import argparse import json import os import faiss import torch from datasets import load_dataset, Dataset from tqdm.auto import tqdm from transformers import AutoTokenizer, DPRQuestionEncoder, DPRContextEncoder from common import articles_to_paragraphs, embed_questions, embed_passages, create_kilt_datapoint, \ kilt_wikipedia_columns from common import kilt_wikipedia_paragraph_columns as columns def generate_support_docs(args): dims = 128 min_chars_per_passage = 200 device = ("cuda" if torch.cuda.is_available() else "cpu") lfqa = load_dataset("vblagoje/lfqa") ctx_tokenizer = AutoTokenizer.from_pretrained(args.ctx_encoder_name) ctx_model = DPRContextEncoder.from_pretrained(args.ctx_encoder_name).to(device) _ = ctx_model.eval() question_tokenizer = AutoTokenizer.from_pretrained(args.question_encoder_name) question_model = DPRQuestionEncoder.from_pretrained(args.question_encoder_name).to(device) _ = question_model.eval() kilt_wikipedia = load_dataset("kilt_wikipedia", split="full") kilt_wikipedia_paragraphs = kilt_wikipedia.map(articles_to_paragraphs, batched=True, remove_columns=kilt_wikipedia_columns, batch_size=512, cache_file_name=f"../data/wiki_kilt_paragraphs_full.arrow", desc="Expanding wiki articles into paragraphs") # use paragraphs that are not simple fragments or very short sentences # Wikipedia Faiss index needs to fit into a 16 Gb GPU kilt_wikipedia_paragraphs = kilt_wikipedia_paragraphs.filter( lambda x: (x["end_character"] - x["start_character"]) > min_chars_per_passage) def query_index(question, topk=7): topk = topk * 3 # grab 3x results and filter for word count question_embedding = embed_questions(question_model, question_tokenizer, [question]) scores, wiki_passages = kilt_wikipedia_paragraphs.get_nearest_examples("embeddings", question_embedding, k=topk) retrieved_examples = [] r = list(zip(wiki_passages[k] for k in columns)) for i in range(topk): retrieved_examples.append({k: v for k, v in zip(columns, [r[j][0][i] for j in range(len(columns))])}) return retrieved_examples def create_support_doc(dataset: Dataset, output_filename: str): progress_bar = tqdm(range(len(dataset)), desc="Creating supporting docs") with open(output_filename, "w") as fp: for example in dataset: wiki_passages = query_index(example["title"]) kilt_dp = create_kilt_datapoint(example, columns, wiki_passages) json.dump(kilt_dp, fp) fp.write("\n") progress_bar.update(1) if not os.path.isfile(args.index_file_name): def embed_passages_for_retrieval(examples): return embed_passages(ctx_model, ctx_tokenizer, examples, max_length=128) paragraphs_embeddings = kilt_wikipedia_paragraphs.map(embed_passages_for_retrieval, batched=True, batch_size=512, cache_file_name=args.encoded_kilt_file_name, desc="Creating faiss index") paragraphs_embeddings.add_faiss_index(column="embeddings", custom_index=faiss.IndexFlatIP(dims)) paragraphs_embeddings.save_faiss_index("embeddings", args.index_file_name) kilt_wikipedia_paragraphs.load_faiss_index("embeddings", args.index_file_name, device=0) create_support_doc(lfqa["train"], "lfqa_dpr_train_precomputed_dense_docs.json") create_support_doc(lfqa["validation"], "lfqa_dpr_validation_precomputed_dense_docs.json") if __name__ == "__main__": parser = argparse.ArgumentParser(description="Creates support docs for seq2seq model training") parser.add_argument( "--ctx_encoder_name", default="vblagoje/dpr-ctx_encoder-single-lfqa-base", help="Question encoder to use", ) parser.add_argument( "--question_encoder_name", default="vblagoje/dpr-question_encoder-single-lfqa-base", help="Question encoder to use", ) parser.add_argument( "--index_file_name", default="../data/kilt_dpr_wikipedia_first.faiss", help="Faiss index with passage embeddings", ) parser.add_argument( "--encoded_kilt_file_name", default="../data/kilt_embedded.arrow", help="Encoded KILT file name", ) main_args, _ = parser.parse_known_args() generate_support_docs(main_args)