run_clip.py

#!/usr/bin/env python
# Copyright 2022 The HuggingFace 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.
"""
Training a CLIP like dual encoder models using text and vision encoders in the library.
The script can be used to train CLIP like models for languages other than English by using
a text encoder pre-trained in the desired language. Currently this script supports the following vision
and text models:
Vision models: ViT(https://huggingface.co/models?filter=vit), CLIP (https://huggingface.co/models?filter=clip)
Text models: BERT, ROBERTa (https://huggingface.co/models?filter=fill-mask)
"""

import logging
import os
import sys
from dataclasses import dataclass, field
from typing import Optional

import torch
import transformers
from datasets import Image, load_dataset
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
from transformers import (
    AutoImageProcessor,
    AutoModel,
    AutoTokenizer,
    HfArgumentParser,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version

from optimum.habana import GaudiConfig, GaudiTrainer, GaudiTrainingArguments
from optimum.habana.utils import set_seed


try:
    from optimum.habana.utils import check_optimum_habana_min_version
except ImportError:

    def check_optimum_habana_min_version(*a, **b):
        return ()


logger = logging.getLogger(__name__)

# Will error if the minimal version of Transformers and Optimum Habana are not installed. Remove at your own risks.
check_min_version("4.55.0")
check_optimum_habana_min_version("1.19.0.dev0")

require_version("datasets>=4.0.0", "To fix: pip install -r examples/pytorch/contrastive-image-text/requirements.txt")


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

    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"}
    )
    image_processor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."})
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    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 `hf auth login` (stored in `~/.huggingface`)."
            )
        },
    )
    trust_remote_code: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to trust the execution of code from datasets/models defined on the Hub."
                " 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."
            )
        },
    )
    freeze_vision_model: bool = field(
        default=False, metadata={"help": "Whether to freeze the vision model parameters or not."}
    )
    freeze_text_model: bool = field(
        default=False, metadata={"help": "Whether to freeze the text model parameters or not."}
    )


@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",
        metadata={"help": "The name of the column in the datasets containing the image file."},
    )
    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=128,
        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."},
    )
    mediapipe_dataloader: bool = field(
        default=False, metadata={"help": "Use gaudi2/gaudi3 HW mediapipe over regular dataloader."}
    )

    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."


# Torchvision preprocessing for images
class Transform(torch.nn.Module):
    def __init__(self, image_size, mean, std):
        super().__init__()
        self.transforms = transforms.Compose(
            [
                transforms.Lambda(lambda img: img.convert("RGB")),
                transforms.Resize((image_size, image_size), interpolation=InterpolationMode.BICUBIC),
                transforms.ToTensor(),
                transforms.ConvertImageDtype(torch.float),
                transforms.Normalize(mean, std),
            ]
        )

    def forward(self, x) -> torch.Tensor:
        """`x` is a PIL.Image.Image provided by datasets.Image"""
        with torch.no_grad():
            return self.transforms(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 main():
    # 1. Parse input arguments
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, GaudiTrainingArguments))
    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()

    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
    # information sent is the one passed as arguments along with your Python/PyTorch versions.
    send_example_telemetry("run_clip", model_args, data_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)],
    )

    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)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    gaudi_config = GaudiConfig.from_pretrained(
        training_args.gaudi_config_name,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        token=model_args.token,
    )

    # Log on each process the small summary:
    mixed_precision = training_args.bf16 or gaudi_config.use_torch_autocast
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, "
        + f"distributed training: {training_args.parallel_mode.value == 'distributed'}, "
        + f"mixed-precision training: {mixed_precision}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    # 3. Detecting last checkpoint and eventually continue from 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"Resuming training from {last_checkpoint}")

    # Load dataset
    if data_args.dataset_name is not None:
        dataset = load_dataset(
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            token=model_args.token,
            trust_remote_code=False,
        )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        dataset = load_dataset(
            "json" if data_args.train_file and data_args.train_file.endswith("json") else "csv",
            data_files=data_files,
            cache_dir=model_args.cache_dir,
            token=model_args.token,
        )
    dataset = dataset.cast_column("image", Image(decode=True))

    # Load model, tokenizer, processor
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name or model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        use_fast=model_args.use_fast_tokenizer,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
    )
    image_processor = AutoImageProcessor.from_pretrained(
        model_args.image_processor_name or model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
    )
    model = AutoModel.from_pretrained(
        model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        token=model_args.token,
        trust_remote_code=model_args.trust_remote_code,
    )

    if model_args.freeze_vision_model:
        for p in model.vision_model.parameters():
            p.requires_grad = False
    if model_args.freeze_text_model:
        for p in model.text_model.parameters():
            p.requires_grad = False

    set_seed(training_args.seed)

    # Ensure there is at least one task to do
    if not (training_args.do_train or training_args.do_eval):
        logger.info("Nothing to do.")
        return

    image_column = data_args.image_column or "image"
    caption_column = data_args.caption_column or "caption"

    image_transformations = Transform(
        model.config.vision_config.image_size,
        image_processor.image_mean,
        image_processor.image_std,
    )

    # Tokenization + transform
    def tokenize_captions(examples):
        captions = list(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):
        examples["pixel_values"] = [image_transformations(image) for image in examples[image_column]]
        return examples

    # Apply preprocessing
    if training_args.do_train:
        train_dataset = dataset["train"]
        if data_args.max_train_samples:
            train_dataset = train_dataset.select(range(min(len(train_dataset), data_args.max_train_samples)))
        train_dataset = train_dataset.map(
            tokenize_captions,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[],
            desc="Tokenizing train captions",
        )
        train_dataset.set_transform(transform_images)

    if training_args.do_eval:
        eval_dataset = dataset["validation"]
        if data_args.max_eval_samples:
            eval_dataset = eval_dataset.select(range(min(len(eval_dataset), data_args.max_eval_samples)))
        eval_dataset = eval_dataset.map(
            tokenize_captions,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[],
            desc="Tokenizing eval captions",
        )
        eval_dataset.set_transform(transform_images)

    # Trainer
    trainer = GaudiTrainer(
        model=model,
        gaudi_config=gaudi_config,
        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,
    )

    # Train
    if training_args.do_train:
        checkpoint = training_args.resume_from_checkpoint or last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()
        tokenizer.save_pretrained(training_args.output_dir)
        image_processor.save_pretrained(training_args.output_dir)
        trainer.log_metrics("train", train_result.metrics)
        trainer.save_metrics("train", train_result.metrics)
        trainer.save_state()

    # Eval
    if training_args.do_eval:
        metrics = trainer.evaluate()
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    # Model card
    finetuned_from = None if os.path.isdir(model_args.model_name_or_path) else model_args.model_name_or_path
    kwargs = {"finetuned_from": finetuned_from, "tasks": "contrastive-image-text-modeling"}
    if data_args.dataset_name:
        kwargs["dataset_tags"] = data_args.dataset_name
        kwargs["dataset"] = data_args.dataset_name

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


if __name__ == "__main__":
    main()