def _classify(self, doc: Document) -> DataFrame:
# Only one candidate class is defined.
candidate_class = self.candidate_extractor.candidate_classes[0]
test_cands = getattr(doc, candidate_class.__tablename__ + "s")
if self.model_type == "emmental":
# Featurization
features_list = self.featurizer.apply(doc)
# Convert features into a sparse matrix
F_test = _F_matrix(features_list[0], self.key_names)
# Dataloader for test
ATTRIBUTE = "wiki"
test_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(ATTRIBUTE, test_cands, F_test,
self.word2id, 2),
split="test",
batch_size=100,
shuffle=False,
)
test_preds = self.emmental_model.predict(test_dataloader,
return_preds=True)
positive = np.where(
np.array(test_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6)
true_preds = [test_cands[_] for _ in positive[0]]
else:
labels_list = self.labeler.apply(doc, lfs=self.lfs)
L_test = _L_matrix(labels_list[0], self.key_names)
marginals = self.label_models[0].predict_proba(L_test)
for cand, prob in zip(test_cands, marginals[:, 1]):
cand.prob = prob
true_preds = sorted(test_cands,
key=lambda cand: cand.prob,
reverse=True)
df = DataFrame()
for entity_relation in get_unique_entity_relations(true_preds):
df = df.append(
DataFrame(
[entity_relation],
columns=[m.__name__ for m in candidate_class.mentions]))
return df
def _classify(self, doc: Document) -> DataFrame:
# Only one candidate class is used.
candidate_class = self.candidate_extractor.candidate_classes[0]
test_cands = getattr(doc, candidate_class.__tablename__ + "s")
features_list = self.featurizer.apply(doc)
# Convert features into a sparse matrix
F_test = FonduerModel.convert_features_to_matrix(
features_list[0], self.key_names)
test_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(ATTRIBUTE, test_cands, F_test, self.word2id,
2),
split="test",
batch_size=100,
shuffle=False,
)
test_preds = self.emmental_model.predict(test_dataloader,
return_preds=True)
positive = np.where(
np.array(test_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.7)
true_preds = [test_cands[_] 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)
df = DataFrame()
for c in true_preds:
part = c[0].context.get_span()
doc = c[0].context.sentence.document.name.upper()
val = c[1].context.get_span()
for p in get_implied_parts(part, doc, parts_by_doc):
entity_relation = (doc, p, val)
df = df.append(
DataFrame([entity_relation],
columns=["doc", "part", "val"]))
return df
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 train_model(cands, F, align_type, model_type="LogisticRegression"):
# Extract candidates and features based on the align type (row/column)
align_val = 0 if align_type == "row" else 1
train_cands = cands[align_val][0]
F_train = F[align_val][0]
train_marginals = np.array([[0, 1] if gold[align_val](x) else [1, 0]
for x in train_cands[0]])
# 1.) Setup training config
config = {
"meta_config": {
"verbose": True
},
"model_config": {
"model_path": None,
"device": 0,
"dataparallel": False
},
"learner_config": {
"n_epochs": 50,
"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.init(Meta.log_path)
emmental.Meta.update_config(config=config)
# 2.) Collect word counter from training data
word_counter = collect_word_counter(train_cands)
# 3.) Generate word embedding module for LSTM model
# (in Logistic Regression, we generate it since Fonduer dataset requires word2id dict)
# Geneate special tokens
arity = 2
specials = []
for i in range(arity):
specials += [f"~~[[{i}", f"{i}]]~~"]
emb_layer = EmbeddingModule(word_counter=word_counter,
word_dim=300,
specials=specials)
# 4.) Generate dataloader for training set
# No noise in Gold labels
train_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(
ATTRIBUTE,
train_cands[0],
F_train[0],
emb_layer.word2id,
train_marginals,
),
split="train",
batch_size=100,
shuffle=True,
)
# 5.) Training
tasks = create_task(
ATTRIBUTE,
2,
F_train[0].shape[1],
2,
emb_layer,
model=model_type # "LSTM"
)
model = EmmentalModel(name=f"{ATTRIBUTE}_task")
for task in tasks:
model.add_task(task)
emmental_learner = EmmentalLearner()
emmental_learner.learn(model, [train_dataloader])
return (model, emb_layer)
def eval_model(model, emb_layer, cands, F, align_type="row"):
# Extract candidates and features based on the align type (row/column)
align_val = 0 if align_type == "row" else 1
train_cands = cands[align_val][0]
dev_cands = cands[align_val][1]
test_cands = cands[align_val][2]
F_train = F[align_val][0]
F_dev = F[align_val][1]
F_test = F[align_val][2]
row_on = True if align_type == "row" else False
col_on = True if align_type == "col" else False
# Generate dataloader for test data
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]]
test_results = entity_level_f1(true_pred,
gold_file,
ATTRIBUTE,
test_docs,
row_on=row_on,
col_on=col_on)
# Run on dev and train set for validation
# We run the predictions also on our training and dev set, to validate that everything seems to work smoothly
# Generate dataloader for dev data
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(ATTRIBUTE, dev_cands[0], F_dev[0],
emb_layer.word2id, 2),
split="test",
batch_size=100,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
positive_dev = np.where(
np.array(dev_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6)
true_dev_pred = [dev_cands[0][_] for _ in positive_dev[0]]
dev_results = entity_level_f1(true_dev_pred,
gold_file,
ATTRIBUTE,
dev_docs,
row_on=row_on,
col_on=col_on)
# Generate dataloader for train data
train_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(ATTRIBUTE, train_cands[0], F_train[0],
emb_layer.word2id, 2),
split="test",
batch_size=100,
shuffle=False,
)
train_preds = model.predict(train_dataloader, return_preds=True)
positive_train = np.where(
np.array(train_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6)
true_train_pred = [train_cands[0][_] for _ in positive_train[0]]
train_results = entity_level_f1(true_train_pred,
gold_file,
ATTRIBUTE,
train_docs,
row_on=row_on,
col_on=col_on)
return [train_results, dev_results, test_results]
def main(
conn_string,
gain=False,
current=False,
max_docs=float("inf"),
parse=False,
first_time=False,
re_label=False,
parallel=8,
log_dir="logs",
verbose=False,
):
# Setup initial configuration
if not log_dir:
log_dir = "logs"
if verbose:
level = logging.INFO
else:
level = logging.WARNING
dirname = os.path.dirname(os.path.abspath(__file__))
init_logging(log_dir=os.path.join(dirname, log_dir), level=level)
rel_list = []
if gain:
rel_list.append("gain")
if current:
rel_list.append("current")
logger.info(f"=" * 30)
logger.info(f"Running with parallel: {parallel}, max_docs: {max_docs}")
session = Meta.init(conn_string).Session()
# Parsing
start = timer()
logger.info(f"Starting parsing...")
docs, train_docs, dev_docs, test_docs = parse_dataset(session,
dirname,
first_time=parse,
parallel=parallel,
max_docs=max_docs)
logger.debug(f"Done")
end = timer()
logger.warning(f"Parse Time (min): {((end - start) / 60.0):.1f}")
logger.info(f"# of Documents: {len(docs)}")
logger.info(f"# of train Documents: {len(train_docs)}")
logger.info(f"# of dev Documents: {len(dev_docs)}")
logger.info(f"# of test Documents: {len(test_docs)}")
logger.info(f"Documents: {session.query(Document).count()}")
logger.info(f"Sections: {session.query(Section).count()}")
logger.info(f"Paragraphs: {session.query(Paragraph).count()}")
logger.info(f"Sentences: {session.query(Sentence).count()}")
logger.info(f"Figures: {session.query(Figure).count()}")
# Mention Extraction
start = timer()
mentions = []
ngrams = []
matchers = []
# Only do those that are enabled
if gain:
Gain = mention_subclass("Gain")
gain_matcher = get_gain_matcher()
gain_ngrams = MentionNgrams(n_max=2)
mentions.append(Gain)
ngrams.append(gain_ngrams)
matchers.append(gain_matcher)
if current:
Current = mention_subclass("SupplyCurrent")
current_matcher = get_supply_current_matcher()
current_ngrams = MentionNgramsCurrent(n_max=3)
mentions.append(Current)
ngrams.append(current_ngrams)
matchers.append(current_matcher)
mention_extractor = MentionExtractor(session, mentions, ngrams, matchers)
if first_time:
mention_extractor.apply(docs, parallelism=parallel)
logger.info(f"Total Mentions: {session.query(Mention).count()}")
if gain:
logger.info(f"Total Gain: {session.query(Gain).count()}")
if current:
logger.info(f"Total Current: {session.query(Current).count()}")
cand_classes = []
if gain:
GainCand = candidate_subclass("GainCand", [Gain])
cand_classes.append(GainCand)
if current:
CurrentCand = candidate_subclass("CurrentCand", [Current])
cand_classes.append(CurrentCand)
candidate_extractor = CandidateExtractor(session, cand_classes)
if first_time:
for i, docs in enumerate([train_docs, dev_docs, test_docs]):
candidate_extractor.apply(docs, split=i, parallelism=parallel)
# These must be sorted for deterministic behavior.
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)
logger.info(
f"Total train candidate: {len(train_cands[0]) + len(train_cands[1])}")
logger.info(
f"Total dev candidate: {len(dev_cands[0]) + len(dev_cands[1])}")
logger.info(
f"Total test candidate: {len(test_cands[0]) + len(test_cands[1])}")
logger.info("Done w/ candidate extraction.")
end = timer()
logger.warning(f"CE Time (min): {((end - start) / 60.0):.1f}")
# First, check total recall
# result = entity_level_scores(
# candidates_to_entities(dev_cands[0], is_gain=True),
# corpus=dev_docs,
# is_gain=True,
# )
# logger.info(f"Gain Total Dev Recall: {result.rec:.3f}")
# logger.info(f"\n{pformat(result.FN)}")
# result = entity_level_scores(
# candidates_to_entities(test_cands[0], is_gain=True),
# corpus=test_docs,
# is_gain=True,
# )
# logger.info(f"Gain Total Test Recall: {result.rec:.3f}")
# logger.info(f"\n{pformat(result.FN)}")
#
# result = entity_level_scores(
# candidates_to_entities(dev_cands[1], is_gain=False),
# corpus=dev_docs,
# is_gain=False,
# )
# logger.info(f"Current Total Dev Recall: {result.rec:.3f}")
# logger.info(f"\n{pformat(result.FN)}")
# result = entity_level_scores(
# candidates_to_entities(test_cands[1], is_gain=False),
# corpus=test_docs,
# is_gain=False,
# )
# logger.info(f"Current Test Recall: {result.rec:.3f}")
# logger.info(f"\n{pformat(result.FN)}")
start = timer()
# Using parallelism = 1 for deterministic behavior.
featurizer = Featurizer(session, cand_classes, parallelism=1)
if first_time:
logger.info("Starting featurizer...")
# Set feature space based on dev set, which we use for training rather
# than the large train set.
featurizer.apply(split=1, train=True)
featurizer.apply(split=0)
featurizer.apply(split=2)
logger.info("Done")
logger.info("Getting feature matrices...")
# Serialize feature matrices on first run
if first_time:
F_train = featurizer.get_feature_matrices(train_cands)
F_dev = featurizer.get_feature_matrices(dev_cands)
F_test = featurizer.get_feature_matrices(test_cands)
end = timer()
logger.warning(
f"Featurization Time (min): {((end - start) / 60.0):.1f}")
F_train_dict = {}
F_dev_dict = {}
F_test_dict = {}
for idx, relation in enumerate(rel_list):
F_train_dict[relation] = F_train[idx]
F_dev_dict[relation] = F_dev[idx]
F_test_dict[relation] = F_test[idx]
pickle.dump(F_train_dict,
open(os.path.join(dirname, "F_train_dict.pkl"), "wb"))
pickle.dump(F_dev_dict,
open(os.path.join(dirname, "F_dev_dict.pkl"), "wb"))
pickle.dump(F_test_dict,
open(os.path.join(dirname, "F_test_dict.pkl"), "wb"))
else:
F_train_dict = pickle.load(
open(os.path.join(dirname, "F_train_dict.pkl"), "rb"))
F_dev_dict = pickle.load(
open(os.path.join(dirname, "F_dev_dict.pkl"), "rb"))
F_test_dict = pickle.load(
open(os.path.join(dirname, "F_test_dict.pkl"), "rb"))
F_train = []
F_dev = []
F_test = []
for relation in rel_list:
F_train.append(F_train_dict[relation])
F_dev.append(F_dev_dict[relation])
F_test.append(F_test_dict[relation])
logger.info("Done.")
start = timer()
logger.info("Labeling training data...")
# labeler = Labeler(session, cand_classes)
# lfs = []
# if gain:
# lfs.append(gain_lfs)
#
# if current:
# lfs.append(current_lfs)
#
# if first_time:
# logger.info("Applying LFs...")
# labeler.apply(split=0, lfs=lfs, train=True, parallelism=parallel)
# elif re_label:
# logger.info("Re-applying LFs...")
# labeler.update(split=0, lfs=lfs, parallelism=parallel)
#
# logger.info("Done...")
# logger.info("Getting label matrices...")
# L_train = labeler.get_label_matrices(train_cands)
# logger.info("Done...")
if first_time:
marginals_dict = {}
for idx, relation in enumerate(rel_list):
# Manually create marginals from human annotations
marginal = []
dev_gold_entities = get_gold_set(is_gain=(relation == "gain"))
for c in dev_cands[idx]:
flag = False
for entity in cand_to_entity(c, is_gain=(relation == "gain")):
if entity in dev_gold_entities:
flag = True
if flag:
marginal.append([0.0, 1.0])
else:
marginal.append([1.0, 0.0])
marginals_dict[relation] = np.array(marginal)
pickle.dump(marginals_dict,
open(os.path.join(dirname, "marginals_dict.pkl"), "wb"))
else:
marginals_dict = pickle.load(
open(os.path.join(dirname, "marginals_dict.pkl"), "rb"))
marginals = []
for relation in rel_list:
marginals.append(marginals_dict[relation])
end = timer()
logger.warning(
f"Weak Supervision Time (min): {((end - start) / 60.0):.1f}")
start = timer()
word_counter = collect_word_counter(train_cands)
# Training config
config = {
"meta_config": {
"verbose": True,
"seed": 30
},
"model_config": {
"model_path": None,
"device": 0,
"dataparallel": False
},
"learner_config": {
"n_epochs": 500,
"optimizer_config": {
"lr": 0.001,
"l2": 0.005
},
"task_scheduler": "round_robin",
},
"logging_config": {
"evaluation_freq": 1,
"counter_unit": "epoch",
"checkpointing": False,
"checkpointer_config": {
"checkpoint_metric": {
"model/all/train/loss": "min"
},
"checkpoint_freq": 1,
"checkpoint_runway": 2,
"clear_intermediate_checkpoints": True,
"clear_all_checkpoints": True,
},
},
}
emmental.init(log_dir=Meta.log_path, 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)
train_idxs = []
train_dataloader = []
for idx, relation in enumerate(rel_list):
diffs = marginals[idx].max(axis=1) - marginals[idx].min(axis=1)
train_idxs.append(np.where(diffs > 1e-6)[0])
# only uses dev set as training data, with human annotations
train_dataloader.append(
EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(
relation,
dev_cands[idx],
F_dev[idx],
emb_layer.word2id,
marginals[idx],
train_idxs[idx],
),
split="train",
batch_size=256,
shuffle=True,
))
num_feature_keys = len(featurizer.get_keys())
model = EmmentalModel(name=f"opamp_tasks")
# List relation names, arities, list of classes
tasks = create_task(
rel_list,
[2] * len(rel_list),
num_feature_keys,
[2] * len(rel_list),
emb_layer,
model="LogisticRegression",
)
for task in tasks:
model.add_task(task)
emmental_learner = EmmentalLearner()
# If given a list of multi, will train on multiple
emmental_learner.learn(model, train_dataloader)
# List of dataloader for each relation
for idx, relation in enumerate(rel_list):
test_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, test_cands[idx], F_test[idx],
emb_layer.word2id, 2),
split="test",
batch_size=256,
shuffle=False,
)
test_preds = model.predict(test_dataloader, return_preds=True)
best_result, best_b = scoring(
test_preds,
test_cands[idx],
test_docs,
is_gain=(relation == "gain"),
num=100,
)
# Dump CSV files for analysis
if relation == "gain":
train_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, train_cands[idx],
F_train[idx], emb_layer.word2id, 2),
split="train",
batch_size=256,
shuffle=False,
)
train_preds = model.predict(train_dataloader, return_preds=True)
Y_prob = np.array(train_preds["probs"][relation])[:, TRUE]
output_csv(train_cands[idx], Y_prob, is_gain=True)
Y_prob = np.array(test_preds["probs"][relation])[:, TRUE]
output_csv(test_cands[idx], Y_prob, is_gain=True, append=True)
dump_candidates(test_cands[idx],
Y_prob,
"gain_test_probs.csv",
is_gain=True)
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, dev_cands[idx], F_dev[idx],
emb_layer.word2id, 2),
split="dev",
batch_size=256,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
Y_prob = np.array(dev_preds["probs"][relation])[:, TRUE]
output_csv(dev_cands[idx], Y_prob, is_gain=True, append=True)
dump_candidates(dev_cands[idx],
Y_prob,
"gain_dev_probs.csv",
is_gain=True)
if relation == "current":
train_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, train_cands[idx],
F_train[idx], emb_layer.word2id, 2),
split="train",
batch_size=256,
shuffle=False,
)
train_preds = model.predict(train_dataloader, return_preds=True)
Y_prob = np.array(train_preds["probs"][relation])[:, TRUE]
output_csv(train_cands[idx], Y_prob, is_gain=False)
Y_prob = np.array(test_preds["probs"][relation])[:, TRUE]
output_csv(test_cands[idx], Y_prob, is_gain=False, append=True)
dump_candidates(test_cands[idx],
Y_prob,
"current_test_probs.csv",
is_gain=False)
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, dev_cands[idx], F_dev[idx],
emb_layer.word2id, 2),
split="dev",
batch_size=256,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
Y_prob = np.array(dev_preds["probs"][relation])[:, TRUE]
output_csv(dev_cands[idx], Y_prob, is_gain=False, append=True)
dump_candidates(dev_cands[idx],
Y_prob,
"current_dev_probs.csv",
is_gain=False)
end = timer()
logger.warning(
f"Classification AND dump data Time (min): {((end - start) / 60.0):.1f}"
)
def main(
conn_string,
stg_temp_min=False,
stg_temp_max=False,
polarity=False,
ce_v_max=False,
max_docs=float("inf"),
parse=False,
first_time=False,
re_label=False,
parallel=4,
log_dir=None,
verbose=False,
):
if not log_dir:
log_dir = "logs"
if verbose:
level = logging.INFO
else:
level = logging.WARNING
dirname = os.path.dirname(os.path.abspath(__file__))
init_logging(log_dir=os.path.join(dirname, log_dir), level=level)
rel_list = []
if stg_temp_min:
rel_list.append("stg_temp_min")
if stg_temp_max:
rel_list.append("stg_temp_max")
if polarity:
rel_list.append("polarity")
if ce_v_max:
rel_list.append("ce_v_max")
session = Meta.init(conn_string).Session()
# Parsing
logger.info(f"Starting parsing...")
start = timer()
docs, train_docs, dev_docs, test_docs = parse_dataset(session,
dirname,
first_time=parse,
parallel=parallel,
max_docs=max_docs)
end = timer()
logger.warning(f"Parse Time (min): {((end - start) / 60.0):.1f}")
logger.info(f"# of train Documents: {len(train_docs)}")
logger.info(f"# of dev Documents: {len(dev_docs)}")
logger.info(f"# of test Documents: {len(test_docs)}")
logger.info(f"Documents: {session.query(Document).count()}")
logger.info(f"Sections: {session.query(Section).count()}")
logger.info(f"Paragraphs: {session.query(Paragraph).count()}")
logger.info(f"Sentences: {session.query(Sentence).count()}")
logger.info(f"Figures: {session.query(Figure).count()}")
# Mention Extraction
start = timer()
mentions = []
ngrams = []
matchers = []
# Only do those that are enabled
Part = mention_subclass("Part")
part_matcher = get_matcher("part")
part_ngrams = MentionNgramsPart(parts_by_doc=None, n_max=3)
mentions.append(Part)
ngrams.append(part_ngrams)
matchers.append(part_matcher)
if stg_temp_min:
StgTempMin = mention_subclass("StgTempMin")
stg_temp_min_matcher = get_matcher("stg_temp_min")
stg_temp_min_ngrams = MentionNgramsTemp(n_max=2)
mentions.append(StgTempMin)
ngrams.append(stg_temp_min_ngrams)
matchers.append(stg_temp_min_matcher)
if stg_temp_max:
StgTempMax = mention_subclass("StgTempMax")
stg_temp_max_matcher = get_matcher("stg_temp_max")
stg_temp_max_ngrams = MentionNgramsTemp(n_max=2)
mentions.append(StgTempMax)
ngrams.append(stg_temp_max_ngrams)
matchers.append(stg_temp_max_matcher)
if polarity:
Polarity = mention_subclass("Polarity")
polarity_matcher = get_matcher("polarity")
polarity_ngrams = MentionNgrams(n_max=1)
mentions.append(Polarity)
ngrams.append(polarity_ngrams)
matchers.append(polarity_matcher)
if ce_v_max:
CeVMax = mention_subclass("CeVMax")
ce_v_max_matcher = get_matcher("ce_v_max")
ce_v_max_ngrams = MentionNgramsVolt(n_max=1)
mentions.append(CeVMax)
ngrams.append(ce_v_max_ngrams)
matchers.append(ce_v_max_matcher)
mention_extractor = MentionExtractor(session, mentions, ngrams, matchers)
if first_time:
mention_extractor.apply(docs, parallelism=parallel)
logger.info(f"Total Mentions: {session.query(Mention).count()}")
logger.info(f"Total Part: {session.query(Part).count()}")
if stg_temp_min:
logger.info(f"Total StgTempMin: {session.query(StgTempMin).count()}")
if stg_temp_max:
logger.info(f"Total StgTempMax: {session.query(StgTempMax).count()}")
if polarity:
logger.info(f"Total Polarity: {session.query(Polarity).count()}")
if ce_v_max:
logger.info(f"Total CeVMax: {session.query(CeVMax).count()}")
# Candidate Extraction
cands = []
throttlers = []
if stg_temp_min:
PartStgTempMin = candidate_subclass("PartStgTempMin",
[Part, StgTempMin])
stg_temp_min_throttler = stg_temp_filter
cands.append(PartStgTempMin)
throttlers.append(stg_temp_min_throttler)
if stg_temp_max:
PartStgTempMax = candidate_subclass("PartStgTempMax",
[Part, StgTempMax])
stg_temp_max_throttler = stg_temp_filter
cands.append(PartStgTempMax)
throttlers.append(stg_temp_max_throttler)
if polarity:
PartPolarity = candidate_subclass("PartPolarity", [Part, Polarity])
polarity_throttler = polarity_filter
cands.append(PartPolarity)
throttlers.append(polarity_throttler)
if ce_v_max:
PartCeVMax = candidate_subclass("PartCeVMax", [Part, CeVMax])
ce_v_max_throttler = ce_v_max_filter
cands.append(PartCeVMax)
throttlers.append(ce_v_max_throttler)
candidate_extractor = CandidateExtractor(session,
cands,
throttlers=throttlers)
if first_time:
for i, docs in enumerate([train_docs, dev_docs, test_docs]):
candidate_extractor.apply(docs, split=i, parallelism=parallel)
num_cands = session.query(Candidate).filter(
Candidate.split == i).count()
logger.info(f"Candidates in split={i}: {num_cands}")
# These must be sorted for deterministic behavior.
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)
end = timer()
logger.warning(
f"Candidate Extraction Time (min): {((end - start) / 60.0):.1f}")
logger.info(f"Total train candidate: {sum(len(_) for _ in train_cands)}")
logger.info(f"Total dev candidate: {sum(len(_) for _ in dev_cands)}")
logger.info(f"Total test candidate: {sum(len(_) for _ in test_cands)}")
pickle_file = os.path.join(dirname, "data/parts_by_doc_new.pkl")
with open(pickle_file, "rb") as f:
parts_by_doc = pickle.load(f)
# Check total recall
for i, name in enumerate(rel_list):
logger.info(name)
result = entity_level_scores(
candidates_to_entities(dev_cands[i], parts_by_doc=parts_by_doc),
attribute=name,
corpus=dev_docs,
)
logger.info(f"{name} Total Dev Recall: {result.rec:.3f}")
result = entity_level_scores(
candidates_to_entities(test_cands[i], parts_by_doc=parts_by_doc),
attribute=name,
corpus=test_docs,
)
logger.info(f"{name} Total Test Recall: {result.rec:.3f}")
# Featurization
start = timer()
cands = []
if stg_temp_min:
cands.append(PartStgTempMin)
if stg_temp_max:
cands.append(PartStgTempMax)
if polarity:
cands.append(PartPolarity)
if ce_v_max:
cands.append(PartCeVMax)
# Using parallelism = 1 for deterministic behavior.
featurizer = Featurizer(session, cands, parallelism=1)
if first_time:
logger.info("Starting featurizer...")
featurizer.apply(split=0, train=True)
featurizer.apply(split=1)
featurizer.apply(split=2)
logger.info("Done")
logger.info("Getting feature matrices...")
if first_time:
F_train = featurizer.get_feature_matrices(train_cands)
F_dev = featurizer.get_feature_matrices(dev_cands)
F_test = featurizer.get_feature_matrices(test_cands)
end = timer()
logger.warning(
f"Featurization Time (min): {((end - start) / 60.0):.1f}")
F_train_dict = {}
F_dev_dict = {}
F_test_dict = {}
for idx, relation in enumerate(rel_list):
F_train_dict[relation] = F_train[idx]
F_dev_dict[relation] = F_dev[idx]
F_test_dict[relation] = F_test[idx]
pickle.dump(F_train_dict,
open(os.path.join(dirname, "F_train_dict.pkl"), "wb"))
pickle.dump(F_dev_dict,
open(os.path.join(dirname, "F_dev_dict.pkl"), "wb"))
pickle.dump(F_test_dict,
open(os.path.join(dirname, "F_test_dict.pkl"), "wb"))
else:
F_train_dict = pickle.load(
open(os.path.join(dirname, "F_train_dict.pkl"), "rb"))
F_dev_dict = pickle.load(
open(os.path.join(dirname, "F_dev_dict.pkl"), "rb"))
F_test_dict = pickle.load(
open(os.path.join(dirname, "F_test_dict.pkl"), "rb"))
F_train = []
F_dev = []
F_test = []
for relation in rel_list:
F_train.append(F_train_dict[relation])
F_dev.append(F_dev_dict[relation])
F_test.append(F_test_dict[relation])
logger.info("Done.")
for i, cand in enumerate(cands):
logger.info(f"{cand} Train shape: {F_train[i].shape}")
logger.info(f"{cand} Test shape: {F_test[i].shape}")
logger.info(f"{cand} Dev shape: {F_dev[i].shape}")
logger.info("Labeling training data...")
# Labeling
start = timer()
lfs = []
if stg_temp_min:
lfs.append(stg_temp_min_lfs)
if stg_temp_max:
lfs.append(stg_temp_max_lfs)
if polarity:
lfs.append(polarity_lfs)
if ce_v_max:
lfs.append(ce_v_max_lfs)
# Using parallelism = 1 for deterministic behavior.
labeler = Labeler(session, cands, parallelism=1)
if first_time:
logger.info("Applying LFs...")
labeler.apply(split=0, lfs=lfs, train=True)
logger.info("Done...")
# Uncomment if debugging LFs
# load_transistor_labels(session, cands, ["ce_v_max"])
# labeler.apply(split=1, lfs=lfs, train=False, parallelism=parallel)
# labeler.apply(split=2, lfs=lfs, train=False, parallelism=parallel)
elif re_label:
logger.info("Updating LFs...")
labeler.update(split=0, lfs=lfs)
logger.info("Done...")
# Uncomment if debugging LFs
# labeler.apply(split=1, lfs=lfs, train=False, parallelism=parallel)
# labeler.apply(split=2, lfs=lfs, train=False, parallelism=parallel)
logger.info("Getting label matrices...")
L_train = labeler.get_label_matrices(train_cands)
# Uncomment if debugging LFs
# L_dev = labeler.get_label_matrices(dev_cands)
# L_dev_gold = labeler.get_gold_labels(dev_cands, annotator="gold")
#
# L_test = labeler.get_label_matrices(test_cands)
# L_test_gold = labeler.get_gold_labels(test_cands, annotator="gold")
logger.info("Done.")
if first_time:
marginals_dict = {}
for idx, relation in enumerate(rel_list):
marginals_dict[relation] = generative_model(L_train[idx])
pickle.dump(marginals_dict,
open(os.path.join(dirname, "marginals_dict.pkl"), "wb"))
else:
marginals_dict = pickle.load(
open(os.path.join(dirname, "marginals_dict.pkl"), "rb"))
marginals = []
for relation in rel_list:
marginals.append(marginals_dict[relation])
end = timer()
logger.warning(f"Supervision Time (min): {((end - start) / 60.0):.1f}")
start = timer()
word_counter = collect_word_counter(train_cands)
# Training config
config = {
"meta_config": {
"verbose": True,
"seed": 17
},
"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": {
"model/all/train/loss": "min"
},
"checkpoint_freq": 1,
"checkpoint_runway": 2,
"clear_intermediate_checkpoints": True,
"clear_all_checkpoints": True,
},
},
}
emmental.init(log_dir=Meta.log_path, 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)
train_idxs = []
train_dataloader = []
for idx, relation in enumerate(rel_list):
diffs = marginals[idx].max(axis=1) - marginals[idx].min(axis=1)
train_idxs.append(np.where(diffs > 1e-6)[0])
train_dataloader.append(
EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(
relation,
train_cands[idx],
F_train[idx],
emb_layer.word2id,
marginals[idx],
train_idxs[idx],
),
split="train",
batch_size=100,
shuffle=True,
))
num_feature_keys = len(featurizer.get_keys())
model = EmmentalModel(name=f"transistor_tasks")
# List relation names, arities, list of classes
tasks = create_task(
rel_list,
[2] * len(rel_list),
num_feature_keys,
[2] * len(rel_list),
emb_layer,
model="LogisticRegression",
)
for task in tasks:
model.add_task(task)
emmental_learner = EmmentalLearner()
# If given a list of multi, will train on multiple
emmental_learner.learn(model, train_dataloader)
# List of dataloader for each rlation
for idx, relation in enumerate(rel_list):
test_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, test_cands[idx], F_test[idx],
emb_layer.word2id, 2),
split="test",
batch_size=100,
shuffle=False,
)
test_preds = model.predict(test_dataloader, return_preds=True)
best_result, best_b = scoring(
relation,
test_preds,
test_cands[idx],
test_docs,
F_test[idx],
parts_by_doc,
num=100,
)
# Dump CSV files for CE_V_MAX for digi-key analysis
if relation == "ce_v_max":
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, dev_cands[idx], F_dev[idx],
emb_layer.word2id, 2),
split="dev",
batch_size=100,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
Y_prob = np.array(test_preds["probs"][relation])[:, TRUE]
dump_candidates(test_cands[idx], Y_prob, "ce_v_max_test_probs.csv")
Y_prob = np.array(dev_preds["probs"][relation])[:, TRUE]
dump_candidates(dev_cands[idx], Y_prob, "ce_v_max_dev_probs.csv")
# Dump CSV files for POLARITY for digi-key analysis
if relation == "polarity":
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={relation: "labels"},
dataset=FonduerDataset(relation, dev_cands[idx], F_dev[idx],
emb_layer.word2id, 2),
split="dev",
batch_size=100,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
Y_prob = np.array(test_preds["probs"][relation])[:, TRUE]
dump_candidates(test_cands[idx], Y_prob, "polarity_test_probs.csv")
Y_prob = np.array(dev_preds["probs"][relation])[:, TRUE]
dump_candidates(dev_cands[idx], Y_prob, "polarity_dev_probs.csv")
end = timer()
logger.warning(f"Classification Time (min): {((end - start) / 60.0):.1f}")
def eval_model(model, emb_layer, cands, F, schema_filter=False):
# Extract candidates and features
train_cands = cands[0]
dev_cands = cands[1]
test_cands = cands[2]
F_train = F[0]
F_dev = F[1]
F_test = F[2]
# apply schema filter
def apply(cands):
return schema_match_filter(
cands,
"station",
"price",
price_col_keywords,
stations_mapping_dict,
0.05,
DEBUG,
)
# Generate dataloader for test data
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]]
true_pred = apply(true_pred) if schema_filter else true_pred
test_results = entity_level_f1(true_pred, gold_file, ATTRIBUTE, test_docs, stations_mapping_dict=stations_mapping_dict)
# Run on dev and train set for validation
# We run the predictions also on our training and dev set, to validate that everything seems to work smoothly
# Generate dataloader for dev data
dev_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(
ATTRIBUTE, dev_cands[0], F_dev[0], emb_layer.word2id, 2
),
split="test",
batch_size=100,
shuffle=False,
)
dev_preds = model.predict(dev_dataloader, return_preds=True)
positive_dev = np.where(np.array(dev_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6)
true_dev_pred = [dev_cands[0][_] for _ in positive_dev[0]]
true_dev_pred = apply(true_dev_pred) if schema_filter else true_dev_pred
dev_results = entity_level_f1(true_dev_pred, gold_file, ATTRIBUTE, dev_docs, stations_mapping_dict=stations_mapping_dict)
# Generate dataloader for train data
train_dataloader = EmmentalDataLoader(
task_to_label_dict={ATTRIBUTE: "labels"},
dataset=FonduerDataset(
ATTRIBUTE, train_cands[0], F_train[0], emb_layer.word2id, 2
),
split="test",
batch_size=100,
shuffle=False,
)
train_preds = model.predict(train_dataloader, return_preds=True)
positive_train = np.where(np.array(train_preds["probs"][ATTRIBUTE])[:, TRUE] > 0.6)
true_train_pred = [train_cands[0][_] for _ in positive_train[0]]
true_train_pred = apply(true_train_pred) if schema_filter else true_train_pred
train_results = entity_level_f1(true_train_pred, gold_file, ATTRIBUTE, train_docs, stations_mapping_dict=stations_mapping_dict)
return [train_results, dev_results, test_results]
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"
)
emmental_model = EmmentalModel()
for task in tasks: