import os import re import time import shutil from pathlib import Path from functools import partial from typing import Union, Dict, List import torch from torch.utils.data import DataLoader import datasets from datasets import load_dataset, Dataset from transformers import AutoTokenizer, PreTrainedTokenizer from huggingface_hub import Repository, create_repo, HfApi from optimum.onnxruntime import ( AutoOptimizationConfig, ORTModelForFeatureExtraction, ORTOptimizer, ) os.environ["TOKENIZERS_PARALLELISM"] = "false" opt_configs = { "O2": AutoOptimizationConfig.O2(), "O3": AutoOptimizationConfig.O3(), "O4": AutoOptimizationConfig.O4(), } def get_batch_size(device_name: str, model_name: str, opt_level: str): """ TODO: run actual tests T4 has 16GB A10 has 24GB Args: device_name (`str`): The name of the GPU device in use. model_name (`str`): The name of the model in use. opt_level (`str`): The optimization level in use. Returns: `int`: The batch size to use. """ if "small" in model_name: bs = 192 elif "base" in model_name: bs = 128 elif "large" in model_name: bs = 64 else: bs = 32 if "A10" in device_name: bs *= 2 if opt_level == "O4": bs *= 2 return bs def mean_pooling(last_hidden_state: torch.Tensor, attention_mask: torch.Tensor): """ Mean pool the token embeddings. Args: last_hidden_state (`tuple`): The output of the model. attention_mask (`torch.Tensor`): The attention mask. Returns: `torch.Tensor`: The mean pooled embeddings. """ input_mask_expanded = ( attention_mask.unsqueeze(-1).expand(last_hidden_state.size()).float() ) return torch.sum(last_hidden_state * input_mask_expanded, 1) / torch.clamp( input_mask_expanded.sum(1), min=1e-9 ) def load_hf_dataset(ds_name: str, ds_config: str = None, ds_split: str = "train"): """ Load a dataset from the HuggingFace Hub. Will be streaming so as to not load the whole dataset to local storage. Args: ds_name (`str`): The name of the dataset to load. ds_config (`str`, *optional*, Defaults to `None`): The configuration of the dataset to load. ds_split (`str`, *optional*, Defaults to `"train"`): The split of the dataset to load. Returns: ds (`datasets.IterableDataset`): The loaded dataset. """ if ds_config == "": ds_config = None if ds_name == "wikipedia": pattern = re.compile(r"[^a-zA-Z0-9]") folder = Path("/data") / pattern.sub("", ds_name+ds_config) files = list(map(str, folder.glob("chunk_*"))) return load_dataset("parquet", data_files=files, split="train") ds = load_dataset(ds_name, ds_config, split=ds_split) return ds def download_wikipedia(ds_name, ds_config, num2skip, num2embed): ds = load_dataset(ds_name, ds_config, streaming=True, split="train") def gen(): if num2embed > 0: for example in ds.skip(num2skip).take(num2embed): yield {"text": example["text"]} else: for example in ds.skip(num2skip): yield {"text": example["text"]} ds2 = Dataset.from_generator(gen) chunk_size = 20_000 filenames = [] pattern = re.compile(r"[^a-zA-Z0-9]") folder = Path("/data") / pattern.sub("", ds_name+ds_config) folder.mkdir(exist_ok=True, parents=True) for chunk_num, start_idx in enumerate(range(0, len(ds2), chunk_size)): end_idx = min(start_idx + chunk_size, len(ds2)) temp = ds2.select(range(start_idx, end_idx)) temp.to_parquet(str(folder / f"chunk_{chunk_num}")) filenames.append(str(folder / f"chunk_{chunk_num}")) return load_dataset("parquet", data_files=filenames, split="train") def get_model_and_tokenizer(model_name: str, optimization_level: str, progress): """ Load the model and tokenizer from the HuggingFace Hub. If the model is not already optimized, optimize it and save it to the local directory. Args: model_name (`str`): The name of the model to load. optimization_level (`str`): The optimization level to use. Should be one of `"O2"`, `"O3"`, or `"O4"`. Returns: model (`ORTModelForFeatureExtraction`): The optimized model. tokenizer (`PreTrainedTokenizer`): The tokenizer. """ optimized_model_name = f"model_optimized_{optimization_level}.onnx" model_dir = Path(model_name.replace("/", "_")) if not (model_dir / optimized_model_name).exists(): if progress is not None: progress(0.2, "Downloading tokenizer...") tokenizer = AutoTokenizer.from_pretrained(model_name) tokenizer.save_pretrained(model_dir) if progress is not None: progress(0.4, "Downloading model...") model = ORTModelForFeatureExtraction.from_pretrained(model_name, export=True) model.save_pretrained(model_dir) optimizer = ORTOptimizer.from_pretrained(model) optimization_config = opt_configs[optimization_level] if progress is not None: progress(0.6, "Optimizing model...") optimizer.optimize(save_dir=model_dir, optimization_config=optimization_config) Path(model_dir / "model_optimized.onnx").rename( model_dir / optimized_model_name ) else: tokenizer = AutoTokenizer.from_pretrained(model_dir) if progress is not None: progress(0.8, "Loading optimized model and tokenizer...") return ( ORTModelForFeatureExtraction.from_pretrained( model_dir, file_name=optimized_model_name, provider="CUDAExecutionProvider", ), tokenizer, ) def tokenize( examples: Dict[str, List[str]], tokenizer: PreTrainedTokenizer, column_name: str = "text", padding: Union[bool, str] = True, max_length: int = 512, ): """ Tokenize the examples using the tokenizer. Args: examples (`Dict[str, List[str]]`): examples to tokenize tokenizer (`PreTrainedTokenizer`): tokenizer to use column_name (`str`, *optional*, defaults to `text`): column name to use for tokenization. Defaults to `text` padding (`bool`, *optional*, defaults to `True`): whether to pad the examples. Defaults to `True` Use `"max_length"` if using `O4` optimization level If `True`, the batch will be padded to the longest in the batch. max_length (`int`, *optional*, Defaults to `512`): max length to use for the model. Defaults to `512`. Any sequences longer will be truncated. If padding is `"max_length"`, the padding will be added until the sequence is of length `max_length`. Returns: `Dict[str, List[List[int]]]`: tokenized examples """ # TODO: add lengths, sort by length, use dynamic padding # TODO: option for controlling length for models that can go shorter/longer than 512 return tokenizer( examples[column_name], truncation=True, padding=padding, max_length=max_length ) def collate_fn(examples, tokenizer=None, padding=None, column_name="text"): try: keys = examples[0].keys() except KeyError: print(examples) else: batch = {k: [] for k in examples[0].keys()} tokenized = tokenizer( [x[column_name] for x in examples], truncation=True, padding=padding, max_length=512, return_tensors="pt" ) tokenized[column_name] = [x[column_name] for x in examples] return tokenized # for example in examples: # for k, v in example.items(): # batch[k].append(v) # return { # k: torch.tensor(v, dtype=torch.long) if k in {"attention_mask", "input_ids"} else v for k, v in batch.items() # } @torch.inference_mode() def batch_embed( ds: datasets.IterableDataset, model: ORTModelForFeatureExtraction, tokenizer: PreTrainedTokenizer, model_name: str, column_name: str, new_dataset_id: str, opt_level: str, upload_batch_size: int = 10_000, map_batch_size: int = 2000, num2skip: int = 0, num2embed: int = -1, progress=None, ): """ Run the model on the dataset and upload the embeddings to the hub. Args: ds (`datasets.Dataset`): dataset to embed. From `load_hf_dataset` model (`ORTModelForFeatureExtraction`): model to use for embedding. From `get_model_and_tokenizer` tokenizer (`AutoTokenizer`): tokenizer to use for embedding. From `get_model_and_tokenizer` model_name (`str`): name of the model to use. Used to determine batch size. column_name (`str`): column name to use for embedding. Default option in gradio app is `text` new_dataset_id (`str`): id of the new dataset to create. Should include username or organization. e.g. nbroad/new-embeddings opt_level (`str`): optimization level to use. Should be one of `O2`, `O3`, `O4` See here for more details on optimization levels: https://huggingface.co/docs/optimum/onnxruntime/usage_guides/optimization#optimization-configuration upload_batch_size (`int`, *optional*, defaults to `10_000`): number of embeddings to upload at once. Defaults to 10,000. map_batch_size (`int`, *optional*, defaults to `2000`): number of examples to tokenize at once. Defaults to 2000. num2skip (`int`, *optional*, defaults to `0`): number of examples to skip. Defaults to 0. num2embed (`int`, *optional*, defaults to `-1`): number of examples to embed. Defaults to -1, which means all examples. Returns: current_count (`int`): number of examples embedded so far time_taken (`float`): time taken to embed the examples in seconds """ api = HfApi( token=os.environ["HF_TOKEN"], ) username = api.whoami()["name"] if "/" not in new_dataset_id: new_dataset_id = username + "/" + new_dataset_id repo = init_git_repo(new_dataset_id) # ds = ds.map( # tokenize, # batched=True, # batch_size=map_batch_size, # fn_kwargs={ # "tokenizer": tokenizer, # "column_name": column_name, # "padding": "max_length" if opt_level == "O4" else True, # }, # ) embeds = [] texts = [] # last_count keeps track of how many had been embedded since last push last_count = 0 # current count keeps track of how many have been embedded in total current_count = 0 device = torch.device("cuda" if torch.cuda.is_available() else "cpu") inference_bs = get_batch_size(torch.cuda.get_device_name(0), model_name, opt_level) # skip through some examples if specified if num2skip > 0: ds = ds.skip(num2skip) start_time = time.time() for batch in DataLoader( ds, batch_size=inference_bs, shuffle=False, num_workers=2, pin_memory=True, drop_last=False, collate_fn=partial( collate_fn, column_name=column_name, tokenizer=tokenizer, padding="max_length" if opt_level == "O4" else True ) ): ids = batch["input_ids"].to(device) mask = batch["attention_mask"].to(device) t_ids = torch.zeros_like(ids) outputs = model(input_ids=ids, attention_mask=mask, token_type_ids=t_ids) embeds.extend(mean_pooling(outputs[0], mask).cpu().tolist()) texts.extend(batch[column_name]) current_count += ids.shape[0] # Check if we have embedded enough examples if current_count >= num2embed: diff = current_count - num2embed embeds = embeds[:-diff] texts = texts[:-diff] current_count = num2embed break # Periodically upload to the hub if len(embeds) > upload_batch_size: push_to_repo(new_dataset_id, last_count, current_count, embeds, texts, api) embeds = [] texts = [] last_count = current_count # Provide updates if progress is not None: progress( (current_count, None), "Embedding docs...", total=None, unit="Docs Embedded", ) time_taken = time.time() - start_time # If there are any remaining embeddings, upload them if len(embeds) > 0: push_to_repo(new_dataset_id, last_count, current_count, embeds, texts, api) return current_count - num2skip, time_taken def init_git_repo(repo_id: str): """ Initialize a git repo for the new dataset. ***Removes existing local folder if exists*** Args: repo_id (`str`): id of the new dataset to create. Should include username or organization. e.g. nbroad/new-embeddings """ local_dir = repo_id.replace("/", "_") create_repo( repo_id, repo_type="dataset", token=os.environ["HF_TOKEN"], private=True, exist_ok=True, ) try: repo = Repository( local_dir=local_dir, clone_from=repo_id, repo_type="dataset", token=os.environ["HF_TOKEN"], skip_lfs_files=True, ) except EnvironmentError: shutil.rmtree(local_dir) repo = Repository( local_dir=local_dir, clone_from=repo_id, repo_type="dataset", token=os.environ["HF_TOKEN"], skip_lfs_files=True, ) if repo is not None: repo.git_pull() return repo def push_to_repo( repo_id: str, last_count: int, current_count: int, embeds: List[List[float]], texts: List[str], api: HfApi, ): """ Push embeddings to the repo. Args: repo_id (`str`): id of the new dataset to create. Should include username or organization. last_count (`int`): last count of embeddings. This is the number of embeddings that have already been pushed. current_count (`int`): current count of embeddings. This is the number of embeddings that have been pushed after this batch. embeds (`List[List[float]]`): list of embeddings to push to the repo texts (`List[str]`): list of texts to push to the repo api (`huggingface_hub.HfApi`): api to use to push to the repo """ temp_ds = Dataset.from_dict( { "embedding": embeds, "text": texts, } ) local_dir = repo_id.replace("/", "_") data_dir = Path(local_dir) / "data" data_dir.mkdir(exist_ok=True, parents=True) # use zfill so sorting puts the files in order filename = f"embeddings_{str(last_count).zfill(8)}_{current_count}.parquet" filepath = str(data_dir / filename) temp_ds.to_parquet(filepath) files = sorted(list(data_dir.glob("*.parquet"))) api.upload_file( path_or_fileobj=filepath, path_in_repo=f"data/{filename}", repo_id=repo_id, repo_type="dataset", run_as_future=True, token=os.environ["HF_TOKEN"], commit_message=f"Embedded examples {last_count} thru {current_count}", ) # Delete old files if len(files) > 4: for file in files[:2]: file.unlink()