run_ner_no_trainer_.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning a 🤗 Transformers model on token classification tasks (NER, POS, CHUNKS) relying on the accelerate library
without using a Trainer.
"""

import argparse
import logging
import math
import os
import random
import numpy as np


import datasets
import torch
from datasets import ClassLabel, load_dataset, load_metric
from torch.utils.data.dataloader import DataLoader
from tqdm.auto import tqdm

import transformers
from accelerate import Accelerator
from transformers import (
    CONFIG_MAPPING,
    MODEL_MAPPING,
    AdamW,
    AutoConfig,
    AutoModelForTokenClassification,
    AutoTokenizer,
    DataCollatorForTokenClassification,
    SchedulerType,
    default_data_collator,
    get_scheduler,
    set_seed,
    BertForTokenClassification
)
from transformers.utils.versions import require_version
import time
# from classifier import MyBertForTokenClassification_clusterloss_test
from classifier_fine import MyBertForTokenClassification_prototype
#from classifier_ori import OriBertForTokenClassification

logger = logging.getLogger(__name__)
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/token-classification/requirements.txt")

# You should update this to your particular problem to have better documentation of `model_type`
MODEL_CONFIG_CLASSES = list(MODEL_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)


def parse_args():
    parser = argparse.ArgumentParser(
        description="Finetune a transformers model on a text classification task (NER) with accelerate library"
    )
    parser.add_argument(
        "--dataset_name",
        type=str,
        default=None,
        help="The name of the dataset to use (via the datasets library).",
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The configuration name of the dataset to use (via the datasets library).",
    )
    parser.add_argument(
        "--sample_path",
        type=str,
        default=None,
        help="The configuration name of the dataset to use (via the datasets library).",
    )
    parser.add_argument(
        "--train_file", type=str, default=None, help="A csv or a json file containing the training data."
    )
    parser.add_argument(
        "--validation_file", type=str, default=None, help="A csv or a json file containing the validation data."
    )

    parser.add_argument(
        "--text_column_name",
        type=str,
        default=None,
        help="The column name of text to input in the file (a csv or JSON file).",
    )
    parser.add_argument(
        "--label_column_name",
        type=str,
        default=None,
        help="The column name of label to input in the file (a csv or JSON file).",
    )
    parser.add_argument(
        "--max_length",
        type=int,
        default=128,
        help=(
            "The maximum total input sequence length after tokenization. Sequences longer than this will be truncated,"
            " sequences shorter will be padded if `--pad_to_max_lenght` is passed."
        ),
    )
    parser.add_argument(
        "--pad_to_max_length",
        action="store_true",
        help="If passed, pad all samples to `max_length`. Otherwise, dynamic padding is used.",
    )
    parser.add_argument(
        "--model_name_or_path",
        type=str,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
        required=True,
    )
    parser.add_argument(
        "--config_name",
        type=str,
        default=None,
        help="Pretrained config name or path if not the same as model_name",
    )
    parser.add_argument(
        "--tokenizer_name",
        type=str,
        default=None,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--per_device_train_batch_size",
        type=int,
        default=4,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument(
        "--per_device_eval_batch_size",
        type=int,
        default=128,
        help="Batch size (per device) for the evaluation dataloader.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.")
    parser.add_argument("--num_train_epochs", type=int, default=3, help="Total number of training epochs to perform.")
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--lr_scheduler_type",
        type=SchedulerType,
        default="linear",
        help="The scheduler type to use.",
        choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
    )
    parser.add_argument(
        "--num_warmup_steps", type=int, default=0, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.")
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--model_type",
        type=str,
        default=None,
        help="Model type to use if training from scratch.",
        choices=MODEL_TYPES,
    )
    parser.add_argument(
        "--label_all_tokens",
        action="store_true",
        help="Setting labels of all special tokens to -100 and thus PyTorch will ignore them.",
    )
    parser.add_argument(
        "--return_entity_level_metrics",
        action="store_true",
        help="Indication whether entity level metrics are to be returner.",
    )
    parser.add_argument(
        "--task_name",
        type=str,
        default="ner",
        choices=["ner", "pos", "chunk"],
        help="The name of the task.",
    )
    parser.add_argument(
        "--save_mode",
        type=str,
        default="best",
        choices=["best", "last"],
        help="Save the best or the final model.",
    )
    parser.add_argument(
        "--debug",
        action="store_true",
        help="Activate debug mode and run training only with a subset of data.",
    )
    parser.add_argument(
        "--eval_last_epoch",
        action="store_true",
        help="Only evaluate the last epoch.",
    )
    parser.add_argument(
        "--label_schema",
        type=str,
        default="BIO",
    )
    parser.add_argument(
        "--label_list", type=str, default=None, help="Path of label list."
    )
    parser.add_argument(
        "--probe",
        action="store_true",
        help="Whether probing.",
    )
    parser.add_argument(
        "--only_eval",
        action="store_true",
        help="Whether only evaluate.",
    )
    parser.add_argument(
        "--use_ori_cls",
        action="store_true",
        help="Whether use the vallina classifier.",
    )
    parser.add_argument(
        "--no_save",
        action="store_true",
        help="Whether save model.",
    )
    parser.add_argument(
        "--sample_id", type=str, default=None, help="Sample ids."
    )
    parser.add_argument(
        "--use_aug",
        action="store_true",
        help="Whether use augmented data.",
    )

    args = parser.parse_args()

    args.save = not args.no_save
    if args.probe:
        args.save = False

    # Sanity checks
    if args.task_name is None and args.train_file is None and args.validation_file is None:
        raise ValueError("Need either a task name or a training/validation file.")
    else:
        if args.train_file is not None:
            extension = args.train_file.split(".")[-1]
            assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
        if args.validation_file is not None:
            extension = args.validation_file.split(".")[-1]
            assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."

    if args.output_dir is not None:
        os.makedirs(args.output_dir, exist_ok=True)

    return args


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)
    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

    if args.text_column_name is not None:
        text_column_name = args.text_column_name
    elif "tokens" in column_names:
        text_column_name = "tokens"
    else:
        text_column_name = column_names[0]

    if args.label_column_name is not None:
        label_column_name = args.label_column_name
    elif f"{args.task_name}_tags" in column_names:
        label_column_name = f"{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()

        ## conll
        # label_list = ["B-PER", "I-PER", "B-LOC", "I-LOC", "B-ORG", "I-ORG", "B-MISC", "I-MISC", "O"]
        # label_list.sort()

        ## ontonotes dNUM
        # IO
        if args.label_schema == "IO":
            with open(args.label_list, "r") as f:
                label_list = [l.strip() for l in f.readlines() if not l.strip().startswith("B-")]
        else:
            with open(args.label_list, "r") as f:
                label_list = [l.strip() for l in f.readlines()]


        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)}

    # IO
    num_labels = len(label_list)

    # if BIO TODO
    if args.label_schema == "IO":
        import copy
        ori_label_to_id = copy.deepcopy(label_to_id)
        ori_label_list = copy.deepcopy(label_list)
        for label, id in label_to_id.items():
            if label != "O" and "B-" + label[2:] not in label_to_id:
                ori_label_to_id["B-" + label[2:]] = len(ori_label_to_id)
                ori_label_list.append("B-" + label[2:])
    else:
        ori_label_list = label_list
        ori_label_to_id = label_to_id
    # 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.sample_id is not None:
        config.sample_id = args.sample_id
    if args.sample_path is not None:
        config.sample_path=args.sample_path
    tokenizer_name_or_path = args.tokenizer_name if args.tokenizer_name else args.model_name_or_path
    if not tokenizer_name_or_path:
        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 config.model_type in {"gpt2", "roberta"}:
        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, use_fast=True, add_prefix_space=True)
    else:
        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, use_fast=True, do_lower_case=False)

    if args.only_eval and args.output_dir:
        args.model_name_or_path = args.output_dir

    if args.model_name_or_path:

        if args.use_ori_cls:
            model = BertForTokenClassification.from_pretrained(
                args.model_name_or_path,
                from_tf=bool(".ckpt" in args.model_name_or_path),
                config=config,
                ignore_mismatched_sizes=True
            )
        else:
            config.model_name = args.model_name_or_path.split("/")[-1]
            config.use_aug = args.use_aug
            model = MyBertForTokenClassification_prototype.from_pretrained(
                args.model_name_or_path,
                from_tf=bool(".ckpt" in args.model_name_or_path),
                config=config,
                ignore_mismatched_sizes=True
            )

    else:
        logger.info("Training new model from scratch")
        model = AutoModelForTokenClassification.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

    # 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 = []
        ori_labels = []
        # gold_labels = examples[label_column_name]
        # if args.label_schema=='IO':
        #     gold_labels = [['I-{}'.format(l[2:]) if l !='O' else 'O' for l in label] for label in gold_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 = []
            ori_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.
                # ori_label = label[word_idx]
                # if args.label_schema=='IO':
                #     label[word_idx] = "I-"+label[word_idx][2:] if label[word_idx] and label[word_idx] != "O" else label[word_idx]
                if word_idx is None:
                    label_ids.append(-100)
                    ori_label_ids.append(-100)
                # We set the label for the first token of each word.
                elif word_idx != previous_word_idx:
                    ori_label_ids.append(ori_label_to_id[label[word_idx]])

                    if args.label_schema == "IO":
                        label_ids.append(
                            label_to_id["I-" + label[word_idx][2:] if label[word_idx] != "O" else label[word_idx]])
                    else:
                        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:
                    ori_label_ids.append(ori_label_to_id[label[word_idx]] if args.label_all_tokens else -100)
                    if args.label_schema == "IO":
                        label_ids.append(label_to_id["I-" + label[word_idx][2:] if label[word_idx] != "O" else label[
                            word_idx]] if args.label_all_tokens else -100)
                    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)
            ori_labels.append(ori_label_ids)
        tokenized_inputs["labels"] = labels
        tokenized_inputs["ori_labels"] = ori_labels
        return tokenized_inputs

    processed_raw_datasets = raw_datasets.map(
        tokenize_and_align_labels,
        batched=True,
        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]}.")

    # 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 = DataCollatorForLMTokanClassification(
            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"]
    if args.probe:
        print("probing!")
        retain_key = "classifier"  # classifier
    else:
        retain_key = ""
    a = [n for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) and retain_key in n]
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in model.named_parameters() if
                       not any(nd in n for nd in no_decay) and retain_key in n],
            "weight_decay": args.weight_decay,
        },
        {
            "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) and retain_key in n],
            "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
    # print('loading metric.....')
    metric = load_metric("seqeval_metric.py")
    # print('loading metric.....done!')

    def switch_to_BIO(labels):
        past_label = 'O'
        labels_BIO = []
        for label in labels:
            if label.startswith('I-') and (past_label == 'O' or past_label[2:] != label[2:]):
                labels_BIO.append('B-' + label[2:])
            else:
                labels_BIO.append(label)
            past_label = label
        return labels_BIO

    def get_labels(predictions, references, tokens):
        # Transform predictions and references tensos to numpy arrays
        if device.type == "cpu":
            y_pred = predictions.detach().clone().numpy()
            y_true = references.detach().clone().numpy()
            x_tokens = tokens.detach().clone().numpy()
        else:
            y_pred = predictions.detach().cpu().clone().numpy()
            y_true = references.detach().cpu().clone().numpy()
            x_tokens = tokens.detach().cpu().clone().tolist()

        # Remove ignored index (special tokens)
        true_predictions = [
            [ori_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 = [
            [ori_label_list[l] for (p, l) in zip(pred, gold_label) if l != -100]
            for pred, gold_label in zip(y_pred, y_true)
        ]

        if args.label_schema == "IO":
            true_predictions = list(map(switch_to_BIO, true_predictions))
            true_labels = list(map(switch_to_BIO, true_labels))

        ori_tokens = [
            [tokenizer.convert_ids_to_tokens(t) for (p, l, t) in zip(pred, gold_label, token) if l != -100]
            for pred, gold_label, token in zip(y_pred, y_true, x_tokens)
        ]

        return true_predictions, true_labels, ori_tokens


    def compute_metrics():
        results = metric.compute()
        if args.return_entity_level_metrics:
            # Unpack nested dictionaries
            final_results = {}
            macro_results = {"macro_f1": 0., "macro_precision": 0., "macro_recall": 0., "class_num": 0}
            for key, value in results.items():
                if isinstance(value, dict):
                    for n, v in value.items():
                        final_results[f"{key}_{n}"] = v
                    macro_results["macro_f1"] += value["f1"]
                    macro_results["macro_precision"] += value["precision"]
                    macro_results["macro_recall"] += value["recall"]
                    macro_results["class_num"] += 1
                else:
                    final_results[key] = value

            for key, value in macro_results.items():
                if key != "class_num":
                    final_results[key] = macro_results[key] / macro_results["class_num"]
            return final_results
        else:
            return {
                "precision": results["overall_precision"],
                "recall": results["overall_recall"],
                "f1": results["overall_f1"],
                "accuracy": results["overall_accuracy"],
            }

    def evaluate(best_metric=-1, save=True):
        model.eval()
        start = time.time()
        token_list = []
        y_true = []
        y_pred = []
        hiddens = None
        all_labels = None
        pro_bar = tqdm(range(len(eval_dataloader)))
        for step, batch in enumerate(eval_dataloader):
            labels = batch.pop("ori_labels")
            _ = batch.pop("labels")
            with torch.no_grad():
                outputs = model(**batch)
            predictions = outputs.logits.argmax(dim=-1)

            token_labels = batch.pop("input_ids")
            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)
                token_labels = accelerator.pad_across_processes(token_labels, dim=1, pad_index=-100)

            predictions_gathered = accelerator.gather(predictions)
            labels_gathered = accelerator.gather(labels)
            token_labels_gathered = accelerator.gather(token_labels)
            preds, refs, tokens = get_labels(predictions_gathered, labels_gathered, token_labels_gathered)
            token_list.extend(tokens)
            y_true.extend(refs)
            y_pred.extend(preds)
            metric.add_batch(
                predictions=preds,
                references=refs,
            )  # predictions and preferences are expected to be a nested list of labels, not label_ids

            pro_bar.update(1)

        # eval_metric = metric.compute()
        eval_metric = compute_metrics()
        # accelerator.print(f"epoch {epoch}:", eval_metric)
        print("Decoding time: {}".format(time.time() - start))
        for key in eval_metric.keys():
            if "f1" in key and "overall" not in key and "macro" not in key:
                label = key[:-3]
                print("{}: {}, {}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                                              label + "_recall", eval_metric[label + "_recall"],
                                                              label + "_f1", eval_metric[label + "_f1"],
                                                              label + "_number", eval_metric[label + "_number"]))
        label = "overall"
        print("{}: {}, {}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                                      label + "_recall", eval_metric[label + "_recall"],
                                                      label + "_f1", eval_metric[label + "_f1"],
                                                      label + "_accuracy", eval_metric[label + "_accuracy"]))
        label = "macro"
        print("{}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                                      label + "_recall", eval_metric[label + "_recall"],
                                                      label + "_f1", eval_metric[label + "_f1"]))

        if best_metric == -1 or best_metric["overall_f1"] < eval_metric["overall_f1"]:
            best_metric = eval_metric

            with open(os.path.join(args.output_dir, "predictions.txt"), "w") as f:
                for i in range(len(y_true)):
                    for j in range(len(y_true[i])):
                        f.write(f"{token_list[i][j]} {y_true[i][j]} {y_pred[i][j]}\n")
                    f.write("\n")

            if args.save_mode == "best" and save:
                print("Saving model to {}".format(args.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, save_function=accelerator.save)
                    tokenizer.save_pretrained(args.output_dir)

            

        return best_metric


    def evaluate_with_dump(best_metric=-1, save=True):
        model.eval()
        start = time.time()
        token_list = []
        y_true = []
        y_pred = []
        hiddens = None
        all_labels = None
        for step, batch in enumerate(eval_dataloader):
            labels = batch.pop("ori_labels")
            batch["output_hidden_states"] = True
            with torch.no_grad():
                outputs = model(**batch)
            predictions = outputs.logits.argmax(dim=-1)

            hidden_outputs = outputs.hidden_states[-1].detach().cpu()
            all_label = labels.detach().cpu()
            if hiddens is not None:
                hiddens = torch.cat([hiddens, hidden_outputs.view(-1, hidden_outputs.size(-1))], dim=0)
                all_labels = torch.cat([all_labels, all_label.view(-1)], dim=0)
            else:
                hiddens = hidden_outputs.view(-1, hidden_outputs.size(-1))
                all_labels = all_label.view(-1)

            labels = batch["labels"]
            token_labels = batch.pop("input_ids")
            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)
                token_labels = accelerator.pad_across_processes(token_labels, dim=1, pad_index=-100)

            predictions_gathered = accelerator.gather(predictions)
            labels_gathered = accelerator.gather(labels)
            token_labels_gathered = accelerator.gather(token_labels)
            preds, refs, tokens = get_labels(predictions_gathered, labels_gathered, token_labels_gathered)
            token_list.extend(tokens)
            y_true.extend(refs)
            y_pred.extend(preds)
            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)
        print("Decoding time: {}".format(time.time() - start))
        for key in eval_metric.keys():
            if "f1" in key and "overall" not in key and "macro" not in key:
                label = key[:-3]
                print("{}: {}, {}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                                              label + "_recall", eval_metric[label + "_recall"],
                                                              label + "_f1", eval_metric[label + "_f1"],
                                                              label + "_number", eval_metric[label + "_number"]))
        label = "overall"
        print("{}: {}, {}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                                      label + "_recall", eval_metric[label + "_recall"],
                                                      label + "_f1", eval_metric[label + "_f1"],
                                                      label + "_accuracy", eval_metric[label + "_accuracy"]))
        label = "macro"
        print("{}: {}, {}: {}, {}: {}".format(label + "_precision", eval_metric[label + "_precision"],
                                              label + "_recall", eval_metric[label + "_recall"],
                                              label + "_f1", eval_metric[label + "_f1"]))

        if best_metric == -1 or best_metric["overall_f1"] < eval_metric["overall_f1"]:
            best_metric = eval_metric

            with open(os.path.join(args.output_dir, "predictions.txt"), "w") as f:
                for i in range(len(y_true)):
                    for j in range(len(y_true[i])):
                        f.write(f"{token_list[i][j]} {y_true[i][j]} {y_pred[i][j]}\n")
                    f.write("\n")

            if args.save_mode == "best" and save:
                print("Saving model to {}".format(args.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, save_function=accelerator.save)
                    tokenizer.save_pretrained(args.output_dir)

            hiddens = hiddens.numpy()
            y_true = all_labels.numpy()

            valid_hiddens = []
            valid_labels = []
            for l, h in zip(y_true, hiddens):
                if l != -100:
                    valid_labels.append(ori_label_list[l])
                    valid_hiddens.append(h)

            valid_labels = np.array(valid_labels)
            valid_hiddens = np.array(valid_hiddens)

            np.save(os.path.join(args.output_dir, "embeddings.npy"), valid_hiddens)
            np.save(os.path.join(args.output_dir, "labels.npy"), valid_labels)


        return best_metric


    if args.only_eval:
        # evaluate_with_dump(save=False)
        evaluate(save=False)
    else:
        # 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}")
        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

        best_metric = -1
        for epoch in range(args.num_train_epochs):
            if not args.use_ori_cls and epoch % 10 == 0:
                model.update_cluster_label()
            model.train()
            for step, batch in enumerate(train_dataloader):
                ori_labels = batch.pop("ori_labels")
                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


            # set when to eval
            if args.eval_last_epoch:
                if epoch == args.num_train_epochs - 1:
                    best_metric = evaluate(best_metric, args.save)
            elif epoch >= 0:
                best_metric = evaluate(best_metric, args.save)

        print("Finish training, best metric: ")
        print(best_metric)
        with open(os.path.join(args.output_dir, "results.txt"), "w") as f:
            f.write(f"{best_metric}")
        if args.save_mode == "last":
            print("Saving final model to {}".format(args.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, save_function=accelerator.save)
                tokenizer.save_pretrained(args.output_dir)


class DataCollatorForLMTokanClassification(DataCollatorForTokenClassification):
    def __call__(self, features):
        label_name = "label" if "label" in features[0].keys() else "labels"
        labels = [feature[label_name] for feature in features] if label_name in features[0].keys() else None
        ori_labels = [feature['ori_labels'] for feature in features] if 'ori_labels' in features[0].keys() else None
        batch = self.tokenizer.pad(
            features,
            padding=self.padding,
            max_length=self.max_length,
            pad_to_multiple_of=self.pad_to_multiple_of,
            # Conversion to tensors will fail if we have labels as they are not of the same length yet.
            return_tensors="pt" if labels is None else None,
        )

        if labels is None:
            return batch

        sequence_length = torch.tensor(batch["input_ids"]).shape[1]
        padding_side = self.tokenizer.padding_side
        if padding_side == "right":
            batch["labels"] = [label + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels]
            batch['ori_labels'] = [label + [self.label_pad_token_id] * (sequence_length - len(label)) for label in
                                   ori_labels]
        else:
            batch["labels"] = [[self.label_pad_token_id] * (sequence_length - len(label)) + label for label in labels]
            batch["ori_labels"] = [[self.label_pad_token_id] * (sequence_length - len(label)) + label for label in
                                   ori_labels]
        batch = {k: torch.tensor(v, dtype=torch.int64) for k, v in batch.items()}
        return batch


if __name__ == "__main__":
    main()