a less bad version of the hf run_classification script

run_classification.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2020 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.
""" Finetuning the library models for text classification."""

import logging
import os
from pathlib import Path
import random
import sys
import warnings
from dataclasses import dataclass, field
from typing import List, Optional

import datasets
import evaluate
import numpy as np
import torch
import transformers
from datasets import Sequence, load_dataset
from scipy.special import (
    expit as sigmoid,
)  # More stable than writing sigmoid from scratch
from sklearn.metrics import (
    hamming_loss,
    jaccard_score,
    precision_recall_fscore_support,
)
from transformers import (
    AutoConfig,
    AutoModelForSequenceClassification,
    AutoTokenizer,
    DataCollatorWithPadding,
    EvalPrediction,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    default_data_collator,
    set_seed,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils.versions import require_version

require_version(
    "datasets>=1.8.0",
    "To fix: pip install -r examples/pytorch/text-classification/requirements.txt",
)


logger = logging.getLogger(__name__)


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.

    Using `HfArgumentParser` we can turn this class
    into argparse arguments to be able to specify them on
    the command line.
    """

    dataset_name: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the dataset to use (via the datasets library)."},
    )
    dataset_config_name: Optional[str] = field(
        default=None,
        metadata={
            "help": "The configuration name of the dataset to use (via the datasets library)."
        },
    )
    do_regression: bool = field(
        default=None,
        metadata={
            "help": "Whether to do regression instead of classification. If None, will be inferred from the dataset."
        },
    )
    text_column_names: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The name of the text column in the input dataset or a CSV/JSON file. "
                'If not specified, will use the "sentence" column for single/multi-label classifcation task.'
            )
        },
    )
    text_column_delimiter: Optional[str] = field(
        default=" ",
        metadata={
            "help": "THe delimiter to use to join text columns into a single sentence."
        },
    )
    train_split_name: Optional[str] = field(
        default=None,
        metadata={
            "help": 'The name of the train split in the input dataset. If not specified, will use the "train" split when do_train is enabled'
        },
    )
    validation_split_name: Optional[str] = field(
        default=None,
        metadata={
            "help": 'The name of the validation split in the input dataset. If not specified, will use the "validation" split when do_eval is enabled'
        },
    )
    test_split_name: Optional[str] = field(
        default=None,
        metadata={
            "help": 'The name of the test split in the input dataset. If not specified, will use the "test" split when do_predict is enabled'
        },
    )
    remove_splits: Optional[str] = field(
        default=None,
        metadata={
            "help": "The splits to remove from the dataset. Multiple splits should be separated by commas."
        },
    )
    remove_columns: Optional[str] = field(
        default=None,
        metadata={
            "help": "The columns to remove from the dataset. Multiple columns should be separated by commas."
        },
    )
    label_column_name: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The name of the label column in the input dataset or a CSV/JSON file. "
                'If not specified, will use the "label" column for single/multi-label classifcation task'
            )
        },
    )
    max_seq_length: int = field(
        default=128,
        metadata={
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    overwrite_cache: bool = field(
        default=False,
        metadata={"help": "Overwrite the cached preprocessed datasets or not."},
    )
    pad_to_max_length: bool = field(
        default=True,
        metadata={
            "help": (
                "Whether to pad all samples to `max_seq_length`. "
                "If False, will pad the samples dynamically when batching to the maximum length in the batch."
            )
        },
    )
    shuffle_train_dataset: bool = field(
        default=False, metadata={"help": "Whether to shuffle the train dataset or not."}
    )
    shuffle_seed: int = field(
        default=42,
        metadata={
            "help": "Random seed that will be used to shuffle the train dataset."
        },
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
        },
    )
    max_predict_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                "value if set."
            )
        },
    )
    metric_name: Optional[str] = field(
        default=None, metadata={"help": "The metric to use for evaluation."}
    )
    train_file: Optional[str] = field(
        default=None,
        metadata={"help": "A csv or a json file containing the training data."},
    )
    validation_file: Optional[str] = field(
        default=None,
        metadata={"help": "A csv or a json file containing the validation data."},
    )
    test_file: Optional[str] = field(
        default=None,
        metadata={"help": "A csv or a json file containing the test data."},
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    print_samples: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Whether to print a sample of the dataset before training or not."
        },
    )

    def __post_init__(self):
        if self.dataset_name is None:
            if self.train_file is None or self.validation_file is None:
                raise ValueError(" training/validation file or a dataset name.")

            train_extension = self.train_file.split(".")[-1]
            assert train_extension in [
                "csv",
                "json",
            ], "`train_file` should be a csv or a json file."
            validation_extension = self.validation_file.split(".")[-1]
            assert (
                validation_extension == train_extension
            ), "`validation_file` should have the same extension (csv or json) as `train_file`."


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={
            "help": "Path to pretrained model or model identifier from huggingface.co/models"
        }
    )
    config_name: Optional[str] = field(
        default=None,
        metadata={
            "help": "Pretrained config name or path if not the same as model_name"
        },
    )
    tokenizer_name: Optional[str] = field(
        default=None,
        metadata={
            "help": "Pretrained tokenizer name or path if not the same as model_name"
        },
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={
            "help": "Where do you want to store the pretrained models downloaded from huggingface.co"
        },
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={
            "help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."
        },
    )
    model_revision: str = field(
        default="main",
        metadata={
            "help": "The specific model version to use (can be a branch name, tag name or commit id)."
        },
    )
    token: str = field(
        default=None,
        metadata={
            "help": (
                "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
                "generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
            )
        },
    )
    use_auth_token: bool = field(
        default=None,
        metadata={
            "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead."
        },
    )
    trust_remote_code: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option"
                "should only be set to `True` for repositories you trust and in which you have read the code, as it will "
                "execute code present on the Hub on your local machine."
            )
        },
    )
    ignore_mismatched_sizes: bool = field(
        default=False,
        metadata={
            "help": "Will enable to load a pretrained model whose head dimensions are different."
        },
    )


def get_label_list(raw_dataset, split="train") -> List[str]:
    """Get the list of labels from a mutli-label dataset"""

    if isinstance(raw_dataset[split]["label"][0], list):
        label_list = [
            label for sample in raw_dataset[split]["label"] for label in sample
        ]
        label_list = list(set(label_list))
    else:
        label_list = raw_dataset[split].unique("label")
    # we will treat the label list as a list of string instead of int, consistent with model.config.label2id
    label_list = [str(label) for label in label_list]
    return label_list


def main():
    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 model_args.use_auth_token is not None:
        warnings.warn(
            "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead.",
            FutureWarning,
        )
        if model_args.token is not None:
            raise ValueError(
                "`token` and `use_auth_token` are both specified. Please set only the argument `token`."
            )
        model_args.token = model_args.use_auth_token

    # 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)],
    )

    if training_args.should_log:
        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
        transformers.utils.logging.set_verbosity_info()

    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(logging.ERROR)
    transformers.utils.logging.set_verbosity(logging.ERROR)
    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: {training_args.parallel_mode.value == 'distributed'}, 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, or specify a dataset name
    # to load from huggingface/datasets. In ether case, you can specify a the key of the column(s) containing the text and
    # the key of the column containing the label. If multiple columns are specified for the text, they will be joined togather
    # for the actual text value.
    # 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,
            token=model_args.token,
        )
        # Try print some info about the dataset
        logger.info(f"Dataset loaded: {raw_datasets}")
        logger.info(raw_datasets)
    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
        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 dataset name 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,
                token=model_args.token,
            )
        else:
            # Loading a dataset from local json files
            raw_datasets = load_dataset(
                "json",
                data_files=data_files,
                cache_dir=model_args.cache_dir,
                token=model_args.token,
            )

    # See more about loading any type of standard or custom dataset at
    # https://huggingface.co/docs/datasets/loading_datasets.

    if data_args.remove_splits is not None:
        for split in data_args.remove_splits.split(","):
            logger.info(f"removing split {split}")
            raw_datasets.pop(split)

    if data_args.train_split_name is not None:
        logger.info(f"using {data_args.validation_split_name} as validation set")
        raw_datasets["train"] = raw_datasets[data_args.train_split_name]
        raw_datasets.pop(data_args.train_split_name)

    if data_args.validation_split_name is not None:
        logger.info(f"using {data_args.validation_split_name} as validation set")
        raw_datasets["validation"] = raw_datasets[data_args.validation_split_name]
        raw_datasets.pop(data_args.validation_split_name)

    if data_args.test_split_name is not None:
        logger.info(f"using {data_args.test_split_name} as test set")
        raw_datasets["test"] = raw_datasets[data_args.test_split_name]
        raw_datasets.pop(data_args.test_split_name)

    if data_args.remove_columns is not None:
        for split in raw_datasets.keys():
            for column in data_args.remove_columns.split(","):
                logger.info(f"removing column {column} from split {split}")
                raw_datasets[split].remove_columns(column)

    if (
        data_args.label_column_name is not None
        and data_args.label_column_name != "label"
    ):
        for key in raw_datasets.keys():
            raw_datasets[key] = raw_datasets[key].rename_column(
                data_args.label_column_name, "label"
            )

    # Determine the task type: Classification (single/multi-label) or Regression
    is_regression = data_args.do_regression or raw_datasets["train"].features[
        "label"
    ].dtype in ["float32", "float64"]
    is_multi_label = not is_regression and isinstance(
        raw_datasets["train"].features["label"], Sequence
    )

    def create_label_mappings(datasets, task_type, is_multi_label=False):
        """
        create_label_mappings - Helper function to create label mappings

        :param _type_ datasets: _description_
        :param _type_ task_type: _description_
        :param bool is_multi_label: _description_, defaults to False
        :return _type_: _description_
        """
        label_list = []  # Define label_list to ensure it's always available
        if task_type == "regression":
            return (
                {},
                {},
                label_list,
            )  # Return empty mappings and label_list for regression

        all_labels = set()
        for split in datasets.keys():
            if is_multi_label:
                # Flatten and collect unique labels for multi-label classification
                labels = [
                    label for sublist in datasets[split]["label"] for label in sublist
                ]
            else:
                # Collect unique labels for single-label classification
                labels = datasets[split]["label"]
            all_labels.update(labels)

        # Create sorted list of unique labels
        label_list = sorted(list(all_labels))
        # Generate mappings
        label_to_id = {label: idx for idx, label in enumerate(label_list)}
        id2label = {idx: label for idx, label in enumerate(label_list)}

        return label_to_id, id2label, label_list

    # Use the function to create mappings and label_list
    label_to_id, id2label, label_list = create_label_mappings(
        raw_datasets,
        "regression" if is_regression else "classification",
        is_multi_label,
    )

    num_labels = 1 if is_regression else len(label_to_id)

    # Adjust logging to reflect the task and mappings
    logger.info(
        f"Task type: {'Regression' if is_regression else 'Multi-label Classification' if is_multi_label else 'Single-label Classification'}"
    )
    if not is_regression:
        logger.info(f"Unique labels extracted: {len(label_to_id)}")

    # 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,
        id2label=id2label,
        label2id=label_to_id,
        finetuning_task="text-classification",
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
    )

    if is_regression:
        config.problem_type = "regression"
        logger.info("setting problem type to regression")
    elif is_multi_label:
        config.problem_type = "multi_label_classification"
        logger.info("setting problem type to multi label classification")
    else:
        config.problem_type = "single_label_classification"
        logger.info("setting problem type to single label classification")

    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,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
    )
    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,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
        ignore_mismatched_sizes=model_args.ignore_mismatched_sizes,
    )

    # DATA PROCESSING
    # --------------------------------------------------------------------------

    # 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

    # for training ,we will update the config with label infos,
    # if do_train is not set, we will use the label infos in the config
    if training_args.do_train and not is_regression:  # classification, training
        label_to_id = {v: i for i, v in enumerate(label_list)}
        # update config with label infos
        if model.config.label2id != label_to_id:
            logger.warning(
                "The label2id key in the model config.json is not equal to the label2id key of this "
                "run. You can ignore this if you are doing finetuning."
            )
        model.config.label2id = label_to_id
        model.config.id2label = {id: label for label, id in config.label2id.items()}
    elif not is_regression:  # classification, but not training
        logger.info("using label infos in the model config")
        logger.info("label2id: {}".format(model.config.label2id))
        label_to_id = model.config.label2id
    else:  # regression
        label_to_id = None

    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}). we are not doing shit about this."
        )  # do not assume the user is a dumbass
        tokenizer.model_max_length = data_args.max_seq_length

    max_seq_length = data_args.max_seq_length

    def multi_labels_to_ids(labels: List[str]) -> List[float]:
        ids = [0.0] * len(label_to_id)  # BCELoss requires float as target type
        for label in labels:
            ids[label_to_id[label]] = 1.0
        return ids

    def preprocess_function(examples):
        if data_args.text_column_names is not None:
            text_column_names = data_args.text_column_names.split(",")
            # join together text columns into "sentence" column
            examples["sentence"] = examples[text_column_names[0]]
            for column in text_column_names[1:]:
                for i in range(len(examples[column])):
                    examples["sentence"][i] += (
                        data_args.text_column_delimiter + examples[column][i]
                    )
        # Tokenize the texts
        result = tokenizer(
            examples["sentence"],
            padding=padding,
            max_length=max_seq_length,
            truncation=True,
        )
        if label_to_id is not None and "label" in examples:
            if is_multi_label:
                result["label"] = [multi_labels_to_ids(l) for l in examples["label"]]
            else:
                result["label"] = [
                    (label_to_id[str(l)] if l != -1 else -1) for l in examples["label"]
                ]
        return result

    # Running the preprocessing pipeline on all the datasets
    with training_args.main_process_first(desc="dataset map pre-processing"):
        raw_datasets = raw_datasets.map(
            preprocess_function,
            batched=True,
            batch_size=32 if os.cpu_count() > 16 else 64,
            num_proc=data_args.preprocessing_num_workers,
            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.shuffle_train_dataset:
            logger.info("Shuffling the training dataset")
            train_dataset = train_dataset.shuffle(seed=data_args.shuffle_seed)
        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))

    if training_args.do_eval:
        if (
            "validation" not in raw_datasets
            and "validation_matched" not in raw_datasets
        ):
            if "test" not in raw_datasets and "test_matched" not in raw_datasets:
                raise ValueError(
                    "--do_eval requires a validation or test dataset if validation is not defined."
                )
            else:
                logger.warning(
                    "Validation dataset not found. Falling back to test dataset for validation."
                )
                eval_dataset = raw_datasets["test"]
        else:
            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))

    if training_args.do_predict or data_args.test_file is not None:
        if "test" not in raw_datasets:
            raise ValueError("--do_predict requires a test dataset")
        predict_dataset = raw_datasets["test"]
        # remove label column if it exists
        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))

    # Log a few random samples from the training set:
    if training_args.do_train and data_args.print_samples:
        for index in random.sample(range(len(train_dataset)), 3):
            logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
    else:
        logger.info("Not printing samples, to change this set --print_samples=True")

    # METRICS FOR EVALUATION
    # --------------------------------------------------------------------------

    if data_args.metric_name is not None:
        metric = (
            evaluate.load(data_args.metric_name, config_name="multilabel")
            if is_multi_label
            else evaluate.load(data_args.metric_name)
        )
        logger.info(f"Using metric {data_args.metric_name} for evaluation.")
    else:
        if is_regression:
            metric = evaluate.load("mse")
            logger.info(
                "Using mean squared error (mse) as regression score, you can use --metric_name to overwrite."
            )
        else:
            if is_multi_label:
                metric = evaluate.load("f1", config_name="multilabel")
                logger.info(
                    "Using multilabel F1 for multi-label classification task, you can use --metric_name to overwrite."
                )
            else:
                metric = evaluate.load("accuracy")
                logger.info(
                    "Using accuracy as classification score, you can use --metric_name to overwrite."
                )

    def compute_metrics(p: EvalPrediction):
        preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
        if is_multi_label:
            preds = sigmoid(preds)
            preds_binary = np.array(
                preds > 0.5, dtype=int
            )  # Convert probabilities to binary predictions

            precision, recall, f1, _ = precision_recall_fscore_support(
                p.label_ids, preds_binary, average="micro", zero_division=0
            )
            precision_macro, recall_macro, f1_macro, _ = (
                precision_recall_fscore_support(
                    p.label_ids, preds_binary, average="macro", zero_division=0
                )
            )
            precision_weighted, recall_weighted, f1_weighted, _ = (
                precision_recall_fscore_support(
                    p.label_ids, preds_binary, average="weighted", zero_division=0
                )
            )

            hamming = hamming_loss(p.label_ids, preds_binary)
            jaccard = jaccard_score(
                p.label_ids, preds_binary, average="macro", zero_division=0
            )

            result = {
                "precision_micro": precision,
                "recall_micro": recall,
                "f1_micro": f1,
                "precision_macro": precision_macro,
                "recall_macro": recall_macro,
                "f1_macro": f1_macro,
                "precision_weighted": precision_weighted,
                "recall_weighted": recall_weighted,
                "f1_weighted": f1_weighted,
                "hamming_loss": hamming,
                "jaccard_score": jaccard,
            }
        else:
            preds = np.argmax(preds, axis=1)
            result = metric.compute(predictions=preds, references=p.label_ids)
        return result

    # Data collator will default to DataCollatorWithPadding when the tokenizer is passed to Trainer, so we change it if
    # we already did the padding.
    if data_args.pad_to_max_length:
        data_collator = default_data_collator
    elif training_args.fp16:
        data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
    else:
        data_collator = None

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset if training_args.do_train else None,
        eval_dataset=eval_dataset if training_args.do_eval else None,
        compute_metrics=compute_metrics,
        tokenizer=tokenizer,
        data_collator=data_collator,
    )

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint

        with torch.backends.cuda.sdp_kernel(
            enable_flash=True,
            enable_math=False,
            enable_mem_efficient=False,
        ):
            train_result = trainer.train(resume_from_checkpoint=checkpoint)
        metrics = train_result.metrics
        max_train_samples = (
            data_args.max_train_samples
            if data_args.max_train_samples is not None
            else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))
        trainer.save_model()  # Saves the tokenizer too for easy upload
        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        logger.info("*** Evaluate ***")
        metrics = trainer.evaluate(eval_dataset=eval_dataset)
        max_eval_samples = (
            data_args.max_eval_samples
            if data_args.max_eval_samples is not None
            else len(eval_dataset)
        )
        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    if training_args.do_predict:
        logger.info("*** Predict ***")

        # Assuming 'label' in predict_dataset contains the true labels as one-hot encoded vectors
        true_labels = (
            predict_dataset["label"]
            if "label" in predict_dataset.column_names
            else None
        )
        # logger.info(f"THE TRUE LABELS ARE {true_labels}")

        # Remove the 'label' column if it exists to avoid issues during prediction
        if "label" in predict_dataset.features:
            predict_dataset = predict_dataset.remove_columns("label")

        predictions = trainer.predict(
            predict_dataset, metric_key_prefix="predict"
        ).predictions

        if is_regression:
            predictions = np.squeeze(predictions)
        elif is_multi_label:
            predictions = np.array([np.where(p > 0, 1, 0) for p in predictions])
        else:
            predictions = np.argmax(predictions, axis=1)

        # logger.info(f"THE PREDICTIONS ARE {predictions}")

        output_predict_file = os.path.join(
            training_args.output_dir, "predict_results.txt"
        )
        if trainer.is_world_process_zero():
            with open(output_predict_file, "w") as writer:
                logger.info("***** Predict results *****")
                writer.write("index\tprediction\ttrue_label\n")

                for index, item in enumerate(predictions):
                    # Format predictions as lists
                    prediction_list = (
                        [label_list[i] for i, pred in enumerate(item) if pred]
                        if is_multi_label
                        else [label_list[item]]
                    )

                    # Format true labels as lists, if available
                    true_label_list = []
                    if true_labels is not None:
                        true_label = true_labels[index]
                        true_label_list = [
                            label_list[i]
                            for i, label in enumerate(true_label)
                            if label == 1
                        ]

                    # Convert lists to string representation for writing to file
                    prediction_str = str(prediction_list)
                    true_label_str = (
                        str(true_label_list) if true_labels is not None else "[]"
                    )

                    writer.write(f"{index}\t{prediction_str}\t{true_label_str}\n")

        logger.info(f"Predict results saved at {output_predict_file}")

    kwargs = {
        "finetuned_from": model_args.model_name_or_path,
        "tasks": "text-classification",
        "dataset": data_args.dataset_name,
    }

    # pop kwargs huggingface hub doesnt like
    if Path(model_args.model_name_or_path).exists():
        kwargs.pop("finetuned_from")
    if data_args.train_file is not None:
        kwargs.pop("dataset")

    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)

    logger.info("End of training")


def _mp_fn(index):
    # For xla_spawn (TPUs)
    main()


if __name__ == "__main__":
    main()