def benchmark_reader(): reader_results = [] doc_store = get_document_store("elasticsearch") docs, labels = eval_data_from_file(data_dir / filename) index_to_doc_store(doc_store, docs, None, labels) for reader_name in reader_models: for reader_type in reader_types: try: reader = get_reader(reader_name, reader_type) results = reader.eval(document_store=doc_store, doc_index=doc_index, label_index=label_index, device="cuda") # print(results) results["passages_per_second"] = n_passages / results[ "reader_time"] results["reader"] = reader_name results["error"] = "" reader_results.append(results) except Exception as e: results = { 'EM': 0., 'f1': 0., 'top_n_accuracy': 0., 'top_n': 0, 'reader_time': 0., "passages_per_second": 0., "seconds_per_query": 0., 'reader': reader_name, "error": e } reader_results.append(results) reader_df = pd.DataFrame.from_records(reader_results) reader_df.to_csv("reader_results.csv")
def prepare_data(data_dir, filename_gold, filename_negative, n_docs=None, n_queries=None, add_precomputed=False): """ filename_gold points to a squad format file. filename_negative points to a csv file where the first column is doc_id and second is document text. If add_precomputed is True, this fn will look in the embeddings files for precomputed embeddings to add to each Document """ gold_docs, labels = eval_data_from_file(data_dir / filename_gold) # Reduce number of docs gold_docs = gold_docs[:n_docs] # Remove labels whose gold docs have been removed doc_ids = [x.id for x in gold_docs] labels = [x for x in labels if x.document_id in doc_ids] # Filter labels down to n_queries selected_queries = list(set(f"{x.document_id} | {x.question}" for x in labels)) selected_queries = selected_queries[:n_queries] labels = [x for x in labels if f"{x.document_id} | {x.question}" in selected_queries] n_neg_docs = max(0, n_docs - len(gold_docs)) neg_docs = prepare_negative_passages(data_dir, filename_negative, n_neg_docs) docs = gold_docs + neg_docs if add_precomputed: docs = add_precomputed_embeddings(data_dir / embeddings_dir, embeddings_filenames, docs) return docs, labels
def benchmark_reader(ci=False, update_json=False, save_markdown=False, **kwargs): if ci: reader_models = reader_models_ci max_docs = 100 # heuristic to estimate num of passages for the reduced num of docs n_passages = n_total_passages * (max_docs / n_total_docs) else: reader_models = reader_models_full max_docs = None n_passages = n_total_passages reader_results = [] doc_store = get_document_store("elasticsearch") # download squad data _download_extract_downstream_data(input_file=data_dir / filename) docs, labels = eval_data_from_file(data_dir / filename, max_docs) index_to_doc_store(doc_store, docs, None, labels) for reader_name in reader_models: for reader_type in reader_types: logger.info( f"##### Start reader run - model:{reader_name}, type: {reader_type} ##### " ) try: reader = get_reader(reader_name, reader_type) results = reader.eval(document_store=doc_store, doc_index=doc_index, label_index=label_index, device="cuda") # results = reader.eval_on_file(data_dir, filename, device="cuda") print(results) results["passages_per_second"] = n_passages / results[ "reader_time"] results["reader"] = reader_name results["error"] = "" reader_results.append(results) except Exception as e: results = { 'EM': 0., 'f1': 0., 'top_n_accuracy': 0., 'top_n': 0, 'reader_time': 0., "passages_per_second": 0., "seconds_per_query": 0., 'reader': reader_name, "error": e } reader_results.append(results) reader_df = pd.DataFrame.from_records(reader_results) reader_df.to_csv(results_file) if save_markdown: md_file = results_file.replace(".csv", ".md") with open(md_file, "w") as f: f.write(str(reader_df.to_markdown())) if update_json: populate_reader_json()
def add_eval_data(self, filename: str, doc_index: str = "eval_document", label_index: str = "label"): """ Adds a SQuAD-formatted file to the DocumentStore in order to be able to perform evaluation on it. :param filename: Name of the file containing evaluation data :type filename: str :param doc_index: Elasticsearch index where evaluation documents should be stored :type doc_index: str :param label_index: Elasticsearch index where labeled questions should be stored :type label_index: str """ docs, labels = eval_data_from_file(filename) self.write_documents(docs, index=doc_index) self.write_labels(labels, index=label_index)