generate_passage_embeddings.py

# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
# 
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import os

import argparse
import csv
import logging
import pickle
from pathlib import Path
from tqdm import tqdm

import numpy as np
import torch
from torch.utils.data import DataLoader

from transformers import AutoModel, AutoTokenizer

import src.model
import src.data
import src.util
import src.slurm


logger = logging.getLogger(__name__)

def embed_passages(opt, passages, model, tokenizer):
    batch_size = opt.per_gpu_batch_size * opt.world_size
    collator = src.data.TextCollator(tokenizer, args.passage_maxlength)
    dataset = src.data.TextDataset(passages, title_prefix='title:', passage_prefix='context:')
    dataloader = DataLoader(dataset, batch_size=batch_size, drop_last=False, num_workers=10, collate_fn=collator)
    total = 0
    allids, allembeddings = [], []
    with torch.no_grad():
        for k, (ids, text_ids, text_mask) in enumerate(tqdm(dataloader)):
            embeddings = model.embed_text(
                text_ids=text_ids.cuda(), 
                text_mask=text_mask.cuda(), 
                apply_mask=model.config.apply_passage_mask,
                extract_cls=model.config.extract_cls,
            )
            embeddings = embeddings.cpu()
            total += len(ids)

            allids.append(ids)
            allembeddings.append(embeddings)
            if k % 10000 == 0:
                logger.info('Encoded passages %d', total)

    allembeddings = torch.cat(allembeddings, dim=0).numpy()
    allids = [x for idlist in allids for x in idlist]
    return allids, allembeddings


def main(opt):
    logger = src.util.init_logger(is_main=True)
    tokenizer = AutoTokenizer.from_pretrained('facebook/contriever')
    model_class = src.model.Retriever
    #model, _, _, _, _, _ = src.util.load(model_class, opt.model_path, opt)
    
    config = src.model.RetrieverConfig.from_pretrained(pretrained_model_name_or_path='facebook/contriever')
    #model = model_class.from_pretrained(opt.model_path)
    
    if opt.model_path is None:
        logging.info("Loading pretrained Contriever")
        model = model_class(config, initialize_wContriever=True)
        model = model.to(opt.device)
    else:
        model = model_class.from_pretrained(opt.model_path)
        #model, optimizer, scheduler, opt_checkpoint, global_step, best_eval_loss = \
        #    src.util.load(model_class, opt.model_path, opt, reset_params=False)
        logger.info(f"Model loaded from {opt.model_path}")


    model.eval()
    model = model.to(opt.device)
    if not opt.no_fp16:
        model = model.half()

    passages = src.util.load_passages(args.passages)

    shard_size = len(passages) // args.num_shards
    start_idx = args.shard_id * shard_size
    end_idx = start_idx + shard_size
    if args.shard_id == args.num_shards-1:
        end_idx = len(passages)

    passages = passages[start_idx:end_idx]
    logger.info(f'Embedding generation for {len(passages)} passages from idx {start_idx} to {end_idx}')

    allids, allembeddings = embed_passages(opt, passages, model, tokenizer)

    output_path = Path(args.output_path)
    save_file = output_path.parent / (output_path.name + f'_{args.shard_id:02d}')
    output_path.parent.mkdir(parents=True, exist_ok=True) 
    logger.info(f'Saving {len(allids)} passage embeddings to {save_file}')
    with open(save_file, mode='wb') as f:
        pickle.dump((allids, allembeddings), f, protocol=4)

    logger.info(f'Total passages processed {len(allids)}. Written to {save_file}.')


if __name__ == '__main__':
    parser = argparse.ArgumentParser()

    parser.add_argument('--passages', type=str, default=None, help='Path to passages (.tsv file)')
    parser.add_argument('--output_path', type=str, default='wikipedia_embeddings/passages', help='prefix path to save embeddings')
    parser.add_argument('--shard_id', type=int, default=0, help="Id of the current shard")
    parser.add_argument('--num_shards', type=int, default=1, help="Total number of shards")
    parser.add_argument('--per_gpu_batch_size', type=int, default=32, help="Batch size for the passage encoder forward pass")
    parser.add_argument('--passage_maxlength', type=int, default=200, help="Maximum number of tokens in a passage")
    parser.add_argument('--model_path', type=str, help="path to directory containing model weights and config file")
    parser.add_argument('--no_fp16', action='store_true', help="inference in fp32")
    args = parser.parse_args()

    src.slurm.init_distributed_mode(args)

    main(args)