def test_mention_longest_match(): """Test longest match filtering in mention extraction.""" file_name = "lincoln_short" docs_path = f"tests/data/pure_html/{file_name}.html" doc = parse_doc(docs_path, file_name) # Mention Extraction name_ngrams = MentionNgramsPart(n_max=3) place_ngrams = MentionNgramsTemp(n_max=4) Name = mention_subclass("Name") Place = mention_subclass("Place") def is_birthplace_table_row(mention): if not mention.sentence.is_tabular(): return False ngrams = get_row_ngrams(mention, lower=True) if "birth_place" in ngrams: return True else: return False birthplace_matcher = LambdaFunctionMatcher( func=is_birthplace_table_row, longest_match_only=False ) mention_extractor_udf = MentionExtractorUDF( [Name, Place], [name_ngrams, place_ngrams], [PersonMatcher(), birthplace_matcher], ) doc = mention_extractor_udf.apply(doc) mentions = doc.places mention_spans = [x.context.get_span() for x in mentions] assert "Sinking Spring Farm" in mention_spans assert "Farm" in mention_spans assert len(mention_spans) == 23 # Clear manually for mention in doc.places[:]: doc.places.remove(mention) birthplace_matcher = LambdaFunctionMatcher( func=is_birthplace_table_row, longest_match_only=True ) mention_extractor_udf = MentionExtractorUDF( [Name, Place], [name_ngrams, place_ngrams], [PersonMatcher(), birthplace_matcher], ) doc = mention_extractor_udf.apply(doc) mentions = doc.places mention_spans = [x.context.get_span() for x in mentions] assert "Sinking Spring Farm" in mention_spans assert "Farm" not in mention_spans assert len(mention_spans) == 4
def setup_common_components(): """Set up mention/candidate extractor.""" preprocessor = HTMLDocPreprocessor("tests/data/html/") parser = Parser(None) mention_extractor = MentionExtractor( None, [Part, Temp], [MentionNgramsPart(parts_by_doc=None, n_max=3), MentionNgramsTemp(n_max=2)], [part_matcher, temp_matcher], ) candidate_extractor = CandidateExtractor(None, [PartTemp], [temp_throttler]) return { "preprocessor": preprocessor, "parser": parser, "mention_extractor": mention_extractor, "candidate_extractor": candidate_extractor, }
def test_e2e(): """Run an end-to-end test on documents of the hardware domain.""" PARALLEL = 4 max_docs = 12 fonduer.init_logging( log_dir="log_folder", format="[%(asctime)s][%(levelname)s] %(name)s:%(lineno)s - %(message)s", level=logging.INFO, ) session = fonduer.Meta.init(CONN_STRING).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) assert session.query(Document).count() == max_docs num_docs = session.query(Document).count() logger.info(f"Docs: {num_docs}") assert num_docs == max_docs num_sentences = session.query(Sentence).count() logger.info(f"Sentences: {num_sentences}") # Divide into test and train docs = sorted(corpus_parser.get_documents()) last_docs = sorted(corpus_parser.get_last_documents()) ld = len(docs) assert ld == len(last_docs) assert len(docs[0].sentences) == len(last_docs[0].sentences) assert len(docs[0].sentences) == 799 assert len(docs[1].sentences) == 663 assert len(docs[2].sentences) == 784 assert len(docs[3].sentences) == 661 assert len(docs[4].sentences) == 513 assert len(docs[5].sentences) == 700 assert len(docs[6].sentences) == 528 assert len(docs[7].sentences) == 161 assert len(docs[8].sentences) == 228 assert len(docs[9].sentences) == 511 assert len(docs[10].sentences) == 331 assert len(docs[11].sentences) == 528 # Check table numbers assert len(docs[0].tables) == 9 assert len(docs[1].tables) == 9 assert len(docs[2].tables) == 14 assert len(docs[3].tables) == 11 assert len(docs[4].tables) == 11 assert len(docs[5].tables) == 10 assert len(docs[6].tables) == 10 assert len(docs[7].tables) == 2 assert len(docs[8].tables) == 7 assert len(docs[9].tables) == 10 assert len(docs[10].tables) == 6 assert len(docs[11].tables) == 9 # Check figure numbers assert len(docs[0].figures) == 32 assert len(docs[1].figures) == 11 assert len(docs[2].figures) == 38 assert len(docs[3].figures) == 31 assert len(docs[4].figures) == 7 assert len(docs[5].figures) == 38 assert len(docs[6].figures) == 10 assert len(docs[7].figures) == 31 assert len(docs[8].figures) == 4 assert len(docs[9].figures) == 27 assert len(docs[10].figures) == 5 assert len(docs[11].figures) == 27 # Check caption numbers assert len(docs[0].captions) == 0 assert len(docs[1].captions) == 0 assert len(docs[2].captions) == 0 assert len(docs[3].captions) == 0 assert len(docs[4].captions) == 0 assert len(docs[5].captions) == 0 assert len(docs[6].captions) == 0 assert len(docs[7].captions) == 0 assert len(docs[8].captions) == 0 assert len(docs[9].captions) == 0 assert len(docs[10].captions) == 0 assert len(docs[11].captions) == 0 train_docs = set() dev_docs = set() test_docs = set() splits = (0.5, 0.75) data = [(doc.name, doc) for doc in docs] data.sort(key=lambda x: x[0]) for i, (doc_name, doc) in enumerate(data): if i < splits[0] * ld: train_docs.add(doc) elif i < splits[1] * ld: dev_docs.add(doc) else: test_docs.add(doc) logger.info([x.name for x in train_docs]) # NOTE: With multi-relation support, return values of getting candidates, # mentions, or sparse matrices are formatted as a list of lists. This means # that with a single relation, we need to index into the list of lists to # get the candidates/mentions/sparse matrix for a particular relation or # mention. # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") mention_extractor = MentionExtractor( session, [Part, Temp, Volt], [part_ngrams, temp_ngrams, volt_ngrams], [part_matcher, temp_matcher, volt_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 299 assert session.query(Temp).count() == 138 assert session.query(Volt).count() == 140 assert len(mention_extractor.get_mentions()) == 3 assert len(mention_extractor.get_mentions()[0]) == 299 assert ( len( mention_extractor.get_mentions( docs=[session.query(Document).filter(Document.name == "112823").first()] )[0] ) == 70 ) # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, volt_throttler] ) for i, docs in enumerate([train_docs, dev_docs, test_docs]): candidate_extractor.apply(docs, split=i, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 3493 assert session.query(PartTemp).filter(PartTemp.split == 1).count() == 61 assert session.query(PartTemp).filter(PartTemp.split == 2).count() == 416 assert session.query(PartVolt).count() == 4282 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0, sort=True) dev_cands = candidate_extractor.get_candidates(split=1, sort=True) test_cands = candidate_extractor.get_candidates(split=2, sort=True) assert len(train_cands) == 2 assert len(train_cands[0]) == 3493 assert ( len( candidate_extractor.get_candidates( docs=[session.query(Document).filter(Document.name == "112823").first()] )[0] ) == 1432 ) # Featurization featurizer = Featurizer(session, [PartTemp, PartVolt]) # Test that FeatureKey is properly reset featurizer.apply(split=1, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 214 assert session.query(FeatureKey).count() == 1260 # Test Dropping FeatureKey # Should force a row deletion featurizer.drop_keys(["DDL_e1_W_LEFT_POS_3_[NNP NN IN]"]) assert session.query(FeatureKey).count() == 1259 # Should only remove the part_volt as a relation and leave part_temp assert set( session.query(FeatureKey) .filter(FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]") .one() .candidate_classes ) == {"part_temp", "part_volt"} featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartVolt]) assert session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]" ).one().candidate_classes == ["part_temp"] assert session.query(FeatureKey).count() == 1259 # Inserting the removed key featurizer.upsert_keys( ["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartTemp, PartVolt] ) assert set( session.query(FeatureKey) .filter(FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]") .one() .candidate_classes ) == {"part_temp", "part_volt"} assert session.query(FeatureKey).count() == 1259 # Removing the key again featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartVolt]) # Removing the last relation from a key should delete the row featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartTemp]) assert session.query(FeatureKey).count() == 1258 session.query(Feature).delete(synchronize_session="fetch") session.query(FeatureKey).delete(synchronize_session="fetch") featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 6478 assert session.query(FeatureKey).count() == 4538 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (3493, 4538) assert F_train[1].shape == (2985, 4538) assert len(featurizer.get_keys()) == 4538 featurizer.apply(split=1, parallelism=PARALLEL) assert session.query(Feature).count() == 6692 assert session.query(FeatureKey).count() == 4538 F_dev = featurizer.get_feature_matrices(dev_cands) assert F_dev[0].shape == (61, 4538) assert F_dev[1].shape == (153, 4538) featurizer.apply(split=2, parallelism=PARALLEL) assert session.query(Feature).count() == 8252 assert session.query(FeatureKey).count() == 4538 F_test = featurizer.get_feature_matrices(test_cands) assert F_test[0].shape == (416, 4538) assert F_test[1].shape == (1144, 4538) gold_file = "tests/data/hardware_tutorial_gold.csv" labeler = Labeler(session, [PartTemp, PartVolt]) labeler.apply( docs=last_docs, lfs=[[gold], [gold]], table=GoldLabel, train=True, parallelism=PARALLEL, ) assert session.query(GoldLabel).count() == 8252 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] ce_v_max_lfs = [ LF_bad_keywords_in_row, LF_current_in_row, LF_non_ce_voltages_in_row, ] with pytest.raises(ValueError): labeler.apply(split=0, lfs=stg_temp_lfs, train=True, parallelism=PARALLEL) labeler.apply( docs=train_docs, lfs=[stg_temp_lfs, ce_v_max_lfs], train=True, parallelism=PARALLEL, ) assert session.query(Label).count() == 6478 assert session.query(LabelKey).count() == 9 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3493, 9) assert L_train[1].shape == (2985, 9) assert len(labeler.get_keys()) == 9 # Test Dropping LabelerKey labeler.drop_keys(["LF_storage_row"]) assert len(labeler.get_keys()) == 8 # Test Upserting LabelerKey labeler.upsert_keys(["LF_storage_row"]) assert "LF_storage_row" in [label.name for label in labeler.get_keys()] L_train_gold = labeler.get_gold_labels(train_cands) assert L_train_gold[0].shape == (3493, 1) L_train_gold = labeler.get_gold_labels(train_cands, annotator="gold") assert L_train_gold[0].shape == (3493, 1) gen_model = LabelModel() gen_model.fit(L_train=L_train[0], n_epochs=500, log_freq=100) train_marginals = gen_model.predict_proba(L_train[0]) disc_model = LogisticRegression() disc_model.train( (train_cands[0], F_train[0]), train_marginals, X_dev=(train_cands[0], F_train[0]), Y_dev=L_train_gold[0].reshape(-1), b=0.6, pos_label=TRUE, n_epochs=5, lr=0.001, ) test_score = disc_model.predict((test_cands[0], F_test[0]), b=0.6, pos_label=TRUE) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score == TRUE))] pickle_file = "tests/data/parts_by_doc_dict.pkl" with open(pickle_file, "rb") as f: parts_by_doc = pickle.load(f) (TP, FP, FN) = entity_level_f1( true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc ) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 < 0.7 and f1 > 0.3 stg_temp_lfs_2 = [ LF_to_left, LF_test_condition_aligned, LF_collector_aligned, LF_current_aligned, LF_voltage_row_temp, LF_voltage_row_part, LF_typ_row, LF_complement_left_row, LF_too_many_numbers_row, LF_temp_on_high_page_num, LF_temp_outside_table, LF_not_temp_relevant, ] labeler.update(split=0, lfs=[stg_temp_lfs_2, ce_v_max_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 6478 assert session.query(LabelKey).count() == 16 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3493, 16) gen_model = LabelModel() gen_model.fit(L_train=L_train[0], n_epochs=500, log_freq=100) train_marginals = gen_model.predict_proba(L_train[0]) disc_model = LogisticRegression() disc_model.train( (train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001 ) test_score = disc_model.predict((test_cands[0], F_test[0]), b=0.6, pos_label=TRUE) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score == TRUE))] (TP, FP, FN) = entity_level_f1( true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc ) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7 # Testing LSTM disc_model = LSTM() disc_model.train( (train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001 ) test_score = disc_model.predict((test_cands[0], F_test[0]), b=0.6, pos_label=TRUE) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score == TRUE))] (TP, FP, FN) = entity_level_f1( true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc ) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7 # Testing Sparse Logistic Regression disc_model = SparseLogisticRegression() disc_model.train( (train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001 ) test_score = disc_model.predict((test_cands[0], F_test[0]), b=0.6, pos_label=TRUE) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score == TRUE))] (TP, FP, FN) = entity_level_f1( true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc ) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7 # Testing Sparse LSTM disc_model = SparseLSTM() disc_model.train( (train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001 ) test_score = disc_model.predict((test_cands[0], F_test[0]), b=0.6, pos_label=TRUE) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score == TRUE))] (TP, FP, FN) = entity_level_f1( true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc ) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7 # Evaluate mention level scores L_test_gold = labeler.get_gold_labels(test_cands, annotator="gold") Y_test = L_test_gold[0].reshape(-1) scores = disc_model.score((test_cands[0], F_test[0]), Y_test, b=0.6, pos_label=TRUE) logger.info(scores) assert scores["f1"] > 0.6
def test_mention_longest_match(caplog): """Test longest match filtering in mention extraction.""" caplog.set_level(logging.INFO) # SpaCy on mac has issue on parallel parsing PARALLEL = 1 max_docs = 1 session = Meta.init("postgresql://localhost:5432/" + DB).Session() docs_path = "tests/data/pure_html/lincoln_short.html" # Parsing logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser(session, structural=True, lingual=True) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) docs = session.query(Document).order_by(Document.name).all() # Mention Extraction name_ngrams = MentionNgramsPart(n_max=3) place_ngrams = MentionNgramsTemp(n_max=4) Name = mention_subclass("Name") Place = mention_subclass("Place") def is_birthplace_table_row(mention): if not mention.sentence.is_tabular(): return False ngrams = get_row_ngrams(mention, lower=True) if "birth_place" in ngrams: return True else: return False birthplace_matcher = LambdaFunctionMatcher( func=is_birthplace_table_row, longest_match_only=False ) mention_extractor = MentionExtractor( session, [Name, Place], [name_ngrams, place_ngrams], [PersonMatcher(), birthplace_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) mentions = session.query(Place).all() mention_spans = [x.context.get_span() for x in mentions] assert "Sinking Spring Farm" in mention_spans assert "Farm" in mention_spans assert len(mention_spans) == 23 birthplace_matcher = LambdaFunctionMatcher( func=is_birthplace_table_row, longest_match_only=True ) mention_extractor = MentionExtractor( session, [Name, Place], [name_ngrams, place_ngrams], [PersonMatcher(), birthplace_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) mentions = session.query(Place).all() mention_spans = [x.context.get_span() for x in mentions] assert "Sinking Spring Farm" in mention_spans assert "Farm" not in mention_spans assert len(mention_spans) == 4
def test_cand_gen(caplog): """Test extracting candidates from mentions from documents.""" caplog.set_level(logging.INFO) if platform == "darwin": logger.info("Using single core.") PARALLEL = 1 else: logger.info("Using two cores.") PARALLEL = 2 # Travis only gives 2 cores def do_nothing_matcher(fig): return True max_docs = 1 session = Meta.init("postgresql://localhost:5432/" + DB).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" # Parsing logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, structural=True, lingual=True, visual=True, pdf_path=pdf_path ) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) assert session.query(Document).count() == max_docs assert session.query(Sentence).count() == 799 docs = session.query(Document).order_by(Document.name).all() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) figs = MentionFigures(types="png") Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") Fig = mention_subclass("Fig") fig_matcher = LambdaFunctionFigureMatcher(func=do_nothing_matcher) with pytest.raises(ValueError): mention_extractor = MentionExtractor( session, [Part, Temp, Volt], [part_ngrams, volt_ngrams], # Fail, mismatched arity [part_matcher, temp_matcher, volt_matcher], ) with pytest.raises(ValueError): mention_extractor = MentionExtractor( session, [Part, Temp, Volt], [part_ngrams, temp_matcher, volt_ngrams], [part_matcher, temp_matcher], # Fail, mismatched arity ) mention_extractor = MentionExtractor( session, [Part, Temp, Volt, Fig], [part_ngrams, temp_ngrams, volt_ngrams, figs], [part_matcher, temp_matcher, volt_matcher, fig_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 70 assert session.query(Volt).count() == 33 assert session.query(Temp).count() == 23 assert session.query(Fig).count() == 31 part = session.query(Part).order_by(Part.id).all()[0] volt = session.query(Volt).order_by(Volt.id).all()[0] temp = session.query(Temp).order_by(Temp.id).all()[0] logger.info(f"Part: {part.context}") logger.info(f"Volt: {volt.context}") logger.info(f"Temp: {temp.context}") # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) with pytest.raises(ValueError): candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[ temp_throttler, volt_throttler, volt_throttler, ], # Fail, mismatched arity ) with pytest.raises(ValueError): candidate_extractor = CandidateExtractor( session, [PartTemp], # Fail, mismatched arity throttlers=[temp_throttler, volt_throttler], ) # Test that no throttler in candidate extractor candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt] ) # Pass, no throttler candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 1610 assert session.query(PartVolt).count() == 2310 assert session.query(Candidate).count() == 3920 candidate_extractor.clear_all(split=0) assert session.query(Candidate).count() == 0 assert session.query(PartTemp).count() == 0 assert session.query(PartVolt).count() == 0 # Test with None in throttlers in candidate extractor candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, None] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 1432 assert session.query(PartVolt).count() == 2310 assert session.query(Candidate).count() == 3742 candidate_extractor.clear_all(split=0) assert session.query(Candidate).count() == 0 candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, volt_throttler] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 1432 assert session.query(PartVolt).count() == 1993 assert session.query(Candidate).count() == 3425 assert docs[0].name == "112823" assert len(docs[0].parts) == 70 assert len(docs[0].volts) == 33 assert len(docs[0].temps) == 23 # Test that deletion of a Candidate does not delete the Mention session.query(PartTemp).delete(synchronize_session="fetch") assert session.query(PartTemp).count() == 0 assert session.query(Temp).count() == 23 assert session.query(Part).count() == 70 # Test deletion of Candidate if Mention is deleted assert session.query(PartVolt).count() == 1993 assert session.query(Volt).count() == 33 session.query(Volt).delete(synchronize_session="fetch") assert session.query(Volt).count() == 0 assert session.query(PartVolt).count() == 0
def test_cand_gen_cascading_delete(caplog): """Test cascading the deletion of candidates.""" caplog.set_level(logging.INFO) if platform == "darwin": logger.info("Using single core.") PARALLEL = 1 else: logger.info("Using two cores.") PARALLEL = 2 # Travis only gives 2 cores max_docs = 1 session = Meta.init("postgresql://localhost:5432/" + DB).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" # Parsing logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, structural=True, lingual=True, visual=True, pdf_path=pdf_path ) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) assert session.query(Document).count() == max_docs assert session.query(Sentence).count() == 799 docs = session.query(Document).order_by(Document.name).all() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor( session, [Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher] ) mention_extractor.clear_all() mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Mention).count() == 93 assert session.query(Part).count() == 70 assert session.query(Temp).count() == 23 part = session.query(Part).order_by(Part.id).all()[0] temp = session.query(Temp).order_by(Temp.id).all()[0] logger.info(f"Part: {part.context}") logger.info(f"Temp: {temp.context}") # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor( session, [PartTemp], throttlers=[temp_throttler] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 1432 assert session.query(Candidate).count() == 1432 assert docs[0].name == "112823" assert len(docs[0].parts) == 70 assert len(docs[0].temps) == 23 # Delete from parent class should cascade to child x = session.query(Candidate).first() session.query(Candidate).filter_by(id=x.id).delete(synchronize_session="fetch") assert session.query(Candidate).count() == 1431 assert session.query(PartTemp).count() == 1431 # Clearing Mentions should also delete Candidates mention_extractor.clear() assert session.query(Mention).count() == 0 assert session.query(Part).count() == 0 assert session.query(Temp).count() == 0 assert session.query(PartTemp).count() == 0 assert session.query(Candidate).count() == 0
def test_e2e(): """Run an end-to-end test on documents of the hardware domain.""" # GitHub Actions gives 2 cores # help.github.com/en/actions/reference/virtual-environments-for-github-hosted-runners PARALLEL = 2 max_docs = 12 fonduer.init_logging( format="[%(asctime)s][%(levelname)s] %(name)s:%(lineno)s - %(message)s", level=logging.INFO, ) session = fonduer.Meta.init(CONN_STRING).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) assert session.query(Document).count() == max_docs num_docs = session.query(Document).count() logger.info(f"Docs: {num_docs}") assert num_docs == max_docs num_sentences = session.query(Sentence).count() logger.info(f"Sentences: {num_sentences}") # Divide into test and train docs = sorted(corpus_parser.get_documents()) last_docs = sorted(corpus_parser.get_last_documents()) ld = len(docs) assert ld == len(last_docs) assert len(docs[0].sentences) == len(last_docs[0].sentences) assert len(docs[0].sentences) == 799 assert len(docs[1].sentences) == 663 assert len(docs[2].sentences) == 784 assert len(docs[3].sentences) == 661 assert len(docs[4].sentences) == 513 assert len(docs[5].sentences) == 700 assert len(docs[6].sentences) == 528 assert len(docs[7].sentences) == 161 assert len(docs[8].sentences) == 228 assert len(docs[9].sentences) == 511 assert len(docs[10].sentences) == 331 assert len(docs[11].sentences) == 528 # Check table numbers assert len(docs[0].tables) == 9 assert len(docs[1].tables) == 9 assert len(docs[2].tables) == 14 assert len(docs[3].tables) == 11 assert len(docs[4].tables) == 11 assert len(docs[5].tables) == 10 assert len(docs[6].tables) == 10 assert len(docs[7].tables) == 2 assert len(docs[8].tables) == 7 assert len(docs[9].tables) == 10 assert len(docs[10].tables) == 6 assert len(docs[11].tables) == 9 # Check figure numbers assert len(docs[0].figures) == 32 assert len(docs[1].figures) == 11 assert len(docs[2].figures) == 38 assert len(docs[3].figures) == 31 assert len(docs[4].figures) == 7 assert len(docs[5].figures) == 38 assert len(docs[6].figures) == 10 assert len(docs[7].figures) == 31 assert len(docs[8].figures) == 4 assert len(docs[9].figures) == 27 assert len(docs[10].figures) == 5 assert len(docs[11].figures) == 27 # Check caption numbers assert len(docs[0].captions) == 0 assert len(docs[1].captions) == 0 assert len(docs[2].captions) == 0 assert len(docs[3].captions) == 0 assert len(docs[4].captions) == 0 assert len(docs[5].captions) == 0 assert len(docs[6].captions) == 0 assert len(docs[7].captions) == 0 assert len(docs[8].captions) == 0 assert len(docs[9].captions) == 0 assert len(docs[10].captions) == 0 assert len(docs[11].captions) == 0 train_docs = set() dev_docs = set() test_docs = set() splits = (0.5, 0.75) data = [(doc.name, doc) for doc in docs] data.sort(key=lambda x: x[0]) for i, (doc_name, doc) in enumerate(data): if i < splits[0] * ld: train_docs.add(doc) elif i < splits[1] * ld: dev_docs.add(doc) else: test_docs.add(doc) logger.info([x.name for x in train_docs]) # NOTE: With multi-relation support, return values of getting candidates, # mentions, or sparse matrices are formatted as a list of lists. This means # that with a single relation, we need to index into the list of lists to # get the candidates/mentions/sparse matrix for a particular relation or # mention. # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") mention_extractor = MentionExtractor( session, [Part, Temp, Volt], [part_ngrams, temp_ngrams, volt_ngrams], [part_matcher, temp_matcher, volt_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 299 assert session.query(Temp).count() == 138 assert session.query(Volt).count() == 140 assert len(mention_extractor.get_mentions()) == 3 assert len(mention_extractor.get_mentions()[0]) == 299 assert (len( mention_extractor.get_mentions(docs=[ session.query(Document).filter(Document.name == "112823").first() ])[0]) == 70) # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, volt_throttler]) for i, docs in enumerate([train_docs, dev_docs, test_docs]): candidate_extractor.apply(docs, split=i, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 3493 assert session.query(PartTemp).filter(PartTemp.split == 1).count() == 61 assert session.query(PartTemp).filter(PartTemp.split == 2).count() == 416 assert session.query(PartVolt).count() == 4282 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0, sort=True) dev_cands = candidate_extractor.get_candidates(split=1, sort=True) test_cands = candidate_extractor.get_candidates(split=2, sort=True) assert len(train_cands) == 2 assert len(train_cands[0]) == 3493 assert (len( candidate_extractor.get_candidates(docs=[ session.query(Document).filter(Document.name == "112823").first() ])[0]) == 1432) # Featurization featurizer = Featurizer(session, [PartTemp, PartVolt]) # Test that FeatureKey is properly reset featurizer.apply(split=1, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 214 assert session.query(FeatureKey).count() == 1260 # Test Dropping FeatureKey # Should force a row deletion featurizer.drop_keys(["DDL_e1_W_LEFT_POS_3_[NNP NN IN]"]) assert session.query(FeatureKey).count() == 1259 # Should only remove the part_volt as a relation and leave part_temp assert set( session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]").one().candidate_classes) == { "part_temp", "part_volt" } featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartVolt]) assert session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]").one().candidate_classes == ["part_temp"] assert session.query(FeatureKey).count() == 1259 # Inserting the removed key featurizer.upsert_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartTemp, PartVolt]) assert set( session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]").one().candidate_classes) == { "part_temp", "part_volt" } assert session.query(FeatureKey).count() == 1259 # Removing the key again featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartVolt]) # Removing the last relation from a key should delete the row featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartTemp]) assert session.query(FeatureKey).count() == 1258 session.query(Feature).delete(synchronize_session="fetch") session.query(FeatureKey).delete(synchronize_session="fetch") featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 6478 assert session.query(FeatureKey).count() == 4538 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (3493, 4538) assert F_train[1].shape == (2985, 4538) assert len(featurizer.get_keys()) == 4538 featurizer.apply(split=1, parallelism=PARALLEL) assert session.query(Feature).count() == 6692 assert session.query(FeatureKey).count() == 4538 F_dev = featurizer.get_feature_matrices(dev_cands) assert F_dev[0].shape == (61, 4538) assert F_dev[1].shape == (153, 4538) featurizer.apply(split=2, parallelism=PARALLEL) assert session.query(Feature).count() == 8252 assert session.query(FeatureKey).count() == 4538 F_test = featurizer.get_feature_matrices(test_cands) assert F_test[0].shape == (416, 4538) assert F_test[1].shape == (1144, 4538) gold_file = "tests/data/hardware_tutorial_gold.csv" labeler = Labeler(session, [PartTemp, PartVolt]) labeler.apply( docs=last_docs, lfs=[[gold], [gold]], table=GoldLabel, train=True, parallelism=PARALLEL, ) assert session.query(GoldLabel).count() == 8252 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] ce_v_max_lfs = [ LF_bad_keywords_in_row, LF_current_in_row, LF_non_ce_voltages_in_row, ] with pytest.raises(ValueError): labeler.apply(split=0, lfs=stg_temp_lfs, train=True, parallelism=PARALLEL) labeler.apply( docs=train_docs, lfs=[stg_temp_lfs, ce_v_max_lfs], train=True, parallelism=PARALLEL, ) assert session.query(Label).count() == 6478 assert session.query(LabelKey).count() == 9 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3493, 9) assert L_train[1].shape == (2985, 9) assert len(labeler.get_keys()) == 9 # Test Dropping LabelerKey labeler.drop_keys(["LF_storage_row"]) assert len(labeler.get_keys()) == 8 # Test Upserting LabelerKey labeler.upsert_keys(["LF_storage_row"]) assert "LF_storage_row" in [label.name for label in labeler.get_keys()] L_train_gold = labeler.get_gold_labels(train_cands) assert L_train_gold[0].shape == (3493, 1) L_train_gold = labeler.get_gold_labels(train_cands, annotator="gold") assert L_train_gold[0].shape == (3493, 1) label_model = LabelModel() label_model.fit(L_train=L_train[0], n_epochs=500, log_freq=100) train_marginals = label_model.predict_proba(L_train[0]) # Collect word counter word_counter = collect_word_counter(train_cands) emmental.init(fonduer.Meta.log_path) # Training config config = { "meta_config": { "verbose": False }, "model_config": { "model_path": None, "device": 0, "dataparallel": False }, "learner_config": { "n_epochs": 5, "optimizer_config": { "lr": 0.001, "l2": 0.0 }, "task_scheduler": "round_robin", }, "logging_config": { "evaluation_freq": 1, "counter_unit": "epoch", "checkpointing": False, "checkpointer_config": { "checkpoint_metric": { f"{ATTRIBUTE}/{ATTRIBUTE}/train/loss": "min" }, "checkpoint_freq": 1, "checkpoint_runway": 2, "clear_intermediate_checkpoints": True, "clear_all_checkpoints": True, }, }, } emmental.Meta.update_config(config=config) # Generate word embedding module arity = 2 # Geneate special tokens specials = [] for i in range(arity): specials += [f"~~[[{i}", f"{i}]]~~"] emb_layer = EmbeddingModule(word_counter=word_counter, word_dim=300, specials=specials) diffs = train_marginals.max(axis=1) - train_marginals.min(axis=1) train_idxs = np.where(diffs > 1e-6)[0] train_dataloader = EmmentalDataLoader( task_to_label_dict={ATTRIBUTE: "labels"}, dataset=FonduerDataset( ATTRIBUTE, train_cands[0], F_train[0], emb_layer.word2id, train_marginals, train_idxs, ), split="train", batch_size=100, shuffle=True, ) tasks = create_task(ATTRIBUTE, 2, F_train[0].shape[1], 2, emb_layer, model="LogisticRegression") model = EmmentalModel(name=f"{ATTRIBUTE}_task") for task in tasks: model.add_task(task) emmental_learner = EmmentalLearner() emmental_learner.learn(model, [train_dataloader]) test_dataloader = EmmentalDataLoader( task_to_label_dict={ATTRIBUTE: "labels"}, dataset=FonduerDataset(ATTRIBUTE, test_cands[0], F_test[0], emb_layer.word2id, 2), split="test", batch_size=100, shuffle=False, ) test_preds = model.predict(test_dataloader, return_preds=True) positive = np.where( np.array(test_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6) true_pred = [test_cands[0][_] for _ in positive[0]] pickle_file = "tests/data/parts_by_doc_dict.pkl" with open(pickle_file, "rb") as f: parts_by_doc = pickle.load(f) (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 < 0.7 and f1 > 0.3 stg_temp_lfs_2 = [ LF_to_left, LF_test_condition_aligned, LF_collector_aligned, LF_current_aligned, LF_voltage_row_temp, LF_voltage_row_part, LF_typ_row, LF_complement_left_row, LF_too_many_numbers_row, LF_temp_on_high_page_num, LF_temp_outside_table, LF_not_temp_relevant, ] labeler.update(split=0, lfs=[stg_temp_lfs_2, ce_v_max_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 6478 assert session.query(LabelKey).count() == 16 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3493, 16) label_model = LabelModel() label_model.fit(L_train=L_train[0], n_epochs=500, log_freq=100) train_marginals = label_model.predict_proba(L_train[0]) diffs = train_marginals.max(axis=1) - train_marginals.min(axis=1) train_idxs = np.where(diffs > 1e-6)[0] train_dataloader = EmmentalDataLoader( task_to_label_dict={ATTRIBUTE: "labels"}, dataset=FonduerDataset( ATTRIBUTE, train_cands[0], F_train[0], emb_layer.word2id, train_marginals, train_idxs, ), split="train", batch_size=100, shuffle=True, ) valid_dataloader = EmmentalDataLoader( task_to_label_dict={ATTRIBUTE: "labels"}, dataset=FonduerDataset( ATTRIBUTE, train_cands[0], F_train[0], emb_layer.word2id, np.argmax(train_marginals, axis=1), train_idxs, ), split="valid", batch_size=100, shuffle=False, ) emmental.Meta.reset() emmental.init(fonduer.Meta.log_path) emmental.Meta.update_config(config=config) tasks = create_task(ATTRIBUTE, 2, F_train[0].shape[1], 2, emb_layer, model="LogisticRegression") model = EmmentalModel(name=f"{ATTRIBUTE}_task") for task in tasks: model.add_task(task) emmental_learner = EmmentalLearner() emmental_learner.learn(model, [train_dataloader, valid_dataloader]) test_preds = model.predict(test_dataloader, return_preds=True) positive = np.where( np.array(test_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.7) true_pred = [test_cands[0][_] for _ in positive[0]] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7 # Testing LSTM emmental.Meta.reset() emmental.init(fonduer.Meta.log_path) emmental.Meta.update_config(config=config) tasks = create_task(ATTRIBUTE, 2, F_train[0].shape[1], 2, emb_layer, model="LSTM") model = EmmentalModel(name=f"{ATTRIBUTE}_task") for task in tasks: model.add_task(task) emmental_learner = EmmentalLearner() emmental_learner.learn(model, [train_dataloader]) test_preds = model.predict(test_dataloader, return_preds=True) positive = np.where( np.array(test_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.7) true_pred = [test_cands[0][_] for _ in positive[0]] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info(f"prec: {prec}") logger.info(f"rec: {rec}") logger.info(f"f1: {f1}") assert f1 > 0.7
def test_incremental(caplog): """Run an end-to-end test on incremental additions.""" caplog.set_level(logging.INFO) PARALLEL = 1 max_docs = 1 session = Meta.init("postgresql://localhost:5432/" + DB).Session() docs_path = "tests/data/html/dtc114w.html" pdf_path = "tests/data/pdf/dtc114w.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) num_docs = session.query(Document).count() logger.info(f"Docs: {num_docs}") assert num_docs == max_docs docs = corpus_parser.get_documents() last_docs = corpus_parser.get_documents() assert len(docs[0].sentences) == len(last_docs[0].sentences) # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor(session, [Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher]) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 11 assert session.query(Temp).count() == 8 # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor(session, [PartTemp], throttlers=[temp_throttler]) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 70 assert session.query(Candidate).count() == 70 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 70 # Featurization featurizer = Featurizer(session, [PartTemp]) featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 70 assert session.query(FeatureKey).count() == 512 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (70, 512) assert len(featurizer.get_keys()) == 512 # Test Dropping FeatureKey featurizer.drop_keys(["CORE_e1_LENGTH_1"]) assert session.query(FeatureKey).count() == 512 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] labeler = Labeler(session, [PartTemp]) labeler.apply(split=0, lfs=[stg_temp_lfs], train=True, parallelism=PARALLEL) assert session.query(Label).count() == 70 # Only 5 because LF_operating_row doesn't apply to the first test doc assert session.query(LabelKey).count() == 5 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (70, 5) assert len(labeler.get_keys()) == 5 docs_path = "tests/data/html/112823.html" pdf_path = "tests/data/pdf/112823.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser.apply(doc_preprocessor, pdf_path=pdf_path, clear=False) assert len(corpus_parser.get_documents()) == 2 new_docs = corpus_parser.get_last_documents() assert len(new_docs) == 1 assert new_docs[0].name == "112823" # Get mentions from just the new docs mention_extractor.apply(new_docs, parallelism=PARALLEL, clear=False) assert session.query(Part).count() == 81 assert session.query(Temp).count() == 31 # Just run candidate extraction and assign to split 0 candidate_extractor.apply(new_docs, split=0, parallelism=PARALLEL, clear=False) # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 1502 # Update features featurizer.update(new_docs, parallelism=PARALLEL) assert session.query(Feature).count() == 1502 assert session.query(FeatureKey).count() == 2573 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (1502, 2573) assert len(featurizer.get_keys()) == 2573 # Update Labels labeler.update(new_docs, lfs=[stg_temp_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 1502 assert session.query(LabelKey).count() == 6 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (1502, 6) # Test clear featurizer.clear(train=True) assert session.query(FeatureKey).count() == 0
def test_cand_gen_cascading_delete(): """Test cascading the deletion of candidates.""" # GitHub Actions gives 2 cores # help.github.com/en/actions/reference/virtual-environments-for-github-hosted-runners PARALLEL = 2 max_docs = 1 session = Meta.init(CONN_STRING).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" # Parsing logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, structural=True, lingual=True, visual=True, pdf_path=pdf_path ) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) assert session.query(Document).count() == max_docs assert session.query(Sentence).count() == 799 docs = session.query(Document).order_by(Document.name).all() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor( session, [Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher] ) mention_extractor.clear_all() mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Mention).count() == 93 assert session.query(Part).count() == 70 assert session.query(Temp).count() == 23 part = session.query(Part).order_by(Part.id).all()[0] temp = session.query(Temp).order_by(Temp.id).all()[0] logger.info(f"Part: {part.context}") logger.info(f"Temp: {temp.context}") # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor( session, [PartTemp], throttlers=[temp_throttler] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 1432 assert session.query(Candidate).count() == 1432 assert docs[0].name == "112823" assert len(docs[0].parts) == 70 assert len(docs[0].temps) == 23 # Delete from parent class should cascade to child x = session.query(Candidate).first() session.query(Candidate).filter_by(id=x.id).delete(synchronize_session="fetch") assert session.query(Candidate).count() == 1431 assert session.query(PartTemp).count() == 1431 # Test that deletion of a Candidate does not delete the Mention x = session.query(PartTemp).first() session.query(PartTemp).filter_by(id=x.id).delete(synchronize_session="fetch") assert session.query(PartTemp).count() == 1430 assert session.query(Temp).count() == 23 assert session.query(Part).count() == 70 # Clearing Mentions should also delete Candidates mention_extractor.clear() assert session.query(Mention).count() == 0 assert session.query(Part).count() == 0 assert session.query(Temp).count() == 0 assert session.query(PartTemp).count() == 0 assert session.query(Candidate).count() == 0
def test_too_many_clients_error_should_not_happen(): """Too many clients error should not happens.""" PARALLEL = 32 logger.info("Parallel: {PARALLEL}") def do_nothing_matcher(fig): return True max_docs = 1 session = Meta.init(CONN_STRING).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" # Parsing logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, structural=True, lingual=True, visual=True, pdf_path=pdf_path ) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) docs = session.query(Document).order_by(Document.name).all() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) figs = MentionFigures(types="png") Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") Fig = mention_subclass("Fig") fig_matcher = LambdaFunctionFigureMatcher(func=do_nothing_matcher) mention_extractor = MentionExtractor( session, [Part, Temp, Volt, Fig], [part_ngrams, temp_ngrams, volt_ngrams, figs], [part_matcher, temp_matcher, volt_matcher, fig_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) # Test that no throttler in candidate extractor candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt] ) # Pass, no throttler candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) candidate_extractor.clear_all(split=0) # Test with None in throttlers in candidate extractor candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, None] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL)
def test_e2e(caplog): """Run an end-to-end test on documents of the hardware domain.""" caplog.set_level(logging.INFO) PARALLEL = 4 max_docs = 12 session = Meta.init("postgresql://localhost:5432/" + DB).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) assert session.query(Document).count() == max_docs num_docs = session.query(Document).count() logger.info("Docs: {}".format(num_docs)) assert num_docs == max_docs num_sentences = session.query(Sentence).count() logger.info("Sentences: {}".format(num_sentences)) # Divide into test and train docs = corpus_parser.get_documents() ld = len(docs) assert ld == len(corpus_parser.get_last_documents()) assert len(docs[0].sentences) == 799 assert len(docs[1].sentences) == 663 assert len(docs[2].sentences) == 784 assert len(docs[3].sentences) == 661 assert len(docs[4].sentences) == 513 assert len(docs[5].sentences) == 700 assert len(docs[6].sentences) == 528 assert len(docs[7].sentences) == 161 assert len(docs[8].sentences) == 228 assert len(docs[9].sentences) == 511 assert len(docs[10].sentences) == 331 assert len(docs[11].sentences) == 528 # Check table numbers assert len(docs[0].tables) == 9 assert len(docs[1].tables) == 9 assert len(docs[2].tables) == 14 assert len(docs[3].tables) == 11 assert len(docs[4].tables) == 11 assert len(docs[5].tables) == 10 assert len(docs[6].tables) == 10 assert len(docs[7].tables) == 2 assert len(docs[8].tables) == 7 assert len(docs[9].tables) == 10 assert len(docs[10].tables) == 6 assert len(docs[11].tables) == 9 # Check figure numbers assert len(docs[0].figures) == 32 assert len(docs[1].figures) == 11 assert len(docs[2].figures) == 38 assert len(docs[3].figures) == 31 assert len(docs[4].figures) == 7 assert len(docs[5].figures) == 38 assert len(docs[6].figures) == 10 assert len(docs[7].figures) == 31 assert len(docs[8].figures) == 4 assert len(docs[9].figures) == 27 assert len(docs[10].figures) == 5 assert len(docs[11].figures) == 27 # Check caption numbers assert len(docs[0].captions) == 0 assert len(docs[1].captions) == 0 assert len(docs[2].captions) == 0 assert len(docs[3].captions) == 0 assert len(docs[4].captions) == 0 assert len(docs[5].captions) == 0 assert len(docs[6].captions) == 0 assert len(docs[7].captions) == 0 assert len(docs[8].captions) == 0 assert len(docs[9].captions) == 0 assert len(docs[10].captions) == 0 assert len(docs[11].captions) == 0 train_docs = set() dev_docs = set() test_docs = set() splits = (0.5, 0.75) data = [(doc.name, doc) for doc in docs] data.sort(key=lambda x: x[0]) for i, (doc_name, doc) in enumerate(data): if i < splits[0] * ld: train_docs.add(doc) elif i < splits[1] * ld: dev_docs.add(doc) else: test_docs.add(doc) logger.info([x.name for x in train_docs]) # NOTE: With multi-relation support, return values of getting candidates, # mentions, or sparse matrices are formatted as a list of lists. This means # that with a single relation, we need to index into the list of lists to # get the candidates/mentions/sparse matrix for a particular relation or # mention. # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") mention_extractor = MentionExtractor( session, [Part, Temp, Volt], [part_ngrams, temp_ngrams, volt_ngrams], [part_matcher, temp_matcher, volt_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 299 assert session.query(Temp).count() == 147 assert session.query(Volt).count() == 140 assert len(mention_extractor.get_mentions()) == 3 assert len(mention_extractor.get_mentions()[0]) == 299 assert (len( mention_extractor.get_mentions(docs=[ session.query(Document).filter(Document.name == "112823").first() ])[0]) == 70) # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) candidate_extractor = CandidateExtractor( session, [PartTemp, PartVolt], throttlers=[temp_throttler, volt_throttler]) for i, docs in enumerate([train_docs, dev_docs, test_docs]): candidate_extractor.apply(docs, split=i, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 3684 assert session.query(PartTemp).filter(PartTemp.split == 1).count() == 72 assert session.query(PartTemp).filter(PartTemp.split == 2).count() == 448 assert session.query(PartVolt).count() == 4282 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) dev_cands = candidate_extractor.get_candidates(split=1) test_cands = candidate_extractor.get_candidates(split=2) assert len(train_cands) == 2 assert len(train_cands[0]) == 3684 assert (len( candidate_extractor.get_candidates(docs=[ session.query(Document).filter(Document.name == "112823").first() ])[0]) == 1496) # Featurization featurizer = Featurizer(session, [PartTemp, PartVolt]) # Test that FeatureKey is properly reset featurizer.apply(split=1, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 225 assert session.query(FeatureKey).count() == 1179 # Test Dropping FeatureKey # Should force a row deletion featurizer.drop_keys(["DDL_e1_W_LEFT_POS_3_[NFP NN NFP]"]) assert session.query(FeatureKey).count() == 1178 # Should only remove the part_volt as a relation and leave part_temp assert set( session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]").one().candidate_classes) == { "part_temp", "part_volt" } featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartVolt]) assert session.query(FeatureKey).filter( FeatureKey.name == "DDL_e1_LEMMA_SEQ_[bc182]").one().candidate_classes == ["part_temp"] assert session.query(FeatureKey).count() == 1178 # Removing the last relation from a key should delete the row featurizer.drop_keys(["DDL_e1_LEMMA_SEQ_[bc182]"], candidate_classes=[PartTemp]) assert session.query(FeatureKey).count() == 1177 session.query(Feature).delete() session.query(FeatureKey).delete() featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 6669 assert session.query(FeatureKey).count() == 4161 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (3684, 4161) assert F_train[1].shape == (2985, 4161) assert len(featurizer.get_keys()) == 4161 featurizer.apply(split=1, parallelism=PARALLEL) assert session.query(Feature).count() == 6894 assert session.query(FeatureKey).count() == 4161 F_dev = featurizer.get_feature_matrices(dev_cands) assert F_dev[0].shape == (72, 4161) assert F_dev[1].shape == (153, 4161) featurizer.apply(split=2, parallelism=PARALLEL) assert session.query(Feature).count() == 8486 assert session.query(FeatureKey).count() == 4161 F_test = featurizer.get_feature_matrices(test_cands) assert F_test[0].shape == (448, 4161) assert F_test[1].shape == (1144, 4161) gold_file = "tests/data/hardware_tutorial_gold.csv" load_hardware_labels(session, PartTemp, gold_file, ATTRIBUTE, annotator_name="gold") assert session.query(GoldLabel).count() == 4204 load_hardware_labels(session, PartVolt, gold_file, ATTRIBUTE, annotator_name="gold") assert session.query(GoldLabel).count() == 8486 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] ce_v_max_lfs = [ LF_bad_keywords_in_row, LF_current_in_row, LF_non_ce_voltages_in_row, ] labeler = Labeler(session, [PartTemp, PartVolt]) with pytest.raises(ValueError): labeler.apply(split=0, lfs=stg_temp_lfs, train=True, parallelism=PARALLEL) labeler.apply(split=0, lfs=[stg_temp_lfs, ce_v_max_lfs], train=True, parallelism=PARALLEL) assert session.query(Label).count() == 6669 assert session.query(LabelKey).count() == 9 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3684, 9) assert L_train[1].shape == (2985, 9) assert len(labeler.get_keys()) == 9 L_train_gold = labeler.get_gold_labels(train_cands) assert L_train_gold[0].shape == (3684, 1) L_train_gold = labeler.get_gold_labels(train_cands, annotator="gold") assert L_train_gold[0].shape == (3684, 1) gen_model = LabelModel(k=2) gen_model.train_model(L_train[0], n_epochs=500, print_every=100) train_marginals = gen_model.predict_proba(L_train[0])[:, 1] disc_model = LogisticRegression() disc_model.train((train_cands[0], F_train[0]), train_marginals, n_epochs=20, lr=0.001) test_score = disc_model.predictions((test_cands[0], F_test[0]), b=0.6) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score > 0))] pickle_file = "tests/data/parts_by_doc_dict.pkl" with open(pickle_file, "rb") as f: parts_by_doc = pickle.load(f) (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 < 0.7 and f1 > 0.3 stg_temp_lfs_2 = [ LF_to_left, LF_test_condition_aligned, LF_collector_aligned, LF_current_aligned, LF_voltage_row_temp, LF_voltage_row_part, LF_typ_row, LF_complement_left_row, LF_too_many_numbers_row, LF_temp_on_high_page_num, LF_temp_outside_table, LF_not_temp_relevant, ] labeler.update(split=0, lfs=[stg_temp_lfs_2, ce_v_max_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 6669 assert session.query(LabelKey).count() == 16 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (3684, 16) gen_model = LabelModel(k=2) gen_model.train_model(L_train[0], n_epochs=500, print_every=100) train_marginals = gen_model.predict_proba(L_train[0])[:, 1] disc_model = LogisticRegression() disc_model.train((train_cands[0], F_train[0]), train_marginals, n_epochs=20, lr=0.001) test_score = disc_model.predictions((test_cands[0], F_test[0]), b=0.6) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score > 0))] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 > 0.7 # Testing LSTM disc_model = LSTM() disc_model.train((train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001) test_score = disc_model.predictions((test_cands[0], F_test[0]), b=0.6) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score > 0))] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 > 0.7 # Testing Sparse Logistic Regression disc_model = SparseLogisticRegression() disc_model.train((train_cands[0], F_train[0]), train_marginals, n_epochs=20, lr=0.001) test_score = disc_model.predictions((test_cands[0], F_test[0]), b=0.6) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score > 0))] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 > 0.7 # Testing Sparse LSTM disc_model = SparseLSTM() disc_model.train((train_cands[0], F_train[0]), train_marginals, n_epochs=5, lr=0.001) test_score = disc_model.predictions((test_cands[0], F_test[0]), b=0.6) true_pred = [test_cands[0][_] for _ in np.nditer(np.where(test_score > 0))] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 > 0.7
def test_cand_gen(): """Test extracting candidates from mentions from documents.""" def do_nothing_matcher(fig): return True docs_path = "tests/data/html/112823.html" pdf_path = "tests/data/pdf/112823.pdf" doc = parse_doc(docs_path, "112823", pdf_path) # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) figs = MentionFigures(types="png") Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") Fig = mention_subclass("Fig") fig_matcher = LambdaFunctionFigureMatcher(func=do_nothing_matcher) with pytest.raises(ValueError): MentionExtractor( "dummy", [Part, Temp, Volt], [part_ngrams, volt_ngrams], # Fail, mismatched arity [part_matcher, temp_matcher, volt_matcher], ) with pytest.raises(ValueError): MentionExtractor( "dummy", [Part, Temp, Volt], [part_ngrams, temp_matcher, volt_ngrams], [part_matcher, temp_matcher], # Fail, mismatched arity ) mention_extractor_udf = MentionExtractorUDF( [Part, Temp, Volt, Fig], [part_ngrams, temp_ngrams, volt_ngrams, figs], [part_matcher, temp_matcher, volt_matcher, fig_matcher], ) doc = mention_extractor_udf.apply(doc) assert len(doc.parts) == 70 assert len(doc.volts) == 33 assert len(doc.temps) == 23 assert len(doc.figs) == 31 part = doc.parts[0] volt = doc.volts[0] temp = doc.temps[0] logger.info(f"Part: {part.context}") logger.info(f"Volt: {volt.context}") logger.info(f"Temp: {temp.context}") # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) with pytest.raises(ValueError): CandidateExtractor( "dummy", [PartTemp, PartVolt], throttlers=[ temp_throttler, volt_throttler, volt_throttler, ], # Fail, mismatched arity ) with pytest.raises(ValueError): CandidateExtractor( "dummy", [PartTemp], # Fail, mismatched arity throttlers=[temp_throttler, volt_throttler], ) # Test that no throttler in candidate extractor candidate_extractor_udf = CandidateExtractorUDF( [PartTemp, PartVolt], [None, None], False, False, True # Pass, no throttler ) doc = candidate_extractor_udf.apply(doc, split=0) assert len(doc.part_temps) == 1610 assert len(doc.part_volts) == 2310 # Clear doc.part_temps = [] doc.part_volts = [] # Test with None in throttlers in candidate extractor candidate_extractor_udf = CandidateExtractorUDF( [PartTemp, PartVolt], [temp_throttler, None], False, False, True ) doc = candidate_extractor_udf.apply(doc, split=0) assert len(doc.part_temps) == 1432 assert len(doc.part_volts) == 2310 # Clear doc.part_temps = [] doc.part_volts = [] candidate_extractor_udf = CandidateExtractorUDF( [PartTemp, PartVolt], [temp_throttler, volt_throttler], False, False, True ) doc = candidate_extractor_udf.apply(doc, split=0) assert len(doc.part_temps) == 1432 assert len(doc.part_volts) == 1993 assert len(doc.parts) == 70 assert len(doc.volts) == 33 assert len(doc.temps) == 23
def test_incremental(): """Run an end-to-end test on incremental additions.""" # GitHub Actions gives 2 cores # help.github.com/en/actions/reference/virtual-environments-for-github-hosted-runners PARALLEL = 2 max_docs = 1 session = Meta.init(CONN_STRING).Session() docs_path = "tests/data/html/dtc114w.html" pdf_path = "tests/data/pdf/dtc114w.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) num_docs = session.query(Document).count() logger.info(f"Docs: {num_docs}") assert num_docs == max_docs docs = corpus_parser.get_documents() last_docs = corpus_parser.get_documents() assert len(docs[0].sentences) == len(last_docs[0].sentences) # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor(session, [Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher]) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 11 assert session.query(Temp).count() == 8 # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor(session, [PartTemp], throttlers=[temp_throttler]) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 70 assert session.query(Candidate).count() == 70 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 70 # Featurization featurizer = Featurizer(session, [PartTemp]) featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 70 assert session.query(FeatureKey).count() == 512 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (70, 512) assert len(featurizer.get_keys()) == 512 # Test Dropping FeatureKey featurizer.drop_keys(["CORE_e1_LENGTH_1"]) assert session.query(FeatureKey).count() == 512 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] labeler = Labeler(session, [PartTemp]) labeler.apply(split=0, lfs=[stg_temp_lfs], train=True, parallelism=PARALLEL) assert session.query(Label).count() == 70 # Only 5 because LF_operating_row doesn't apply to the first test doc assert session.query(LabelKey).count() == 5 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (70, 5) assert len(labeler.get_keys()) == 5 docs_path = "tests/data/html/112823.html" pdf_path = "tests/data/pdf/112823.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser.apply(doc_preprocessor, pdf_path=pdf_path, clear=False) assert len(corpus_parser.get_documents()) == 2 new_docs = corpus_parser.get_last_documents() assert len(new_docs) == 1 assert new_docs[0].name == "112823" # Get mentions from just the new docs mention_extractor.apply(new_docs, parallelism=PARALLEL, clear=False) assert session.query(Part).count() == 81 assert session.query(Temp).count() == 31 # Test if existing mentions are skipped. mention_extractor.apply(new_docs, parallelism=PARALLEL, clear=False) assert session.query(Part).count() == 81 assert session.query(Temp).count() == 31 # Just run candidate extraction and assign to split 0 candidate_extractor.apply(new_docs, split=0, parallelism=PARALLEL, clear=False) # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 1502 # Test if existing candidates are skipped. candidate_extractor.apply(new_docs, split=0, parallelism=PARALLEL, clear=False) train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 1502 # Update features featurizer.update(new_docs, parallelism=PARALLEL) assert session.query(Feature).count() == 1502 assert session.query(FeatureKey).count() == 2573 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (1502, 2573) assert len(featurizer.get_keys()) == 2573 # Update LF_storage_row. Now it always returns ABSTAIN. @labeling_function(name="LF_storage_row") def LF_storage_row_updated(c): return ABSTAIN stg_temp_lfs = [ LF_storage_row_updated, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] # Update Labels labeler.update(docs, lfs=[stg_temp_lfs], parallelism=PARALLEL) labeler.update(new_docs, lfs=[stg_temp_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 1502 # Only 5 because LF_storage_row doesn't apply to any doc (always ABSTAIN) assert session.query(LabelKey).count() == 5 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (1502, 5) # Test clear featurizer.clear(train=True) assert session.query(FeatureKey).count() == 0
def test_incremental(caplog): """Run an end-to-end test on incremental additions.""" caplog.set_level(logging.INFO) # SpaCy on mac has issue on parallel parsing if os.name == "posix": logger.info("Using single core.") PARALLEL = 1 else: PARALLEL = 2 # Travis only gives 2 cores max_docs = 1 session = Meta.init("postgres://localhost:5432/" + DB).Session() docs_path = "tests/data/html/dtc114w.html" pdf_path = "tests/data/pdf/dtc114w.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser( session, parallelism=PARALLEL, structural=True, lingual=True, visual=True, pdf_path=pdf_path, ) corpus_parser.apply(doc_preprocessor) num_docs = session.query(Document).count() logger.info("Docs: {}".format(num_docs)) assert num_docs == max_docs docs = corpus_parser.get_documents() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor( session, [Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher] ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 11 assert session.query(Temp).count() == 9 # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor( session, [PartTemp], throttlers=[temp_throttler] ) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 78 # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 78 # Featurization featurizer = Featurizer(session, [PartTemp]) featurizer.apply(split=0, train=True, parallelism=PARALLEL) assert session.query(Feature).count() == 78 assert session.query(FeatureKey).count() == 496 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (78, 496) assert len(featurizer.get_keys()) == 496 stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] labeler = Labeler(session, [PartTemp]) labeler.apply(split=0, lfs=[stg_temp_lfs], train=True, parallelism=PARALLEL) assert session.query(Label).count() == 78 # Only 5 because LF_operating_row doesn't apply to the first test doc assert session.query(LabelKey).count() == 5 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (78, 5) assert len(labeler.get_keys()) == 5 docs_path = "tests/data/html/112823.html" pdf_path = "tests/data/pdf/112823.pdf" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser.apply(doc_preprocessor, pdf_path=pdf_path, clear=False) assert len(corpus_parser.get_documents()) == 2 new_docs = corpus_parser.get_last_documents() assert len(new_docs) == 1 assert new_docs[0].name == "112823" # Get mentions from just the new docs mention_extractor.apply(new_docs, parallelism=PARALLEL, clear=False) assert session.query(Part).count() == 81 assert session.query(Temp).count() == 33 # Just run candidate extraction and assign to split 0 candidate_extractor.apply(new_docs, split=0, parallelism=PARALLEL, clear=False) # Grab candidate lists train_cands = candidate_extractor.get_candidates(split=0) assert len(train_cands) == 1 assert len(train_cands[0]) == 1574 # Update features featurizer.update(new_docs, parallelism=PARALLEL) assert session.query(Feature).count() == 1574 assert session.query(FeatureKey).count() == 2425 F_train = featurizer.get_feature_matrices(train_cands) assert F_train[0].shape == (1574, 2425) assert len(featurizer.get_keys()) == 2425 # Update Labels labeler.update(new_docs, lfs=[stg_temp_lfs], parallelism=PARALLEL) assert session.query(Label).count() == 1574 assert session.query(LabelKey).count() == 6 L_train = labeler.get_label_matrices(train_cands) assert L_train[0].shape == (1574, 6)
def test_cand_gen(caplog): """Test extracting candidates from mentions from documents.""" caplog.set_level(logging.INFO) # SpaCy on mac has issue on parallel parseing if os.name == "posix": PARALLEL = 1 else: PARALLEL = 2 # Travis only gives 2 cores max_docs = 10 session = Meta.init("postgres://localhost:5432/" + DB).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" # Parsing num_docs = session.query(Document).count() if num_docs != max_docs: logger.info("Parsing...") doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) corpus_parser = Parser(structural=True, lingual=True, visual=True, pdf_path=pdf_path) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) assert session.query(Document).count() == max_docs assert session.query(Sentence).count() == 5892 docs = session.query(Document).order_by(Document.name).all() # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) volt_ngrams = MentionNgramsVolt(n_max=1) Part = mention_subclass("Part") Temp = mention_subclass("Temp") Volt = mention_subclass("Volt") with pytest.raises(ValueError): mention_extractor = MentionExtractor( [Part, Temp, Volt], [part_ngrams, volt_ngrams], # Fail, mismatched arity [part_matcher, temp_matcher, volt_matcher], ) with pytest.raises(ValueError): mention_extractor = MentionExtractor( [Part, Temp, Volt], [part_ngrams, temp_matcher, volt_ngrams], [part_matcher, temp_matcher], # Fail, mismatched arity ) mention_extractor = MentionExtractor( [Part, Temp, Volt], [part_ngrams, temp_ngrams, volt_ngrams], [part_matcher, temp_matcher, volt_matcher], ) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 234 assert session.query(Volt).count() == 108 assert session.query(Temp).count() == 118 part = session.query(Part).order_by(Part.id).all()[0] volt = session.query(Volt).order_by(Volt.id).all()[0] temp = session.query(Temp).order_by(Temp.id).all()[0] logger.info("Part: {}".format(part.span)) logger.info("Volt: {}".format(volt.span)) logger.info("Temp: {}".format(temp.span)) # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) PartVolt = candidate_subclass("PartVolt", [Part, Volt]) candidate_extractor = CandidateExtractor( [PartTemp, PartVolt], throttlers=[temp_throttler, volt_throttler]) candidate_extractor.apply(docs, split=0, parallelism=PARALLEL) assert session.query(PartTemp).count() == 3385 assert session.query(PartVolt).count() == 3364 assert session.query(Candidate).count() == 6749 assert docs[0].name == "112823" assert len(docs[0].parts) == 70 assert len(docs[0].volts) == 33 assert len(docs[0].temps) == 18 # Test that deletion of a Candidate does not delete the Mention session.query(PartTemp).delete() assert session.query(PartTemp).count() == 0 assert session.query(Temp).count() == 118 assert session.query(Part).count() == 234 # Test deletion of Candidate if Mention is deleted assert session.query(PartVolt).count() == 3364 assert session.query(Volt).count() == 108 session.query(Volt).delete() assert session.query(Volt).count() == 0 assert session.query(PartVolt).count() == 0
def test_e2e_logistic_regression(caplog): """Run an end-to-end test on documents of the hardware domain.""" caplog.set_level(logging.INFO) # SpaCy on mac has issue on parallel parseing if os.name == "posix": PARALLEL = 1 else: PARALLEL = 2 # Travis only gives 2 cores max_docs = 12 session = Meta.init("postgres://localhost:5432/" + DB).Session() docs_path = "tests/data/html/" pdf_path = "tests/data/pdf/" doc_preprocessor = HTMLDocPreprocessor(docs_path, max_docs=max_docs) num_docs = session.query(Document).count() if num_docs != max_docs: logger.info("Parsing...") corpus_parser = Parser(structural=True, lingual=True, visual=True, pdf_path=pdf_path) corpus_parser.apply(doc_preprocessor, parallelism=PARALLEL) assert session.query(Document).count() == max_docs num_docs = session.query(Document).count() logger.info("Docs: {}".format(num_docs)) assert num_docs == max_docs num_sentences = session.query(Sentence).count() logger.info("Sentences: {}".format(num_sentences)) # Divide into test and train docs = session.query(Document).order_by(Document.name).all() ld = len(docs) assert len(docs[0].sentences) == 828 assert len(docs[1].sentences) == 706 assert len(docs[2].sentences) == 819 assert len(docs[3].sentences) == 684 assert len(docs[4].sentences) == 552 assert len(docs[5].sentences) == 758 assert len(docs[6].sentences) == 597 assert len(docs[7].sentences) == 165 assert len(docs[8].sentences) == 250 assert len(docs[9].sentences) == 533 assert len(docs[10].sentences) == 354 assert len(docs[11].sentences) == 547 # Check table numbers assert len(docs[0].tables) == 9 assert len(docs[1].tables) == 9 assert len(docs[2].tables) == 14 assert len(docs[3].tables) == 11 assert len(docs[4].tables) == 11 assert len(docs[5].tables) == 10 assert len(docs[6].tables) == 10 assert len(docs[7].tables) == 2 assert len(docs[8].tables) == 7 assert len(docs[9].tables) == 10 assert len(docs[10].tables) == 6 assert len(docs[11].tables) == 9 # Check figure numbers assert len(docs[0].figures) == 32 assert len(docs[1].figures) == 11 assert len(docs[2].figures) == 38 assert len(docs[3].figures) == 31 assert len(docs[4].figures) == 7 assert len(docs[5].figures) == 38 assert len(docs[6].figures) == 10 assert len(docs[7].figures) == 31 assert len(docs[8].figures) == 4 assert len(docs[9].figures) == 27 assert len(docs[10].figures) == 5 assert len(docs[11].figures) == 27 # Check caption numbers assert len(docs[0].captions) == 0 assert len(docs[1].captions) == 0 assert len(docs[2].captions) == 0 assert len(docs[3].captions) == 0 assert len(docs[4].captions) == 0 assert len(docs[5].captions) == 0 assert len(docs[6].captions) == 0 assert len(docs[7].captions) == 0 assert len(docs[8].captions) == 0 assert len(docs[9].captions) == 0 assert len(docs[10].captions) == 0 assert len(docs[11].captions) == 0 train_docs = set() dev_docs = set() test_docs = set() splits = (0.5, 0.75) data = [(doc.name, doc) for doc in docs] data.sort(key=lambda x: x[0]) for i, (doc_name, doc) in enumerate(data): if i < splits[0] * ld: train_docs.add(doc) elif i < splits[1] * ld: dev_docs.add(doc) else: test_docs.add(doc) logger.info([x.name for x in train_docs]) # Mention Extraction part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3) temp_ngrams = MentionNgramsTemp(n_max=2) Part = mention_subclass("Part") Temp = mention_subclass("Temp") mention_extractor = MentionExtractor([Part, Temp], [part_ngrams, temp_ngrams], [part_matcher, temp_matcher]) mention_extractor.apply(docs, parallelism=PARALLEL) assert session.query(Part).count() == 299 assert session.query(Temp).count() == 127 # Candidate Extraction PartTemp = candidate_subclass("PartTemp", [Part, Temp]) candidate_extractor = CandidateExtractor([PartTemp], throttlers=[temp_throttler]) for i, docs in enumerate([train_docs, dev_docs, test_docs]): candidate_extractor.apply(docs, split=i, parallelism=PARALLEL) assert session.query(PartTemp).filter(PartTemp.split == 0).count() == 3201 assert session.query(PartTemp).filter(PartTemp.split == 1).count() == 61 assert session.query(PartTemp).filter(PartTemp.split == 2).count() == 420 train_cands = session.query(PartTemp).filter(PartTemp.split == 0).all() featurizer = FeatureAnnotator(PartTemp) F_train = featurizer.apply(split=0, replace_key_set=True, parallelism=PARALLEL) logger.info(F_train.shape) F_dev = featurizer.apply(split=1, replace_key_set=False, parallelism=PARALLEL) logger.info(F_dev.shape) F_test = featurizer.apply(split=2, replace_key_set=False, parallelism=PARALLEL) logger.info(F_test.shape) gold_file = "tests/data/hardware_tutorial_gold.csv" load_hardware_labels(session, PartTemp, gold_file, ATTRIBUTE, annotator_name="gold") stg_temp_lfs = [ LF_storage_row, LF_operating_row, LF_temperature_row, LF_tstg_row, LF_to_left, LF_negative_number_left, ] labeler = LabelAnnotator(PartTemp, lfs=stg_temp_lfs) L_train = labeler.apply(split=0, clear=True, parallelism=PARALLEL) logger.info(L_train.shape) load_gold_labels(session, annotator_name="gold", split=0) gen_model = GenerativeModel() gen_model.train(L_train, epochs=500, decay=0.9, step_size=0.001 / L_train.shape[0], reg_param=0) logger.info("LF Accuracy: {}".format(gen_model.weights.lf_accuracy)) load_gold_labels(session, annotator_name="gold", split=1) train_marginals = gen_model.marginals(L_train) disc_model = LogisticRegression() disc_model.train((train_cands, F_train), train_marginals, n_epochs=200, lr=0.001) load_gold_labels(session, annotator_name="gold", split=2) test_candidates = [ F_test.get_candidate(session, i) for i in range(F_test.shape[0]) ] test_score = disc_model.predictions((test_candidates, F_test)) true_pred = [ test_candidates[_] for _ in np.nditer(np.where(test_score > 0)) ] pickle_file = "tests/data/parts_by_doc_dict.pkl" with open(pickle_file, "rb") as f: parts_by_doc = pickle.load(f) (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 < 0.7 and f1 > 0.3 stg_temp_lfs_2 = [ LF_test_condition_aligned, LF_collector_aligned, LF_current_aligned, LF_voltage_row_temp, LF_voltage_row_part, LF_typ_row, LF_complement_left_row, LF_too_many_numbers_row, LF_temp_on_high_page_num, LF_temp_outside_table, LF_not_temp_relevant, ] labeler = LabelAnnotator(PartTemp, lfs=stg_temp_lfs_2) L_train = labeler.apply(split=0, clear=False, update_keys=True, update_values=True, parallelism=PARALLEL) gen_model = GenerativeModel() gen_model.train(L_train, epochs=500, decay=0.9, step_size=0.001 / L_train.shape[0], reg_param=0) train_marginals = gen_model.marginals(L_train) disc_model = LogisticRegression() disc_model.train((train_cands, F_train), train_marginals, n_epochs=200, lr=0.001) test_score = disc_model.predictions((test_candidates, F_test)) true_pred = [ test_candidates[_] for _ in np.nditer(np.where(test_score > 0)) ] (TP, FP, FN) = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, parts_by_doc=parts_by_doc) tp_len = len(TP) fp_len = len(FP) fn_len = len(FN) prec = tp_len / (tp_len + fp_len) if tp_len + fp_len > 0 else float("nan") rec = tp_len / (tp_len + fn_len) if tp_len + fn_len > 0 else float("nan") f1 = 2 * (prec * rec) / (prec + rec) if prec + rec > 0 else float("nan") logger.info("prec: {}".format(prec)) logger.info("rec: {}".format(rec)) logger.info("f1: {}".format(f1)) assert f1 > 0.7