diff --git a/examples/pytorch/contrastive-image-text/README.md b/examples/pytorch/contrastive-image-text/README.md new file mode 100644 index 0000000000..969cc56a2d --- /dev/null +++ b/examples/pytorch/contrastive-image-text/README.md @@ -0,0 +1,98 @@ + + +# VisionTextDualEncoder and CLIP model training examples + +The following example showcases how to train a CLIP-like vision-text dual encoder model +using a pre-trained vision and text encoder. + +Such a model can be used for natural language image search and potentially zero-shot image classification. +The model is inspired by [CLIP](https://openai.com/blog/clip/), introduced by Alec Radford et al. +The idea is to train a vision encoder and a text encoder jointly to project the representation of images and their +captions into the same embedding space, such that the caption embeddings are located near the embeddings +of the images they describe. + +### Download COCO dataset (2017) +This example uses COCO dataset (2017) through a custom dataset script, which requires users to manually download the +COCO dataset before training. + +```bash +mkdir data +cd data +wget http://images.cocodataset.org/zips/train2017.zip +wget http://images.cocodataset.org/zips/val2017.zip +wget http://images.cocodataset.org/zips/test2017.zip +wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip +wget http://images.cocodataset.org/annotations/image_info_test2017.zip +cd .. +``` +```suggestion + +Having downloaded COCO dataset manually you should be able to load with the `ydshieh/coc_dataset_script` dataset loading script: + +```py +COCO_DIR = "data" +ds = datasets.load_dataset("ydshieh/coco_dataset_script", "2017", data_dir=COCO_DIR) +### Create a model from a vision encoder model and a text decoder model +Next, we create a [VisionTextDualEncoderModel](https://huggingface.co/docs/transformers/model_doc/vision-text-dual-encoder#visiontextdualencoder). +The `VisionTextDualEncoderModel` class let's you load any vision and text encoder model to create a dual encoder. +Here is an example of how to load the model using pre-trained vision and text models. + +```python3 +from transformers import ( + VisionTextDualEncoderModel, + VisionTextDualEncoderProcessor, + AutoTokenizer, + AutoFeatureExtractor +) + +model = VisionTextDualEncoderModel.from_vision_text_pretrained( + "openai/clip-vit-base-patch32", "roberta-base" +) + +tokenizer = AutoTokenizer.from_pretrained("roberta-base") +feat_ext = AutoFeatureExtractor.from_pretrained("openai/clip-vit-base-patch32") +processor = VisionTextDualEncoderProcessor(feat_ext, tokenizer) + +# save the model and processor +model.save_pretrained("clip-roberta") +processor.save_pretrained("clip-roberta") +``` + +This loads both the text and vision encoders using pre-trained weights, the projection layers are randomly +initialized except for CLIP's vision model. If you use CLIP to initialize the vision model then the vision projection weights are also +loaded using the pre-trained weights. + +### Train the model +Finally, we can run the example script to train the model: + +```bash +python examples/pytorch/contrastive-image-text/run_clip.py \ + --output_dir ./clip-roberta-finetuned \ + --model_name_or_path ./clip-roberta \ + --data_dir ./data \ + --dataset_name ydshieh/coco_dataset_script \ + --dataset_config_name=2017 \ + --image_column image_path \ + --caption_column caption \ + --remove_unused_columns=False \ + --do_train --do_eval \ + --per_device_train_batch_size="64" \ + --per_device_eval_batch_size="64" \ + --learning_rate="5e-5" --warmup_steps="0" --weight_decay 0.1 \ + --overwrite_output_dir \ + --push_to_hub +``` \ No newline at end of file diff --git a/examples/pytorch/contrastive-image-text/requirements.txt b/examples/pytorch/contrastive-image-text/requirements.txt new file mode 100644 index 0000000000..a789fee85e --- /dev/null +++ b/examples/pytorch/contrastive-image-text/requirements.txt @@ -0,0 +1,3 @@ +torch>=1.5.0 +torchvision>=0.6.0 +datasets>=1.8.0 \ No newline at end of file diff --git a/examples/pytorch/contrastive-image-text/run_clip.py b/examples/pytorch/contrastive-image-text/run_clip.py new file mode 100644 index 0000000000..4b57a8e82f --- /dev/null +++ b/examples/pytorch/contrastive-image-text/run_clip.py @@ -0,0 +1,513 @@ +#!/usr/bin/env python +# coding=utf-8 +# 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 +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 + +import transformers +from transformers import ( + AutoFeatureExtractor, + AutoModel, + AutoTokenizer, + HfArgumentParser, + Trainer, + TrainingArguments, + set_seed, +) +from transformers.trainer_utils import get_last_checkpoint +from transformers.utils import check_min_version +from transformers.utils.versions import require_version + + +logger = logging.getLogger(__name__) + +# Will error if the minimal version of Transformers is not installed. Remove at your own risks. +check_min_version("4.17.0.dev0") + +require_version("datasets>=1.8.0", "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"} + ) + feature_extractor_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."}, + ) + use_auth_token: bool = field( + default=False, + metadata={ + "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script " + "with private models)." + }, + ) + 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_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=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"} + ) + 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."}, + ) + + 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." + + +dataset_name_mapping = { + "image_caption_dataset.py": ("image_path", "caption"), +} + + +# We use torchvision for faster image pre-processing. The transforms are implemented as nn.Module, +# so we jit it to be faster. +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: Image) -> torch.Tensor: + 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 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, 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() + + # 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)], + ) + + log_level = training_args.get_process_log_level() + logger.setLevel(log_level) + transformers.utils.logging.set_verbosity(log_level) + transformers.utils.logging.enable_default_handler() + transformers.utils.logging.enable_explicit_format() + + # Log on each process the small summary: + logger.warning( + f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}" + + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}" + ) + logger.info(f"Training/evaluation parameters {training_args}") + + # 3. Detecting last checkpoint and eventualy 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"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " + "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." + ) + + # 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). + # + if data_args.dataset_name is not None: + # Downloading and loading a dataset from the hub. + dataset = load_dataset( + data_args.dataset_name, + data_args.dataset_config_name, + cache_dir=model_args.cache_dir, + keep_in_memory=False, + data_dir=data_args.data_dir, + ) + else: + data_files = {} + if data_args.train_file is not None: + data_files["train"] = data_args.train_file + extension = data_args.train_file.split(".")[-1] + if data_args.validation_file is not None: + data_files["validation"] = data_args.validation_file + extension = data_args.validation_file.split(".")[-1] + if data_args.test_file is not None: + data_files["test"] = data_args.test_file + extension = data_args.test_file.split(".")[-1] + dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir) + # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at + # https://huggingface.co/docs/datasets/loading_datasets.html. + + # 5. Load pretrained model, tokenizer, and feature extractor + if model_args.tokenizer_name: + tokenizer = AutoTokenizer.from_pretrained( + model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer + ) + elif model_args.model_name_or_path: + tokenizer = AutoTokenizer.from_pretrained( + model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer + ) + else: + raise ValueError( + "You are instantiating a new tokenizer from scratch. This is not supported by this script." + "You can do it from another script, save it, and load it from here, using --tokenizer_name." + ) + + # Load feature_extractor, in this script we only use this to get the mean and std for normalization. + feature_extractor = AutoFeatureExtractor.from_pretrained( + model_args.feature_extractor_name or model_args.model_name_or_path, + cache_dir=model_args.cache_dir, + revision=model_args.model_revision, + use_auth_token=True if model_args.use_auth_token else None, + ) + + model = AutoModel.from_pretrained( + model_args.model_name_or_path, + cache_dir=model_args.cache_dir, + revision=model_args.model_revision, + use_auth_token=True if model_args.use_auth_token else None, + ) + config = model.config + + 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) + + # set seed for torch dataloaders + set_seed(training_args.seed) + + # Preprocessing the datasets. + # We need to tokenize inputs and targets. + if training_args.do_train: + column_names = dataset["train"].column_names + elif training_args.do_eval: + column_names = dataset["validation"].column_names + elif training_args.do_predict: + column_names = dataset["test"].column_names + else: + logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.") + return + + # 6. Get the column names for input/target. + 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)}" + ) + + # 7. Preprocessing the datasets. + # Initialize torchvision transforms and jit it for faster processing. + image_transformations = Transform( + config.vision_config.image_size, feature_extractor.image_mean, feature_extractor.image_std + ) + image_transformations = torch.jit.script(image_transformations) + + # Preprocessing the datasets. + # We need to tokenize input captions and transform the images. + 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: + train_dataset = train_dataset.select(range(data_args.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) + + 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: + eval_dataset = eval_dataset.select(range(data_args.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) + + if training_args.do_predict: + if "test" not in dataset: + raise ValueError("--do_predict requires a test dataset") + test_dataset = dataset["test"] + if data_args.max_eval_samples is not None: + test_dataset = test_dataset.select(range(data_args.max_eval_samples)) + + test_dataset = test_dataset.filter( + filter_corrupt_images, batched=True, num_proc=data_args.preprocessing_num_workers + ) + test_dataset = test_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 test dataset", + ) + + # Transform images on the fly as doing it on the whole dataset takes too much time. + test_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, + ) + + # 9. Training + if training_args.do_train: + checkpoint = None + if training_args.resume_from_checkpoint is not None: + checkpoint = training_args.resume_from_checkpoint + elif last_checkpoint is not None: + checkpoint = last_checkpoint + train_result = trainer.train(resume_from_checkpoint=checkpoint) + trainer.save_model() + trainer.log_metrics("train", train_result.metrics) + trainer.save_metrics("train", train_result.metrics) + trainer.save_state() + + # 10. Evaluation + if training_args.do_eval: + metrics = trainer.evaluate() + trainer.log_metrics("eval", metrics) + trainer.save_metrics("eval", metrics) + + # 11. Write Training Stats and push to hub. + kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "contrastive-image-text-modeling"} + if data_args.dataset_name is not None: + kwargs["dataset_tags"] = data_args.dataset_name + if data_args.dataset_config_name is not None: + kwargs["dataset_args"] = data_args.dataset_config_name + kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}" + else: + kwargs["dataset"] = data_args.dataset_name + + if training_args.push_to_hub: + trainer.push_to_hub(**kwargs) + else: + trainer.create_model_card(**kwargs) + + +if __name__ == "__main__": + main()