clip/train_clip_with_hf_trainer.py

import logging
import os
import sys
import transformers
import dataclasses
import ipdb
from dataclasses import dataclass, field
from typing import Optional, List

import torch
from datasets import load_dataset
from PIL import Image
from torchvision.io import ImageReadMode, read_image
from torchvision.transforms import CenterCrop, ConvertImageDtype, Normalize, Resize
from torchvision.transforms.functional import InterpolationMode

from transformers import (
    MODEL_FOR_CAUSAL_LM_MAPPING,
    TrainingArguments,
    default_data_collator,
    is_torch_tpu_available,
    set_seed,
)

# We use torchvision for faster image pre-processing. The transforms are implemented as nn.Module,
# so we jit it to be faster.


logger = logging.getLogger(__name__)

dataset_name_mapping = {
    "image_caption_dataset.py": ("image_path", "caption"),
}


def dataclass_from_dict(klass, d):
    try:
        fieldtypes = {f.name: f.type for f in dataclasses.fields(klass)}
        return klass(**{f: dataclass_from_dict(fieldtypes[f], d[f]) for f in d})
    except:
        return d  # Not a dataclass field


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

    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)."}
    )
    data_dir: Optional[str] = field(default=None, metadata={"help": "The data directory containing input files."})
    image_column: Optional[str] = field(
        default="image_path",
        metadata={"help": "The name of the column in the datasets containing the full image file paths."},
    )
    caption_column: Optional[str] = field(
        default="caption",
        metadata={"help": "The name of the column in the datasets containing the image captions."},
    )
    train_file: Optional[str] = field(
        default=None, metadata={"help": "The input training data file (a jsonlines file)."}
    )
    validation_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input evaluation data file (a jsonlines file)."},
    )
    max_seq_length: Optional[int] = field(
        default=77,
        metadata={
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    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."
            )
        },
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
    image_mean: Optional[str] = field(
        default=None, metadata={"help": "image preprocessing mean"}
    )
    image_std: Optional[str] = field(
        default=None, metadata={"help": "image preprocessing std"}
    )

    def __post_init__(self):
        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
            raise ValueError("Need either a dataset name or a training/validation file.")
        else:
            if self.train_file is not None:
                extension = self.train_file.split(".")[-1]
                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension == "json", "`validation_file` should be a json file."


class Transform(torch.nn.Module):
    def __init__(self, image_size, mean, std):
        super().__init__()
        self.transforms = torch.nn.Sequential(
            Resize([image_size], interpolation=InterpolationMode.BICUBIC),
            CenterCrop(image_size),
            ConvertImageDtype(torch.float),
            Normalize(mean, std),
        )

    def forward(self, x) -> torch.Tensor:
        """`x` should be an instance of `PIL.Image.Image`"""
        with torch.no_grad():
            x = self.transforms(x)
        return x

def collate_fn(examples):
    pixel_values = torch.stack([example["pixel_values"] for example in examples])
    input_ids = torch.tensor([example["input_ids"] for example in examples], dtype=torch.long)
    attention_mask = torch.tensor([example["attention_mask"] for example in examples], dtype=torch.long)
    return {
        "pixel_values": pixel_values,
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "return_loss": True,
    }


def train_with_hf_trainer(model, tokenizer, data_args, training_args, **kwargs):

    import evaluate
    import datasets

    from transformers import Trainer
    from datasets import load_dataset
    from towhee.trainer.training_config import get_dataclasses_help

    print('**** DataTrainingArguments ****')
    get_dataclasses_help(DataTrainingArguments)
    data_args = dataclass_from_dict(DataTrainingArguments, data_args)

    print('**** TrainingArguments ****')
    get_dataclasses_help(TrainingArguments)
    training_args = dataclass_from_dict(TrainingArguments, training_args)

    # 2. 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)],
    ) 

    # 2. Setup logging
    #+ training_args
    #log_level = training_args.get_process_log_level()
    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()

    #!!!!!!!!
    temp_cache_dir = data_args.cache_dir
    # Log on each process the small summary:
    #training_args
    # +local_rank
    # +device
    # +n_gpu
    # +local_rank
    # +fp16
    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}")
    
    # 3. Detecting last checkpoint and eventualy continue from last checkpoint
    ### place holder ###
    ### place holder ###
    ### place holder ###


    # 4. Load dataset
    # 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 image path and the second column for the
    # captions (unless you specify column names for this with the `image_column` and `caption_column` arguments).
    #

    #data_args
    # +dataset_name
    # +dataset_config_name
    # +cache_dir
    # +data_dir
    # +train_file
    # +validation_file
    # +test_file
    #model_args
    # +use_auth_token 
    # +cache_dir
    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=temp_cache_dir,
            keep_in_memory=False,
            data_dir=data_args.data_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]
        dataset = load_dataset(
            extension,
            data_files=data_files,
            cache_dir=temp_cache_dir,
        #    use_auth_token=True if model_args.use_auth_token else None,
        )
    # 5. Load pretrained model, tokenizer, and feature extractor
    #model_args
    # +tokenizer_name
    # +cache_dir
    # +use_fast_tokenizer

    #@if model_args.tokenizer_name:
    #@    tokenizer = AutoTokenizer.from_pretrained(
    #@        model_args.tokenizer_name, cache_dir=temp_cache_dir, use_fast=model_args.use_fast_tokenizer
    #@    )
    #@# --- for CLIP, tokenizer is fixed

    # Load feature_extractor, in this script we only use this to get the mean and std for normalization.
    #model_args
    # +feature_extractor_name
    # +model_name_or_path
    # +cache_dir
    # +model_revision
    # +use_auth_token

    #@feature_extractor = AutoFeatureExtractor.from_pretrained(
    #@    model_args.feature_extractor_name or model_args.model_name_or_path,
    #@    cache_dir=temp_cache_dir,
    #@    revision=model_args.model_revision,
    #@    use_auth_token=True if model_args.use_auth_token else None,
    #@)
    # 只为了拿到mean 和 std

    # load model
    #@model = AutoModel.from_pretrained()
    config = model.config

    #选择text或者vision freeze住
    #model_args
    # +freeze_vision_model
    # +freeze_text_model
    def _freeze_params(module):
        for param in module.parameters():
                param.requires_grad = False

        if model_args.freeze_vision_model:
            _freeze_params(model.vision_model)

        if model_args.freeze_text_model:
            _freeze_params(model.text_model)
    #!!!!!!! freeze
    freeze_vision_model = False
    freeze_text_model = False
    if freeze_vision_model is True:
        _freeze_params(model.vision_model)

    if freeze_text_model is True:
        _freeze_params(model.text_model)

    set_seed(training_args.seed)

    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

    dataset_columns = dataset_name_mapping.get(data_args.dataset_name, None)
    if data_args.image_column is None:
        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
    else:
        image_column = data_args.image_column
        if image_column not in column_names:
            raise ValueError(
                f"--image_column' value '{data_args.image_column}' needs to be one of: {', '.join(column_names)}"
            )
    if data_args.caption_column is None:
        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
    else:
        caption_column = data_args.caption_column
        if caption_column not in column_names:
            raise ValueError(
                f"--caption_column' value '{data_args.caption_column}' needs to be one of: {', '.join(column_names)}"
            )


    #image_column = 'image_path'
    #caption_column = 'caption'

    image_mean, image_std = data_args.image_mean, data_args.image_std
    # 7. Preprocessing the datasets.
    # Initialize torchvision transforms and jit it for faster processing.
    #这个地方需要image_size,image_mean,image_std
    image_transformations = Transform(
        config.vision_config.image_size, image_mean, image_std
    )
    image_transformations = torch.jit.script(image_transformations)

    # Preprocessing the datasets.
    # We need to tokenize input captions and transform the images.
    #data_args
    # +max_seq_length
    #caption_column
    #!!!!!!!!!!!!!!!!!

    #from transformers import CLIPTokenizer, CLIPTextModel ,CLIPModel,CLIPProcessor
    #tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")

    def tokenize_captions(examples):
        captions = [caption for caption in examples[caption_column]]
        text_inputs = tokenizer(captions, max_length=data_args.max_seq_length, padding="max_length", truncation=True)
        examples["input_ids"] = text_inputs.input_ids
        examples["attention_mask"] = text_inputs.attention_mask
        return examples

    def transform_images(examples):
        images = [read_image(image_file, mode=ImageReadMode.RGB) for image_file in examples[image_column]]
        examples["pixel_values"] = [image_transformations(image) for image in images]
        return examples

    def filter_corrupt_images(examples):
        """remove problematic images"""
        valid_images = []
        for image_file in examples[image_column]:
            try:
                Image.open(image_file)
                valid_images.append(True)
            except Exception:
                valid_images.append(False)
        return valid_images
    
    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:
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))
        train_dataset = train_dataset.filter(
            filter_corrupt_images, batched=True, num_proc=data_args.preprocessing_num_workers
        )
        train_dataset = train_dataset.map(
            function=tokenize_captions,
            batched=True,
            remove_columns=[col for col in column_names if col != image_column],
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on train dataset",
        )

        # Transform images on the fly as doing it on the whole dataset takes too much time.
        train_dataset.set_transform(transform_images)
    #training_args 
    # +do_eval
    #data_args
    # +max_eval_samples
    # +preprocessing_num_workers
    # +overwrite_cache
    if training_args.do_eval:
        if "validation" not in dataset:
            raise ValueError("--do_eval requires a train validation")
        eval_dataset = dataset["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))

        eval_dataset = eval_dataset.filter(
            filter_corrupt_images, batched=True, num_proc=data_args.preprocessing_num_workers
        )
        eval_dataset = eval_dataset.map(
            function=tokenize_captions,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[col for col in column_names if col != image_column],
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on validation dataset",
        )

        # Transform images on the fly as doing it on the whole dataset takes too much time.
        eval_dataset.set_transform(transform_images)

    # 8. Initalize 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,
        data_collator=collate_fn,
    )

    # Training
    last_checkpoint = None
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()
        trainer.log_metrics("train", train_result.metrics)
        trainer.save_metrics("train", train_result.metrics)
        trainer.save_state()
    #training_args
    # +do_eval
    # 10. Evaluation
    if training_args.do_eval:
        metrics = trainer.evaluate()
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)