test_dataset = datasets["test"] if data_args.max_test_samples is not None: test_dataset = test_dataset.select( range(data_args.max_test_samples)) test_dataset = test_dataset.map( tokenize_and_align_labels, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, ) # Data collator data_collator = DataCollatorForTokenClassification( tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None) # Metrics metric = load_metric("seqeval") def compute_metrics(p): predictions, labels = p predictions = np.argmax(predictions, axis=2) # Remove ignored index (special tokens) true_predictions = [[ label_list[p] for (p, l) in zip(prediction, label) if l != -100 ] for prediction, label in zip(predictions, labels)] true_labels = [[ label_list[l] for (p, l) in zip(prediction, label) if l != -100 ] for prediction, label in zip(predictions, labels)] results = metric.compute(predictions=true_predictions, references=true_labels) if data_args.return_entity_level_metrics: # Unpack nested dictionaries final_results = {} for key, value in results.items(): if isinstance(value, dict): for n, v in value.items(): final_results[f"{key}_{n}"] = v else: final_results[key] = value return final_results else: return { "precision": results["overall_precision"], "recall": results["overall_recall"], "f1": results["overall_f1"], "accuracy": results["overall_accuracy"], } # 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, tokenizer=tokenizer, data_collator=data_collator, compute_metrics=compute_metrics, ) # Training if training_args.do_train: if last_checkpoint is not None: checkpoint = last_checkpoint elif os.path.isdir(model_args.model_name_or_path): checkpoint = model_args.model_name_or_path else: checkpoint = None train_result = trainer.train(resume_from_checkpoint=checkpoint) metrics = train_result.metrics trainer.save_model() # Saves the tokenizer too for easy upload 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 if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate() max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len( eval_dataset) metrics["eval_samples"] = min(max_val_samples, len(eval_dataset)) trainer.log_metrics("eval", metrics) trainer.save_metrics("eval", metrics) # Predict if training_args.do_predict: logger.info("*** Predict ***") predictions, labels, metrics = trainer.predict(test_dataset) predictions = np.argmax(predictions, axis=2) # Remove ignored index (special tokens) true_predictions = [[ label_list[p] for (p, l) in zip(prediction, label) if l != -100 ] for prediction, label in zip(predictions, labels)] trainer.log_metrics("test", metrics) trainer.save_metrics("test", metrics) # Save predictions output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt") if trainer.is_world_process_zero(): with open(output_test_predictions_file, "w") as writer: for prediction in true_predictions: writer.write(" ".join(prediction) + "\n")
def main(): args = parse_args() # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. accelerator = Accelerator() # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state) # Setup logging, we only want one process per machine to log things on the screen. # accelerator.is_local_main_process is only True for one process per machine. logger.setLevel( logging.INFO if accelerator.is_local_main_process else logging.ERROR) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below) # or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/ # (the dataset will be downloaded automatically from the datasets Hub). # # For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called # 'tokens' is found. 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 args.dataset_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name, data_dir=args.dataset_dir) else: data_files = {} if args.train_file is not None: data_files["train"] = args.train_file if args.validation_file is not None: data_files["validation"] = args.validation_file extension = args.train_file.split(".")[-1] raw_datasets = load_dataset(extension, data_files=data_files) # Trim a number of training examples if args.debug: for split in raw_datasets.keys(): raw_datasets[split] = raw_datasets[split].select(range(100)) # 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. if raw_datasets["train"] is not None: column_names = raw_datasets["train"].column_names features = raw_datasets["train"].features else: column_names = raw_datasets["validation"].column_names features = raw_datasets["validation"].features text_column_name = "tokens" if "tokens" in column_names else column_names[0] label_column_name = f"{args.task_name}_tags" if f"{args.task_name}_tags" in column_names else column_names[ 1] # In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the # unique labels. def get_label_list(labels): unique_labels = set() for label in labels: unique_labels = unique_labels | set(label) label_list = list(unique_labels) label_list.sort() return label_list if isinstance(features[label_column_name].feature, ClassLabel): label_list = features[label_column_name].feature.names # No need to convert the labels since they are already ints. label_to_id = {i: i for i in range(len(label_list))} else: label_list = get_label_list(raw_datasets["train"][label_column_name]) label_to_id = {l: i for i, l in enumerate(label_list)} 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. if args.config_name: config = AutoConfig.from_pretrained(args.config_name, num_labels=num_labels) elif args.model_name_or_path: config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels) else: config = CONFIG_MAPPING[args.model_type]() logger.warning( "You are instantiating a new config instance from scratch.") if args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=True) elif args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=True) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if args.model_name_or_path: model = AutoModelForTokenClassification.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ) else: logger.info("Training new model from scratch") model = AutoModelForTokenClassification.from_config(config) model.resize_token_embeddings(len(tokenizer)) # Preprocessing the raw_datasets. # First we tokenize all the texts. padding = "max_length" if args.pad_to_max_length else False # Tokenize all texts and align the labels with them. def tokenize_and_align_labels(examples): tokenized_inputs = tokenizer( examples[text_column_name], max_length=args.max_length, padding=padding, truncation=True, # We use this argument because the texts in our dataset are lists of words (with a label for each word). is_split_into_words=True, ) labels = [] for i, label in enumerate(examples[label_column_name]): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label_to_id[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label_to_id[label[word_idx]] if args. label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs processed_raw_datasets = raw_datasets.map( tokenize_and_align_labels, batched=True, remove_columns=raw_datasets["train"].column_names) train_dataset = processed_raw_datasets["train"] eval_dataset = processed_raw_datasets["validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # DataLoaders creation: if args.pad_to_max_length: # If padding was already done ot max length, we use the default data collator that will just convert everything # to tensors. data_collator = default_data_collator else: # Otherwise, `DataCollatorForTokenClassification` will apply dynamic padding for us (by padding to the maximum length of # the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple # of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta). data_collator = DataCollatorForTokenClassification( tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)) train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size) eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [ p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) ], "weight_decay": args.weight_decay, }, { "params": [ p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) ], "weight_decay": 0.0, }, ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Use the device given by the `accelerator` object. device = accelerator.device model.to(device) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader) # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be # shorter in multiprocess) # Scheduler and math around the number of training steps. num_update_steps_per_epoch = math.ceil( len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch else: args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Metrics metric = load_metric("seqeval") def get_labels(predictions, references): # Transform predictions and references tensos to numpy arrays if device.type == "cpu": y_pred = predictions.detach().clone().numpy() y_true = references.detach().clone().numpy() else: y_pred = predictions.detach().cpu().clone().numpy() y_true = references.detach().cpu().clone().numpy() # Remove ignored index (special tokens) true_predictions = [[ label_list[p] for (p, l) in zip(pred, gold_label) if l != -100 ] for pred, gold_label in zip(y_pred, y_true)] true_labels = [[ label_list[l] for (p, l) in zip(pred, gold_label) if l != -100 ] for pred, gold_label in zip(y_pred, y_true)] return true_predictions, true_labels def compute_metrics(): results = metric.compute() if args.return_entity_level_metrics: # Unpack nested dictionaries final_results = {} for key, value in results.items(): if isinstance(value, dict): for n, v in value.items(): final_results[f"{key}_{n}"] = v else: final_results[key] = value return final_results else: return { "precision": results["overall_precision"], "recall": results["overall_recall"], "f1": results["overall_f1"], "accuracy": results["overall_accuracy"], } # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info( f" Instantaneous batch size per device = {args.per_device_train_batch_size}" ) logger.info( f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}" ) logger.info( f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 for epoch in range(args.num_train_epochs): model.train() for step, batch in enumerate(train_dataloader): outputs = model(**batch) loss = outputs.loss loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len( train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if completed_steps >= args.max_train_steps: break model.eval() for step, batch in enumerate(eval_dataloader): with torch.no_grad(): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) labels = batch["labels"] if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered predictions = accelerator.pad_across_processes(predictions, dim=1, pad_index=-100) labels = accelerator.pad_across_processes(labels, dim=1, pad_index=-100) predictions_gathered = accelerator.gather(predictions) labels_gathered = accelerator.gather(labels) preds, refs = get_labels(predictions_gathered, labels_gathered) metric.add_batch( predictions=preds, references=refs, ) # predictions and preferences are expected to be a nested list of labels, not label_ids # eval_metric = metric.compute() eval_metric = compute_metrics() accelerator.print(f"epoch {epoch}:", eval_metric) if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
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() # 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: 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("Training/evaluation parameters %s", training_args) # Set seed before initializing model. set_seed(training_args.seed) # Get the datasets: you can either provide your own CSV/JSON/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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. datasets = load_dataset('squad2.py', data_files={ 'train': 'khadija_data/train.json', 'dev': 'khadija_data/dev.json'}) else: data_files = {} if data_args.train_file is not None: data_files["train"] = data_args.train_file if data_args.validation_file is not None: data_files["validation"] = data_args.validation_file extension = data_args.train_file.split(".")[-1] datasets = load_dataset(extension, data_files=data_files, field="data") # 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, ) 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, use_fast=True, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForQuestionAnswering.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, ) # Tokenizer check: this script requires a fast tokenizer. if not isinstance(tokenizer, PreTrainedTokenizerFast): raise ValueError( "This example script only works for models that have a fast tokenizer. Checkout the big table of models " "at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this " "requirement" ) # Preprocessing the datasets. # Preprocessing is slighlty different for training and evaluation. if training_args.do_train: column_names = datasets["train"].column_names else: column_names = datasets["validation"].column_names question_column_name = "question" if "question" in column_names else column_names[0] context_column_name = "context" if "context" in column_names else column_names[1] answer_column_name = "answers" if "answers" in column_names else column_names[2] # Padding side determines if we do (question|context) or (context|question). pad_on_right = tokenizer.padding_side == "right" if data_args.max_seq_length > tokenizer.model_max_length: logger.warn( 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) # Training preprocessing def prepare_train_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[question_column_name if pad_on_right else context_column_name], examples[context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding="max_length" if data_args.pad_to_max_length else False, ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # The offset mappings will give us a map from token to character position in the original context. This will # help us compute the start_positions and end_positions. offset_mapping = tokenized_examples.pop("offset_mapping") # Let's label those examples! tokenized_examples["start_positions"] = [] tokenized_examples["end_positions"] = [] for i, offsets in enumerate(offset_mapping): # We will label impossible answers with the index of the CLS token. input_ids = tokenized_examples["input_ids"][i] cls_index = input_ids.index(tokenizer.cls_token_id) # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] answers = examples[answer_column_name][sample_index] # If no answers are given, set the cls_index as answer. if len(answers["answer_start"]) == 0: tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Start/end character index of the answer in the text. start_char = answers["answer_start"][0] end_char = start_char + len(answers["text"][0]) # Start token index of the current span in the text. token_start_index = 0 while sequence_ids[token_start_index] != (1 if pad_on_right else 0): token_start_index += 1 # End token index of the current span in the text. token_end_index = len(input_ids) - 1 while sequence_ids[token_end_index] != (1 if pad_on_right else 0): token_end_index -= 1 # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index). if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char): tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) else: # Otherwise move the token_start_index and token_end_index to the two ends of the answer. # Note: we could go after the last offset if the answer is the last word (edge case). while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char: token_start_index += 1 tokenized_examples["start_positions"].append( token_start_index - 1) # while offsets[token_end_index][1] >= end_char: # token_end_index -= 1 tokenized_examples["end_positions"].append( token_end_index + 1) return tokenized_examples if training_args.do_train: train_dataset = datasets["train"].map( prepare_train_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # Validation preprocessing def prepare_validation_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[question_column_name if pad_on_right else context_column_name], examples[context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, padding="max_length" if data_args.pad_to_max_length else False, ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the # corresponding example_id and we will store the offset mappings. tokenized_examples["example_id"] = [] for i in range(len(tokenized_examples["input_ids"])): # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples.sequence_ids(i) context_index = 1 if pad_on_right else 0 # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] tokenized_examples["example_id"].append( examples["id"][sample_index]) # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token # position is part of the context or not. tokenized_examples["offset_mapping"][i] = [ (o if sequence_ids[k] == context_index else None) for k, o in enumerate(tokenized_examples["offset_mapping"][i]) ] return tokenized_examples if training_args.do_eval: validation_dataset = datasets["validation"].map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # Data collator # We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data # collator. data_collator = ( default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None) ) # Post-processing: def post_processing_function(examples, features, predictions): # Post-processing: we match the start logits and end logits to answers in the original context. predictions = postprocess_qa_predictions( examples=examples, features=features, predictions=predictions, version_2_with_negative=data_args.version_2_with_negative, n_best_size=data_args.n_best_size, max_answer_length=data_args.max_answer_length, null_score_diff_threshold=data_args.null_score_diff_threshold, output_dir=training_args.output_dir, is_world_process_zero=trainer.is_world_process_zero(), ) # Format the result to the format the metric expects. if data_args.version_2_with_negative: formatted_predictions = [ {"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in predictions.items() ] else: formatted_predictions = [ {"id": k, "prediction_text": v} for k, v in predictions.items()] references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in datasets["validation"]] return EvalPrediction(predictions=formatted_predictions, label_ids=references) metric = load_metric( "squad_v2" if data_args.version_2_with_negative else "squad") def compute_metrics(p: EvalPrediction): return metric.compute(predictions=p.predictions, references=p.label_ids) # Initialize our Trainer trainer = QuestionAnsweringTrainer( model=model, args=training_args, train_dataset=train_dataset if training_args.do_train else None, eval_dataset=validation_dataset if training_args.do_eval else None, eval_examples=datasets["validation"] if training_args.do_eval else None, tokenizer=tokenizer, data_collator=data_collator, post_process_function=post_processing_function, compute_metrics=compute_metrics, ) # Training if training_args.do_train: if last_checkpoint is not None: checkpoint = last_checkpoint elif os.path.isdir(model_args.model_name_or_path): checkpoint = model_args.model_name_or_path else: checkpoint = None train_result = trainer.train(resume_from_checkpoint=checkpoint) trainer.save_model() # Saves the tokenizer too for easy upload output_train_file = os.path.join( training_args.output_dir, "train_results.txt") if trainer.is_world_process_zero(): with open(output_train_file, "w") as writer: logger.info("***** Train results *****") for key, value in sorted(train_result.metrics.items()): logger.info(f" {key} = {value}") writer.write(f"{key} = {value}\n") # 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")) # Evaluation results = {} if training_args.do_eval: logger.info("*** Evaluate ***") results = trainer.evaluate() output_eval_file = os.path.join( training_args.output_dir, "eval_results.txt") if trainer.is_world_process_zero(): with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key, value in sorted(results.items()): logger.info(f" {key} = {value}") writer.write(f"{key} = {value}\n") return results
def __init__(self): self.metric = load_metric("seqeval")
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, MultiLingAdapterArguments)) 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, adapter_args = parser.parse_json_file( json_file=os.path.abspath(sys.argv[1])) else: model_args, data_args, training_args, adapter_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 CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named # label if at least two columns are provided. # # 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.task_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset("glue", data_args.task_name, cache_dir=model_args.cache_dir) elif 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 if data_args.task_name is not None: is_regression = data_args.task_name == "stsb" if not is_regression: label_list = raw_datasets["train"].features["label"].names num_labels = len(label_list) else: num_labels = 1 else: # Trying to have good defaults here, don't hesitate to tweak to your needs. is_regression = raw_datasets["train"].features["label"].dtype in [ "float32", "float64" ] if is_regression: num_labels = 1 else: # A useful fast method: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique label_list = raw_datasets["train"].unique("label") label_list.sort() # Let's sort it for determinism 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, finetuning_task=data_args.task_name, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) 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, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForSequenceClassification.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, ) # Setup adapters if adapter_args.train_adapter: task_name = data_args.task_name or "glue" # check if adapter already exists, otherwise add it if task_name not in model.config.adapters: # resolve the adapter config adapter_config = AdapterConfig.load( adapter_args.adapter_config, non_linearity=adapter_args.adapter_non_linearity, reduction_factor=adapter_args.adapter_reduction_factor, ) # load a pre-trained from Hub if specified if adapter_args.load_adapter: model.load_adapter( adapter_args.load_adapter, config=adapter_config, load_as=task_name, ) # otherwise, add a fresh adapter else: model.add_adapter(task_name, config=adapter_config) # optionally load a pre-trained language adapter if adapter_args.load_lang_adapter: # resolve the language adapter config lang_adapter_config = AdapterConfig.load( adapter_args.lang_adapter_config, non_linearity=adapter_args.lang_adapter_non_linearity, reduction_factor=adapter_args.lang_adapter_reduction_factor, ) # load the language adapter from Hub lang_adapter_name = model.load_adapter( adapter_args.load_lang_adapter, config=lang_adapter_config, load_as=adapter_args.language, ) else: lang_adapter_name = None # Freeze all model weights except of those of this adapter model.train_adapter([task_name]) # Set the adapters to be used in every forward pass if lang_adapter_name: model.set_active_adapters(ac.Stack(lang_adapter_name, task_name)) else: model.set_active_adapters([task_name]) else: if adapter_args.load_adapter or adapter_args.load_lang_adapter: raise ValueError( "Adapters can only be loaded in adapters training mode." "Use --train_adapter to enable adapter training") # Preprocessing the datasets if data_args.task_name is not None: sentence1_key, sentence2_key = task_to_keys[data_args.task_name] else: # Again, we try to have some nice defaults but don't hesitate to tweak to your use case. non_label_column_names = [ name for name in raw_datasets["train"].column_names if name != "label" ] if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names: sentence1_key, sentence2_key = "sentence1", "sentence2" else: if len(non_label_column_names) >= 2: sentence1_key, sentence2_key = non_label_column_names[:2] else: sentence1_key, sentence2_key = non_label_column_names[0], 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. label_to_id = None if (model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id and data_args.task_name is not None and not is_regression): # Some have all caps in their config, some don't. label_name_to_id = { k.lower(): v for k, v in model.config.label2id.items() } if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)): label_to_id = { i: int(label_name_to_id[label_list[i]]) for i in range(num_labels) } else: logger.warning( "Your model seems to have been trained with labels, but they don't match the dataset: ", f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}." "\nIgnoring the model labels as a result.", ) elif data_args.task_name is None and not is_regression: label_to_id = {v: i for i, v in enumerate(label_list)} if label_to_id is not None: model.config.label2id = label_to_id model.config.id2label = { id: label for label, id in config.label2id.items() } elif data_args.task_name is not None and not is_regression: model.config.label2id = {l: i for i, l in enumerate(label_list)} model.config.id2label = { id: label for label, id in config.label2id.items() } 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_function(examples): # Tokenize the texts args = ((examples[sentence1_key], ) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])) result = tokenizer(*args, padding=padding, max_length=max_seq_length, truncation=True) # Map labels to IDs (not necessary for GLUE tasks) if label_to_id is not None and "label" in examples: result["label"] = [(label_to_id[l] if l != -1 else -1) for l in examples["label"]] return result with training_args.main_process_first(desc="dataset map pre-processing"): raw_datasets = raw_datasets.map( preprocess_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_matched" if data_args. task_name == "mnli" else "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.task_name is not None 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_matched" if data_args.task_name == "mnli" else "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]}.") # Get the metric function if data_args.task_name is not None: metric = load_metric("glue", data_args.task_name) else: metric = load_metric("accuracy") # 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.squeeze(preds) if is_regression else np.argmax(preds, axis=1) if data_args.task_name is not None: result = metric.compute(predictions=preds, references=p.label_ids) if len(result) > 1: result["combined_score"] = np.mean(list( result.values())).item() return result elif is_regression: return {"mse": ((preds - p.label_ids)**2).mean().item()} else: 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_class = AdapterTrainer if adapter_args.train_adapter else Trainer trainer = trainer_class( 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 ***") # Loop to handle MNLI double evaluation (matched, mis-matched) tasks = [data_args.task_name] eval_datasets = [eval_dataset] if data_args.task_name == "mnli": tasks.append("mnli-mm") eval_datasets.append(raw_datasets["validation_mismatched"]) for eval_dataset, task in zip(eval_datasets, tasks): 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 ***") # Loop to handle MNLI double evaluation (matched, mis-matched) tasks = [data_args.task_name] predict_datasets = [predict_dataset] if data_args.task_name == "mnli": tasks.append("mnli-mm") predict_datasets.append(raw_datasets["test_mismatched"]) for predict_dataset, task in zip(predict_datasets, tasks): # 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.squeeze( predictions) if is_regression else np.argmax(predictions, axis=1) output_predict_file = os.path.join(training_args.output_dir, f"predict_results_{task}.txt") if trainer.is_world_process_zero(): with open(output_predict_file, "w") as writer: logger.info(f"***** Predict results {task} *****") writer.write("index\tprediction\n") for index, item in enumerate(predictions): if is_regression: writer.write(f"{index}\t{item:3.3f}\n") else: item = label_list[item] writer.write(f"{index}\t{item}\n") kwargs = { "finetuned_from": model_args.model_name_or_path, "tasks": "text-classification" } if data_args.task_name is not None: kwargs["language"] = "en" kwargs["dataset_tags"] = "glue" kwargs["dataset_args"] = data_args.task_name kwargs["dataset"] = f"GLUE {data_args.task_name.upper()}" 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, 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( ) 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 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() transformers.utils.logging.enable_default_handler() transformers.utils.logging.enable_explicit_format() 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 specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub). # # For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named # label if at least two columns are provided. # # 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.task_name == 'imdb': datasets = load_dataset('imdb') elif data_args.task_name is not None: # Downloading and loading a dataset from the hub. datasets = load_dataset("glue", data_args.task_name) 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 datasets = load_dataset("csv", data_files=data_files) else: # Loading a dataset from local json files datasets = load_dataset("json", data_files=data_files) # See more about loading any type of standard or custom dataset at # https://huggingface.co/docs/datasets/loading_datasets.html. # Labels if data_args.task_name is not None: is_regression = data_args.task_name == "stsb" if not is_regression: label_list = datasets["train"].features["label"].names num_labels = len(label_list) else: num_labels = 1 else: # Trying to have good defaults here, don't hesitate to tweak to your needs. is_regression = datasets["train"].features["label"].dtype in [ "float32", "float64" ] if is_regression: num_labels = 1 else: # A useful fast method: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique label_list = datasets["train"].unique("label") label_list.sort() # Let's sort it for determinism 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, finetuning_task=data_args.task_name, cache_dir=model_args.cache_dir, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) 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, use_fast=model_args.use_fast_tokenizer, revision=model_args.model_revision, use_auth_token=True if model_args.use_auth_token else None, ) model = AutoModelForSequenceClassification.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, ) # Preprocessing the datasets if data_args.task_name is not None: sentence1_key, sentence2_key = task_to_keys[data_args.task_name] else: # Again, we try to have some nice defaults but don't hesitate to tweak to your use case. non_label_column_names = [ name for name in datasets["train"].column_names if name != "label" ] if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names: sentence1_key, sentence2_key = "sentence1", "sentence2" else: if len(non_label_column_names) >= 2: sentence1_key, sentence2_key = non_label_column_names[:2] else: sentence1_key, sentence2_key = non_label_column_names[0], 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. label_to_id = None if (model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id and data_args.task_name is not None and is_regression): # Some have all caps in their config, some don't. label_name_to_id = { k.lower(): v for k, v in model.config.label2id.items() } if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)): label_to_id = { i: label_name_to_id[label_list[i]] for i in range(num_labels) } else: logger.warn( "Your model seems to have been trained with labels, but they don't match the dataset: ", f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}." "\nIgnoring the model labels as a result.", ) elif data_args.task_name is None and not is_regression: label_to_id = {v: i for i, v in enumerate(label_list)} def preprocess_function(examples): # Tokenize the texts args = ((examples[sentence1_key], ) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])) result = tokenizer(*args, padding=padding, max_length=data_args.max_seq_length, truncation=True) # Map labels to IDs (not necessary for GLUE tasks) if label_to_id is not None and "label" in examples: result["label"] = [label_to_id[l] for l in examples["label"]] return result datasets = datasets.map(preprocess_function, batched=True, load_from_cache_file=not data_args.overwrite_cache) train_dataset = datasets["train"] if data_args.task_name == 'imdb': eval_dataset = test_dataset = datasets['test'] else: eval_dataset = datasets["validation_matched" if data_args.task_name == "mnli" else "validation"] if data_args.task_name is not None or data_args.test_file is not None: test_dataset = datasets["test_matched" if data_args.task_name == "mnli" else "test"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # Get the metric function if data_args.task_name == 'imdb': metric = load_metric("accuracy") elif data_args.task_name is not None: metric = load_metric("glue", data_args.task_name) # TODO: When datasets metrics include regular accuracy, make an else here and remove special branch from # compute_metrics # 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.squeeze(preds) if is_regression else np.argmax(preds, axis=1) if data_args.task_name is not None: result = metric.compute(predictions=preds, references=p.label_ids) if len(result) > 1: result["combined_score"] = np.mean(list( result.values())).item() return result elif is_regression: return {"mse": ((preds - p.label_ids)**2).mean().item()} else: return { "accuracy": (preds == p.label_ids).astype(np.float32).mean().item() } # Initialize our Trainer trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset if training_args.do_eval else None, compute_metrics=compute_metrics, tokenizer=tokenizer, # Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding. data_collator=default_data_collator if data_args.pad_to_max_length else None, ) # Training if training_args.do_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 trainer.save_model() # Saves the tokenizer too for easy upload output_train_file = os.path.join(training_args.output_dir, "train_results.txt") if trainer.is_world_process_zero(): with open(output_train_file, "w") as writer: logger.info("***** Train results *****") for key, value in sorted(metrics.items()): logger.info(f" {key} = {value}") writer.write(f"{key} = {value}\n") # 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")) # Evaluation eval_results = {} if training_args.do_eval: logger.info("*** Evaluate ***") # Loop to handle MNLI double evaluation (matched, mis-matched) tasks = [data_args.task_name] eval_datasets = [eval_dataset] if data_args.task_name == "mnli": tasks.append("mnli-mm") eval_datasets.append(datasets["validation_mismatched"]) for eval_dataset, task in zip(eval_datasets, tasks): eval_result = trainer.evaluate(eval_dataset=eval_dataset) output_eval_file = os.path.join(training_args.output_dir, f"eval_results_{task}.txt") if trainer.is_world_process_zero(): with open(output_eval_file, "w") as writer: logger.info(f"***** Eval results {task} *****") for key, value in sorted(eval_result.items()): logger.info(f" {key} = {value}") writer.write(f"{key} = {value}\n") eval_results.update(eval_result) if training_args.do_predict: logger.info("*** Test ***") # Loop to handle MNLI double evaluation (matched, mis-matched) tasks = [data_args.task_name] test_datasets = [test_dataset] if data_args.task_name == "mnli": tasks.append("mnli-mm") test_datasets.append(datasets["test_mismatched"]) for test_dataset, task in zip(test_datasets, tasks): # Removing the `label` columns because it contains -1 and Trainer won't like that. test_dataset.remove_columns_("label") predictions = trainer.predict( test_dataset=test_dataset).predictions predictions = np.squeeze( predictions) if is_regression else np.argmax(predictions, axis=1) output_test_file = os.path.join(training_args.output_dir, f"test_results_{task}.txt") if trainer.is_world_process_zero(): with open(output_test_file, "w") as writer: logger.info(f"***** Test results {task} *****") writer.write("index\tprediction\n") for index, item in enumerate(predictions): if is_regression: writer.write(f"{index}\t{item:3.3f}\n") else: item = label_list[item] writer.write(f"{index}\t{item}\n") return eval_results
def __init__(self, context: det_torch.PyTorchTrialContext) -> None: self.logger = logging.getLogger(__name__) self.hparams = attrdict.AttrDict(context.get_hparams()) self.data_config = attrdict.AttrDict(context.get_data_config()) self.context = context # Check to make sure the dataset is configured correctly. if self.data_config.dataset_name is not None: dataset_name = self.data_config.dataset_name if dataset_name == "squad": assert (not self.data_config.version_2_with_negative ), "version_2_with_negative should be false for squad" elif dataset_name == "squad_v2": assert ( self.data_config.version_2_with_negative ), "version_2_with_negative should be true for squad_v2" self.data_processors = data_beam_search # Get the datasets: you can either provide your own CSV or 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 CSV/JSON files, this script will use the column called 'text' or the first column if # no column called 'text' is found. You can easily tweak this behavior (see below). # 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. self.raw_datasets = hf.default_load_dataset(self.data_config) self.column_names = self.raw_datasets["train"].column_names # For beam search, we need to use a different model from the default model returned by # AutoModelForQuestionAnswering. We will use a custom init in this case that is a slight # modification of the BaseTransformerTrial init method. self.exp_config = attrdict.AttrDict(context.get_experiment_config()) # Check to make sure all expected hyperparameters are set. self.check_hparams() # Parse hparams and data_config. ( self.config_kwargs, self.tokenizer_kwargs, self.model_kwargs, ) = hf.default_parse_config_tokenizer_model_kwargs(self.hparams) optimizer_kwargs, scheduler_kwargs = hf.default_parse_optimizer_lr_scheduler_kwargs( self.hparams) self.config = transformers.XLNetConfig.from_pretrained( **self.config_kwargs) self.tokenizer = transformers.XLNetTokenizerFast.from_pretrained( **self.tokenizer_kwargs) # We need to use XLNetForQuestionAnswering instead of XLNetForQuestionAnsweringSimple # which is the default returned by AutoModelForQuestionAnswering. if self.hparams.use_pretrained_weights: self.model_kwargs["config"] = self.config self.model = transformers.XLNetForQuestionAnswering.from_pretrained( **self.model_kwargs) else: self.model = transformers.XLNetForQuestionAnswering(self.config) self.model = self.context.wrap_model(self.model) # The rest is the same as the parent init method. self.optimizer = self.context.wrap_optimizer( hf.build_default_optimizer(self.model, optimizer_kwargs)) if self.hparams.use_apex_amp: self.model, self.optimizer = self.context.configure_apex_amp( models=self.model, optimizers=self.optimizer, ) self.lr_scheduler = self.context.wrap_lr_scheduler( hf.build_default_lr_scheduler(self.optimizer, scheduler_kwargs), det_torch.LRScheduler.StepMode.STEP_EVERY_BATCH, ) self.grad_clip_fn = ( lambda x: torch.nn.utils.clip_grad_norm_( x, optimizer_kwargs.max_grad_norm) if optimizer_kwargs.max_grad_norm > 0 # type: ignore else None) self.logger.info(self.config) if not isinstance(self.tokenizer, transformers.PreTrainedTokenizerFast): raise ValueError( "This example script only works for models that have a fast tokenizer. Checkout " "the big table of models at " "https://huggingface.co/transformers/index.html#bigtable to find the model types " "that meet this requirement") # We need to create the tokenized dataset after init because we need to model and # tokenizer to be available. self.tokenized_datasets = self.build_datasets() train_length = len(self.tokenized_datasets["train"]) self.logger.info("training records: {}".format(train_length)) if ("records_per_epoch" in self.exp_config and train_length != self.exp_config["records_per_epoch"]): self.logger.warning( "number of train records {} does not match records_per_epoch of {}" .format(train_length, self.exp_config["records_per_epoch"])) # Create metric reducer metric = datasets.load_metric("squad_v2" if self.data_config. version_2_with_negative else "squad") def compute_metrics(predictions): predictions = zip(*predictions) predictions = [utils.expand_like(p) for p in predictions] # We need to add back in columns needed for validation. self.tokenized_datasets["validation"].set_format( type=self.tokenized_datasets["validation"].format["type"], columns=list( self.tokenized_datasets["validation"].features.keys()), ) output = self.data_processors.post_processing_function( examples=self.raw_datasets["validation"], features=self.tokenized_datasets["validation"], predictions=predictions, data_args=self.data_config, column_names=self.column_names, prefix="eval", model=self.model, ) result = metric.compute(predictions=output.predictions, references=output.label_ids) # Then remove them again so that data collation doesn't break. hf.remove_unused_columns(self.model, self.tokenized_datasets["validation"]) return result self.reducer = context.experimental.wrap_reducer(compute_metrics, for_training=False)
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( ) 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.") # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) # Setup logging, we only want one process per machine to log things on the screen. logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR) if jax.process_index() == 0: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # Set the verbosity to info of the Transformers logger (on main process only): logger.info(f"Training/evaluation parameters {training_args}") # 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 CSV/JSON files this script will use the first column for the full texts and the second column for the # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments). # if data_args.dataset_name is not None: # Downloading and loading a dataset from the hub. dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False) 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] dataset = 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 if model_args.config_name: config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir) elif model_args.model_name_or_path: config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir) else: config = CONFIG_MAPPING[model_args.model_type]() logger.warning( "You are instantiating a new config instance from scratch.") if model_args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer) elif model_args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if model_args.model_name_or_path: model = FlaxAutoModelForSeq2SeqLM.from_pretrained( model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)) else: model = FlaxAutoModelForSeq2SeqLM.from_config(config, seed=training_args.seed, dtype=getattr( jnp, model_args.dtype)) if model.config.decoder_start_token_id is None: raise ValueError( "Make sure that `config.decoder_start_token_id` is correctly defined" ) prefix = data_args.source_prefix if data_args.source_prefix is not None else "" # Preprocessing the datasets. # We need to tokenize inputs and targets. if training_args.do_train: column_names = dataset["train"].column_names elif training_args.do_eval: column_names = dataset["validation"].column_names elif training_args.do_predict: column_names = dataset["test"].column_names else: logger.info( "There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`." ) return # Get the column names for input/target. dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None) if data_args.text_column is None: text_column = dataset_columns[ 0] if dataset_columns is not None else column_names[0] else: text_column = data_args.text_column if text_column not in column_names: raise ValueError( f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}" ) if data_args.summary_column is None: summary_column = dataset_columns[ 1] if dataset_columns is not None else column_names[1] else: summary_column = data_args.summary_column if summary_column not in column_names: raise ValueError( f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}" ) # Temporarily set max_target_length for training. max_target_length = data_args.max_target_length # In Flax, for seq2seq models we need to pass `decoder_input_ids` # as the Flax models don't accept `labels`, we need to prepare the decoder_input_ids here # for that dynamically import the `shift_tokens_right` function from the model file model_module = __import__(model.__module__, fromlist=["shift_tokens_tight"]) shift_tokens_right_fn = getattr(model_module, "shift_tokens_right") # Setting padding="max_length" as we need fixed length inputs for jitted functions def preprocess_function(examples): inputs = examples[text_column] targets = examples[summary_column] inputs = [prefix + inp for inp in inputs] model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding="max_length", truncation=True, return_tensors="np") # Setup the tokenizer for targets with tokenizer.as_target_tokenizer(): labels = tokenizer(targets, max_length=max_target_length, padding="max_length", truncation=True, return_tensors="np") model_inputs["labels"] = labels["input_ids"] decoder_input_ids = shift_tokens_right_fn( jnp.array(labels["input_ids"]), config.pad_token_id, config.decoder_start_token_id) model_inputs["decoder_input_ids"] = np.asarray(decoder_input_ids) # We need decoder_attention_mask so we can ignore pad tokens from loss model_inputs["decoder_attention_mask"] = labels["attention_mask"] return model_inputs if training_args.do_train: if "train" not in dataset: raise ValueError("--do_train requires a train dataset") train_dataset = dataset["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_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on train dataset", ) if training_args.do_eval: max_target_length = data_args.val_max_target_length if "validation" not in dataset: raise ValueError("--do_eval requires a validation dataset") eval_dataset = dataset["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_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on validation dataset", ) if training_args.do_predict: max_target_length = data_args.val_max_target_length if "test" not in dataset: raise ValueError("--do_predict requires a test dataset") predict_dataset = dataset["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_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on prediction dataset", ) # Metric metric = load_metric("rouge") def postprocess_text(preds, labels): preds = [pred.strip() for pred in preds] labels = [label.strip() for label in labels] # rougeLSum expects newline after each sentence preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds] labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels] return preds, labels def compute_metrics(preds, labels): decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True) decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True) # Some simple post-processing decoded_preds, decoded_labels = postprocess_text( decoded_preds, decoded_labels) result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True) # Extract a few results from ROUGE result = { key: value.mid.fmeasure * 100 for key, value in result.items() } prediction_lens = [ np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds ] result["gen_len"] = np.mean(prediction_lens) result = {k: round(v, 4) for k, v in result.items()} return result # Enable tensorboard only on the master node has_tensorboard = is_tensorboard_available() if has_tensorboard and jax.process_index() == 0: try: from flax.metrics.tensorboard import SummaryWriter summary_writer = SummaryWriter(log_dir=Path( training_args.output_dir).joinpath("logs").as_posix()) except ImportError as ie: has_tensorboard = False logger.warning( f"Unable to display metrics through TensorBoard because some package are not installed: {ie}" ) else: logger.warning( "Unable to display metrics through TensorBoard because the package is not installed: " "Please run pip install tensorboard to enable.") # Initialize our training rng = jax.random.PRNGKey(training_args.seed) rng, dropout_rng = jax.random.split(rng) # Store some constant num_epochs = int(training_args.num_train_epochs) train_batch_size = int( training_args.per_device_train_batch_size) * jax.device_count() eval_batch_size = int( training_args.per_device_eval_batch_size) * jax.device_count() steps_per_epoch = len(train_dataset) // train_batch_size total_train_steps = steps_per_epoch * num_epochs # Create learning rate schedule linear_decay_lr_schedule_fn = create_learning_rate_fn( len(train_dataset), train_batch_size, training_args.num_train_epochs, training_args.warmup_steps, training_args.learning_rate, ) # We use Optax's "masking" functionality to not apply weight decay # to bias and LayerNorm scale parameters. decay_mask_fn returns a # mask boolean with the same structure as the parameters. # The mask is True for parameters that should be decayed. def decay_mask_fn(params): flat_params = traverse_util.flatten_dict(params) flat_mask = { path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params } return traverse_util.unflatten_dict(flat_mask) # create adam optimizer adamw = optax.adamw( learning_rate=linear_decay_lr_schedule_fn, b1=training_args.adam_beta1, b2=training_args.adam_beta2, eps=training_args.adam_epsilon, weight_decay=training_args.weight_decay, mask=decay_mask_fn, ) # Setup train state state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng) # label smoothed cross entropy def loss_fn(logits, labels, padding_mask, label_smoothing_factor=0.0): """ The label smoothing implementation is adapted from Flax's official example: https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104 """ vocab_size = logits.shape[-1] confidence = 1.0 - label_smoothing_factor low_confidence = (1.0 - confidence) / (vocab_size - 1) normalizing_constant = -(confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)) soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence) loss = optax.softmax_cross_entropy(logits, soft_labels) loss = loss - normalizing_constant # ignore padded tokens from loss loss = loss * padding_mask loss = loss.sum() / padding_mask.sum() return loss # Define gradient update step fn def train_step(state, batch, label_smoothing_factor=0.0): dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng) def compute_loss(params): labels = batch.pop("labels") logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0] loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor) return loss grad_fn = jax.value_and_grad(compute_loss) loss, grad = grad_fn(state.params) grad = jax.lax.pmean(grad, "batch") new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng) metrics = { "loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step) } metrics = jax.lax.pmean(metrics, axis_name="batch") return new_state, metrics # Define eval fn def eval_step(params, batch, label_smoothing_factor=0.0): labels = batch.pop("labels") logits = model(**batch, params=params, train=False)[0] loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor) # summarize metrics metrics = {"loss": loss} metrics = jax.lax.pmean(metrics, axis_name="batch") return metrics # Define generation function max_length = (data_args.val_max_target_length if data_args.val_max_target_length is not None else model.config.max_length) num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams gen_kwargs = {"max_length": max_length, "num_beams": num_beams} def generate_step(params, batch): model.params = params output_ids = model.generate(batch["input_ids"], attention_mask=batch["attention_mask"], **gen_kwargs) return output_ids.sequences # Create parallel version of the train and eval step p_train_step = jax.pmap(partial( train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0, )) p_eval_step = jax.pmap( partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch") p_generate_step = jax.pmap(generate_step, "batch") # Replicate the train state on each device state = state.replicate() logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {num_epochs}") logger.info( f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}" ) logger.info( f" Total train batch size (w. parallel & distributed) = {train_batch_size}" ) logger.info(f" Total optimization steps = {total_train_steps}") train_time = 0 epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0) for epoch in epochs: # ======================== Training ================================ train_start = time.time() # Create sampling rng rng, input_rng = jax.random.split(rng) train_metrics = [] # Generate an epoch by shuffling sampling indices from the train dataset train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True) steps_per_epoch = len(train_dataset) // train_batch_size # train for _ in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False): batch = next(train_loader) state, train_metric = p_train_step(state, batch) train_metrics.append(train_metric) train_time += time.time() - train_start train_metric = unreplicate(train_metric) epochs.write( f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})" ) # ======================== Evaluating ============================== eval_metrics = [] eval_preds = [] eval_labels = [] eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size) eval_steps = len(eval_dataset) // eval_batch_size for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False): # Model forward batch = next(eval_loader) labels = batch["labels"] metrics = p_eval_step(state.params, batch) eval_metrics.append(metrics) # generation if data_args.predict_with_generate: generated_ids = p_generate_step(state.params, batch) eval_preds.extend( jax.device_get( generated_ids.reshape(-1, gen_kwargs["max_length"]))) eval_labels.extend( jax.device_get(labels.reshape(-1, labels.shape[-1]))) # normalize eval metrics eval_metrics = get_metrics(eval_metrics) eval_metrics = jax.tree_map(jnp.mean, eval_metrics) # compute ROUGE metrics rouge_desc = "" if data_args.predict_with_generate: rouge_metrics = compute_metrics(eval_preds, eval_labels) eval_metrics.update(rouge_metrics) rouge_desc = " ".join([ f"Eval {key}: {value} |" for key, value in rouge_metrics.items() ]) # Print metrics and update progress bar desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {rouge_desc})" epochs.write(desc) epochs.desc = desc # Save metrics if has_tensorboard and jax.process_index() == 0: cur_step = epoch * (len(train_dataset) // train_batch_size) write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step) # ======================== Prediction loop ============================== if training_args.do_predict: logger.info("*** Predict ***") pred_metrics = [] pred_generations = [] pred_labels = [] pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size) pred_steps = len(predict_dataset) // eval_batch_size for _ in tqdm(range(pred_steps), desc="Predicting...", position=2, leave=False): # Model forward batch = next(pred_loader) labels = batch["labels"] metrics = p_eval_step(state.params, batch) pred_metrics.append(metrics) # generation if data_args.predict_with_generate: generated_ids = p_generate_step(state.params, batch) pred_generations.extend( jax.device_get( generated_ids.reshape(-1, gen_kwargs["max_length"]))) pred_labels.extend( jax.device_get(labels.reshape(-1, labels.shape[-1]))) # normalize prediction metrics pred_metrics = get_metrics(pred_metrics) pred_metrics = jax.tree_map(jnp.mean, pred_metrics) # compute ROUGE metrics rouge_desc = "" if data_args.predict_with_generate: rouge_metrics = compute_metrics(pred_generations, pred_labels) pred_metrics.update(rouge_metrics) rouge_desc = " ".join([ f"Predict {key}: {value} |" for key, value in rouge_metrics.items() ]) # Print metrics desc = f"Predict Loss: {pred_metrics['loss']} | {rouge_desc})" logger.info(desc) # save last checkpoint if jax.process_index() == 0: params = jax.device_get(unreplicate(state.params)) model.save_pretrained(training_args.output_dir, params=params)
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/TXT 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 CSV/JSON files, this script will use the column called 'text' or the first column if no column called # 'text' is found. 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: 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] raw_datasets = load_dataset(extension, data_files=data_files, field="data", 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 = XLNetConfig.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, ) tokenizer = XLNetTokenizerFast.from_pretrained( model_args.tokenizer_name if model_args.tokenizer_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, ) model = XLNetForQuestionAnswering.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, ) # Preprocessing the datasets. # Preprocessing is slighlty different for training and evaluation. if training_args.do_train: column_names = raw_datasets["train"].column_names elif training_args.do_eval: column_names = raw_datasets["validation"].column_names else: column_names = raw_datasets["test"].column_names question_column_name = "question" if "question" in column_names else column_names[ 0] context_column_name = "context" if "context" in column_names else column_names[ 1] answer_column_name = "answers" if "answers" in column_names else column_names[ 2] # Padding side determines if we do (question|context) or (context|question). pad_on_right = tokenizer.padding_side == "right" 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) # Training preprocessing def prepare_train_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[ question_column_name if pad_on_right else context_column_name], examples[ context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, return_special_tokens_mask=True, return_token_type_ids=True, padding="max_length", ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # The offset mappings will give us a map from token to character position in the original context. This will # help us compute the start_positions and end_positions. offset_mapping = tokenized_examples.pop("offset_mapping") # The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers). special_tokens = tokenized_examples.pop("special_tokens_mask") # Let's label those examples! tokenized_examples["start_positions"] = [] tokenized_examples["end_positions"] = [] tokenized_examples["is_impossible"] = [] tokenized_examples["cls_index"] = [] tokenized_examples["p_mask"] = [] for i, offsets in enumerate(offset_mapping): # We will label impossible answers with the index of the CLS token. input_ids = tokenized_examples["input_ids"][i] cls_index = input_ids.index(tokenizer.cls_token_id) tokenized_examples["cls_index"].append(cls_index) # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples["token_type_ids"][i] for k, s in enumerate(special_tokens[i]): if s: sequence_ids[k] = 3 context_idx = 1 if pad_on_right else 0 # Build the p_mask: non special tokens and context gets 0.0, the others get 1.0. # The cls token gets 1.0 too (for predictions of empty answers). tokenized_examples["p_mask"].append([ 0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0 for k, s in enumerate(sequence_ids) ]) # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] answers = examples[answer_column_name][sample_index] # If no answers are given, set the cls_index as answer. if len(answers["answer_start"]) == 0: tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) tokenized_examples["is_impossible"].append(1.0) else: # Start/end character index of the answer in the text. start_char = answers["answer_start"][0] end_char = start_char + len(answers["text"][0]) # Start token index of the current span in the text. token_start_index = 0 while sequence_ids[token_start_index] != context_idx: token_start_index += 1 # End token index of the current span in the text. token_end_index = len(input_ids) - 1 while sequence_ids[token_end_index] != context_idx: token_end_index -= 1 # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index). if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char): tokenized_examples["start_positions"].append(cls_index) tokenized_examples["end_positions"].append(cls_index) tokenized_examples["is_impossible"].append(1.0) else: # Otherwise move the token_start_index and token_end_index to the two ends of the answer. # Note: we could go after the last offset if the answer is the last word (edge case). while token_start_index < len(offsets) and offsets[ token_start_index][0] <= start_char: token_start_index += 1 tokenized_examples["start_positions"].append( token_start_index - 1) while offsets[token_end_index][1] >= end_char: token_end_index -= 1 tokenized_examples["end_positions"].append( token_end_index + 1) tokenized_examples["is_impossible"].append(0.0) return tokenized_examples 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: # Select samples from Dataset, This will help to decrease processing time train_dataset = train_dataset.select( range(data_args.max_train_samples)) # Create Training Features with training_args.main_process_first( desc="train dataset map pre-processing"): train_dataset = train_dataset.map( prepare_train_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on train dataset", ) if data_args.max_train_samples is not None: # Select samples from dataset again since Feature Creation might increase number of features train_dataset = train_dataset.select( range(data_args.max_train_samples)) # Validation preprocessing def prepare_validation_features(examples): # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results # in one example possible giving several features when a context is long, each of those features having a # context that overlaps a bit the context of the previous feature. tokenized_examples = tokenizer( examples[ question_column_name if pad_on_right else context_column_name], examples[ context_column_name if pad_on_right else question_column_name], truncation="only_second" if pad_on_right else "only_first", max_length=max_seq_length, stride=data_args.doc_stride, return_overflowing_tokens=True, return_offsets_mapping=True, return_special_tokens_mask=True, return_token_type_ids=True, padding="max_length", ) # Since one example might give us several features if it has a long context, we need a map from a feature to # its corresponding example. This key gives us just that. sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") # The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers). special_tokens = tokenized_examples.pop("special_tokens_mask") # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the # corresponding example_id and we will store the offset mappings. tokenized_examples["example_id"] = [] # We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label. tokenized_examples["cls_index"] = [] tokenized_examples["p_mask"] = [] for i, input_ids in enumerate(tokenized_examples["input_ids"]): # Find the CLS token in the input ids. cls_index = input_ids.index(tokenizer.cls_token_id) tokenized_examples["cls_index"].append(cls_index) # Grab the sequence corresponding to that example (to know what is the context and what is the question). sequence_ids = tokenized_examples["token_type_ids"][i] for k, s in enumerate(special_tokens[i]): if s: sequence_ids[k] = 3 context_idx = 1 if pad_on_right else 0 # Build the p_mask: non special tokens and context gets 0.0, the others 1.0. tokenized_examples["p_mask"].append([ 0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0 for k, s in enumerate(sequence_ids) ]) # One example can give several spans, this is the index of the example containing this span of text. sample_index = sample_mapping[i] tokenized_examples["example_id"].append( examples["id"][sample_index]) # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token # position is part of the context or not. tokenized_examples["offset_mapping"][i] = [ (o if sequence_ids[k] == context_idx else None) for k, o in enumerate(tokenized_examples["offset_mapping"][i]) ] return tokenized_examples if training_args.do_eval: if "validation" not in raw_datasets: raise ValueError("--do_eval requires a validation dataset") eval_examples = raw_datasets["validation"] if data_args.max_eval_samples is not None: # Selecting Eval Samples from Dataset eval_examples = eval_examples.select( range(data_args.max_eval_samples)) # Create Features from Eval Dataset with training_args.main_process_first( desc="validation dataset map pre-processing"): eval_dataset = eval_examples.map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on validation dataset", ) if data_args.max_eval_samples is not None: # Selecting Samples from Dataset again since Feature Creation might increase samples size eval_dataset = eval_dataset.select( range(data_args.max_eval_samples)) if training_args.do_predict: if "test" not in raw_datasets: raise ValueError("--do_predict requires a test dataset") predict_examples = raw_datasets["test"] if data_args.max_predict_samples is not None: # We will select sample from whole data predict_examples = predict_examples.select( range(data_args.max_predict_samples)) # Test Feature Creation with training_args.main_process_first( desc="prediction dataset map pre-processing"): predict_dataset = predict_examples.map( prepare_validation_features, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, desc="Running tokenizer on prediction dataset", ) if data_args.max_predict_samples is not None: # During Feature creation dataset samples might increase, we will select required samples again predict_dataset = predict_dataset.select( range(data_args.max_predict_samples)) # Data collator # We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data # collator. data_collator = (default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding( tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)) # Post-processing: def post_processing_function(examples, features, predictions, stage="eval"): # Post-processing: we match the start logits and end logits to answers in the original context. predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search( examples=examples, features=features, predictions=predictions, version_2_with_negative=data_args.version_2_with_negative, n_best_size=data_args.n_best_size, max_answer_length=data_args.max_answer_length, start_n_top=model.config.start_n_top, end_n_top=model.config.end_n_top, output_dir=training_args.output_dir, log_level=log_level, prefix=stage, ) # Format the result to the format the metric expects. if data_args.version_2_with_negative: formatted_predictions = [{ "id": k, "prediction_text": v, "no_answer_probability": scores_diff_json[k] } for k, v in predictions.items()] else: formatted_predictions = [{ "id": k, "prediction_text": v } for k, v in predictions.items()] references = [{ "id": ex["id"], "answers": ex[answer_column_name] } for ex in examples] return EvalPrediction(predictions=formatted_predictions, label_ids=references) metric = load_metric( "squad_v2" if data_args.version_2_with_negative else "squad") def compute_metrics(p: EvalPrediction): return metric.compute(predictions=p.predictions, references=p.label_ids) # Initialize our Trainer trainer = QuestionAnsweringTrainer( 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, eval_examples=eval_examples if training_args.do_eval else None, tokenizer=tokenizer, data_collator=data_collator, post_process_function=post_processing_function, compute_metrics=compute_metrics, ) # 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) 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 if training_args.do_eval: logger.info("*** Evaluate ***") metrics = trainer.evaluate() 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) # Prediction if training_args.do_predict: logger.info("*** Predict ***") results = trainer.predict(predict_dataset, predict_examples) metrics = 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 training_args.push_to_hub: kwargs = { "finetuned_from": model_args.model_name_or_path, "tasks": "question-answering" } if data_args.dataset_name is not None: kwargs["dataset_tags"] = data_args.dataset_name if data_args.dataset_config_name is not None: kwargs["dataset_args"] = data_args.dataset_config_name kwargs[ "dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}" else: kwargs["dataset"] = data_args.dataset_name trainer.push_to_hub(**kwargs)
def main(): args = parse_args() # Initialize the accelerator. We will let the accelerator handle device placement for us in this example. accelerator = Accelerator() # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state) # Setup logging, we only want one process per machine to log things on the screen. # accelerator.is_local_main_process is only True for one process per machine. logger.setLevel( logging.INFO if accelerator.is_local_main_process else logging.ERROR) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(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 CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named # label if at least two columns are provided. # 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 args.task_name is not None: # Downloading and loading a dataset from the hub. raw_datasets = load_dataset("glue", args.task_name) else: # Loading the dataset from local csv or json file. data_files = {} if args.train_file is not None: data_files["train"] = args.train_file if args.validation_file is not None: data_files["validation"] = args.validation_file extension = (args.train_file if args.train_file is not None else args.valid_file).split(".")[-1] raw_datasets = load_dataset(extension, data_files=data_files) # See more about loading any type of standard or custom dataset at # https://huggingface.co/docs/datasets/loading_datasets.html. # Labels if args.task_name is not None: is_regression = args.task_name == "stsb" if not is_regression: label_list = raw_datasets["train"].features["label"].names num_labels = len(label_list) else: num_labels = 1 else: # Trying to have good defaults here, don't hesitate to tweak to your needs. is_regression = raw_datasets["train"].features["label"].dtype in [ "float32", "float64" ] if is_regression: num_labels = 1 else: # A useful fast method: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique label_list = raw_datasets["train"].unique("label") label_list.sort() # Let's sort it for determinism 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(args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name) tokenizer = AutoTokenizer.from_pretrained( args.model_name_or_path, use_fast=not args.use_slow_tokenizer) model = AutoModelForSequenceClassification.from_pretrained( args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config, ) # Preprocessing the datasets if args.task_name is not None: sentence1_key, sentence2_key = task_to_keys[args.task_name] else: # Again, we try to have some nice defaults but don't hesitate to tweak to your use case. non_label_column_names = [ name for name in raw_datasets["train"].column_names if name != "label" ] if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names: sentence1_key, sentence2_key = "sentence1", "sentence2" else: if len(non_label_column_names) >= 2: sentence1_key, sentence2_key = non_label_column_names[:2] else: sentence1_key, sentence2_key = non_label_column_names[0], None # Some models have set the order of the labels to use, so let's make sure we do use it. label_to_id = None if (model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id and args.task_name is not None and not is_regression): # Some have all caps in their config, some don't. label_name_to_id = { k.lower(): v for k, v in model.config.label2id.items() } if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)): logger.info( f"The configuration of the model provided the following label correspondence: {label_name_to_id}. " "Using it!") label_to_id = { i: label_name_to_id[label_list[i]] for i in range(num_labels) } else: logger.warning( "Your model seems to have been trained with labels, but they don't match the dataset: ", f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}." "\nIgnoring the model labels as a result.", ) elif args.task_name is None: label_to_id = {v: i for i, v in enumerate(label_list)} padding = "max_length" if args.pad_to_max_length else False def preprocess_function(examples): # Tokenize the texts texts = ((examples[sentence1_key], ) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])) result = tokenizer(*texts, padding=padding, max_length=args.max_length, truncation=True) if "label" in examples: if label_to_id is not None: # Map labels to IDs (not necessary for GLUE tasks) result["labels"] = [label_to_id[l] for l in examples["label"]] else: # In all cases, rename the column to labels because the model will expect that. result["labels"] = examples["label"] return result processed_datasets = raw_datasets.map( preprocess_function, batched=True, remove_columns=raw_datasets["train"].column_names) train_dataset = processed_datasets["train"] eval_dataset = processed_datasets["validation_matched" if args.task_name == "mnli" else "validation"] # Log a few random samples from the training set: for index in random.sample(range(len(train_dataset)), 3): logger.info( f"Sample {index} of the training set: {train_dataset[index]}.") # DataLoaders creation: if args.pad_to_max_length: # If padding was already done ot max length, we use the default data collator that will just convert everything # to tensors. data_collator = default_data_collator else: # Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of # the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple # of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta). data_collator = DataCollatorWithPadding( tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)) train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size) eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) # Optimizer # Split weights in two groups, one with weight decay and the other not. no_decay = ["bias", "LayerNorm.weight"] optimizer_grouped_parameters = [ { "params": [ p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) ], "weight_decay": args.weight_decay, }, { "params": [ p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) ], "weight_decay": 0.0, }, ] optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate) # Prepare everything with our `accelerator`. model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare( model, optimizer, train_dataloader, eval_dataloader) # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be # shorter in multiprocess) # Scheduler and math around the number of training steps. num_update_steps_per_epoch = math.ceil( len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch else: args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) lr_scheduler = get_scheduler( name=args.lr_scheduler_type, optimizer=optimizer, num_warmup_steps=args.num_warmup_steps, num_training_steps=args.max_train_steps, ) # Get the metric function if args.task_name is not None: metric = load_metric("glue", args.task_name) # Train! total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info( f" Instantaneous batch size per device = {args.per_device_train_batch_size}" ) logger.info( f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}" ) logger.info( f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") # Only show the progress bar once on each machine. progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process) completed_steps = 0 for epoch in range(args.num_train_epochs): model.train() for step, batch in enumerate(train_dataloader): outputs = model(**batch) loss = outputs.loss loss = loss / args.gradient_accumulation_steps accelerator.backward(loss) if step % args.gradient_accumulation_steps == 0 or step == len( train_dataloader) - 1: optimizer.step() lr_scheduler.step() optimizer.zero_grad() progress_bar.update(1) completed_steps += 1 if completed_steps >= args.max_train_steps: break model.eval() for step, batch in enumerate(eval_dataloader): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) metric.add_batch( predictions=accelerator.gather(predictions), references=accelerator.gather(batch["labels"]), ) eval_metric = metric.compute() logger.info(f"epoch {epoch}: {eval_metric}") if args.output_dir is not None: accelerator.wait_for_everyone() unwrapped_model = accelerator.unwrap_model(model) unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save) if args.task_name == "mnli": # Final evaluation on mismatched validation set eval_dataset = processed_datasets["validation_mismatched"] eval_dataloader = DataLoader( eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size) eval_dataloader = accelerator.prepare(eval_dataloader) model.eval() for step, batch in enumerate(eval_dataloader): outputs = model(**batch) predictions = outputs.logits.argmax(dim=-1) metric.add_batch( predictions=accelerator.gather(predictions), references=accelerator.gather(batch["labels"]), ) eval_metric = metric.compute() logger.info(f"mnli-mm: {eval_metric}")
def test_hf(): """tests example pipeline""" datasets = load_dataset("ag_news") tokenizer = AutoTokenizer.from_pretrained( "google/bert_uncased_L-4_H-128_A-2") datasets = datasets.map( lambda e: tokenizer( e["text"], truncation=True, padding="max_length", max_length=128), batched=True, ) datasets = datasets.map(lambda e: {"labels": e["label"]}, batched=True) datasets.set_format( type="torch", columns=["input_ids", "token_type_ids", "attention_mask", "labels"], ) loaders = { "train": DataLoader(datasets["train"], batch_size=32, shuffle=True), "valid": DataLoader(datasets["test"], batch_size=32), } metric_callback = LoaderMetricCallback( metric=HFMetric(metric=load_metric("accuracy")), input_key="logits", target_key="labels", ) teacher_model = AutoModelForSequenceClassification.from_pretrained( "google/bert_uncased_L-4_H-128_A-2", num_labels=4) runner = HFRunner() runner.train( model=teacher_model, loaders=loaders, optimizer=torch.optim.Adam(teacher_model.parameters(), lr=1e-4), callbacks=[metric_callback], num_epochs=3, valid_metric="accuracy", minimize_valid_metric=False, check=True, ) metric_callback = LoaderMetricCallback( metric=HFMetric(metric=load_metric("accuracy")), input_key="s_logits", target_key="labels", ) slct_callback = ControlFlowCallback( HiddenStatesSelectCallback(hiddens_key="t_hidden_states", layers=[1, 3]), loaders="train", ) lambda_hiddens_callback = ControlFlowCallback( LambdaPreprocessCallback(lambda s_hiddens, t_hiddens: ( [c_s[:, 0] for c_s in s_hiddens], [t_s[:, 0] for t_s in t_hiddens], # tooks only CLS token )), loaders="train", ) mse_hiddens = ControlFlowCallback(MSEHiddenStatesCallback(), loaders="train") kl_div = ControlFlowCallback(KLDivCallback(), loaders="train") aggregator = ControlFlowCallback( MetricAggregationCallback( prefix="loss", metrics={ "kl_div_loss": 0.2, "mse_loss": 0.2, "task_loss": 0.6 }, mode="weighted_sum", ), loaders="train", ) runner = HFDistilRunner() student_model = AutoModelForSequenceClassification.from_pretrained( "google/bert_uncased_L-2_H-128_A-2", num_labels=4) runner.train( model=torch.nn.ModuleDict({ "teacher": teacher_model, "student": student_model }), loaders=loaders, optimizer=torch.optim.Adam(student_model.parameters(), lr=1e-4), callbacks=[ metric_callback, slct_callback, lambda_hiddens_callback, mse_hiddens, kl_div, aggregator, OptimizerCallback(metric_key="loss"), ], check=True, num_epochs=3, valid_metric="accuracy", minimize_valid_metric=False, valid_loader="valid", )