def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(
os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
if hasattr(Config, 'is_snopes'):
is_snopes = Config.is_snopes
else:
is_snopes = False
logger.debug("is_snopes: " + str(is_snopes))
if mode == RTERunPhase.train:
# training mode
if hasattr(Config, 'training_dump') and os.path.exists(
Config.training_dump):
with open(Config.training_dump, 'rb') as f:
(X_train, Y_labels_train, X_valid,
Y_labels_valid) = pickle.load(f)
else:
# process training JSONL file
X_train, Y_labels_train = read_data_set_from_jsonl(
Config.training_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
X_valid, Y_labels_valid = read_data_set_from_jsonl(
Config.dev_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
b_train = X_train['b']
X_train['b_sizes'] = get_num_sents_of_bodies(b_train)
for i, sample in enumerate(b_train):
if len(sample) < Config.max_sentences:
for _ in range(Config.max_sentences - len(sample)):
sample.append(" ")
b_train[i] = np.asarray(sample)
b_train = np.asarray(b_train)
X_train['b'] = b_train
logger.debug("b_train.shape: " + str(b_train.shape))
b_valid = X_valid['b']
X_valid['b_sizes'] = get_num_sents_of_bodies(b_valid)
for i, sample in enumerate(b_valid):
if len(sample) < Config.max_sentences:
for _ in range(Config.max_sentences - len(sample)):
sample.append(" ")
b_valid[i] = np.asarray(sample)
b_valid = np.asarray(b_valid)
X_valid['b'] = b_valid
logger.debug("b_valid.shape: " + str(b_valid.shape))
if hasattr(Config, 'training_dump'):
with open(Config.training_dump, 'wb') as f:
pickle.dump(
(X_train, Y_labels_train, X_valid, Y_labels_valid),
f,
protocol=pickle.HIGHEST_PROTOCOL)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
estimator.fit(X_train, Y_labels_train, X_valid, Y_labels_valid)
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
X_test, Y_labels_test = read_data_set_from_jsonl(
Config.test_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
b_test = X_test['b']
X_test['b_sizes'] = get_num_sents_of_bodies(b_test)
for i, sample in enumerate(b_test):
if len(sample) < Config.max_sentences:
for _ in range(Config.max_sentences - len(sample)):
sample.append(" ")
b_test[i] = np.asarray(sample)
b_test = np.asarray(b_test)
X_test['b'] = b_test
logger.debug("b_test.shape: " + str(b_test.shape))
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
predictions = estimator.predict(X_test, restore_param_required)
generate_submission(predictions, X_test['id'], Config.test_set_file,
Config.submission_file)
if Y_labels_test:
print_metrics(Y_labels_test, predictions, logger)
return estimator
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(
os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
if hasattr(Config, 'use_inter_evidence_comparison'):
use_inter_evidence_comparison = Config.use_inter_evidence_comparison
else:
use_inter_evidence_comparison = False
# 'esim_inter_evidence' model and 'esim_inter_evidence_claim_evidences_comparison' models need inter evidence inputs
use_inter_evidence_comparison = use_inter_evidence_comparison or Config.estimator_name in {
'esim_inter_evidence', 'esim_inter_evidence_claim_evidences_comparison'
}
if hasattr(Config, 'use_claim_evidences_comparison'):
use_claim_evidences_comparison = Config.use_claim_evidences_comparison
else:
use_claim_evidences_comparison = False
# 'esim_inter_evidence_claim_evidences_comparison' model needs claim-evidence inputs
use_claim_evidences_comparison = use_claim_evidences_comparison or Config.estimator_name in {
'esim_inter_evidence_claim_evidences_comparison'
}
if hasattr(Config, 'use_extra_features'):
use_extra_features = Config.use_extra_features
else:
use_extra_features = False
if hasattr(Config, 'use_numeric_feature'):
use_numeric_feature = Config.use_numeric_feature
else:
use_numeric_feature = False
# 'esim_num_feature' model needs numeric feature inputs
use_numeric_feature = use_numeric_feature or Config.estimator_name in {
'esim_num_feature'
}
if hasattr(Config, 'is_snopes'):
is_snopes = Config.is_snopes
else:
is_snopes = False
logger.debug("is_snopes: " + str(is_snopes))
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("ESIM arguments: " + str(Config.esim_end_2_end_hyper_param))
logger.info("use_inter_sentence_comparison: " +
str(use_inter_evidence_comparison))
logger.info("use_extra_features: " + str(use_extra_features))
logger.info("use_numeric_feature: " + str(use_numeric_feature))
logger.info("use_claim_evidences_comparison: " +
str(use_claim_evidences_comparison))
if mode == RTERunPhase.train:
# # training mode
if hasattr(Config, 'training_dump') and os.path.exists(
Config.training_dump):
with open(Config.training_dump, 'rb') as f:
(X_dict, y_train) = pickle.load(f)
else:
training_set, vocab, embeddings, _, _ = embed_data_set_with_glove_2(
Config.training_set_file,
Config.db_path,
glove_path=Config.glove_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=is_snopes)
h_sent_sizes = training_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
training_set['data']['h_sent_sizes'] = np.expand_dims(
h_sent_sizes, 1)
training_set['data']['h_sizes'] = h_sizes
training_set['data']['h_np'] = np.expand_dims(
training_set['data']['h_np'], 1)
valid_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.dev_set_file,
Config.db_path,
vocab_dict=vocab,
glove_embeddings=embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=is_snopes)
h_sent_sizes = valid_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
valid_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
valid_set['data']['h_sizes'] = h_sizes
valid_set['data']['h_np'] = np.expand_dims(
valid_set['data']['h_np'], 1)
if use_extra_features:
assert hasattr(
Config, 'feature_path'
), "Config should has feature_path if Config.use_feature is True"
training_claim_features, training_evidence_features = load_feature_by_data_set(
Config.training_set_file, Config.feature_path,
Config.max_sentences)
valid_claim_features, valid_evidence_features = load_feature_by_data_set(
Config.dev_set_file, Config.feature_path,
Config.max_sentences)
training_set['data']['h_feats'] = training_claim_features
training_set['data']['b_feats'] = training_evidence_features
valid_set['data']['h_feats'] = valid_claim_features
valid_set['data']['b_feats'] = valid_evidence_features
if use_numeric_feature:
training_num_feat = number_feature(Config.training_set_file,
Config.db_path,
Config.max_sentences,
is_snopes)
valid_num_feat = number_feature(Config.dev_set_file,
Config.db_path,
Config.max_sentences,
is_snopes)
training_set['data']['num_feat'] = training_num_feat
valid_set['data']['num_feat'] = valid_num_feat
if use_inter_evidence_comparison:
training_concat_sent_indices, training_concat_sent_sizes = generate_concat_indices_for_inter_evidence(
training_set['data']['b_np'],
training_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
training_set['data'][
'b_concat_indices'] = training_concat_sent_indices
training_set['data'][
'b_concat_sizes'] = training_concat_sent_sizes
valid_concat_sent_indices, valid_concat_sent_sizes = generate_concat_indices_for_inter_evidence(
valid_set['data']['b_np'],
valid_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
valid_set['data'][
'b_concat_indices'] = valid_concat_sent_indices
valid_set['data']['b_concat_sizes'] = valid_concat_sent_sizes
if use_claim_evidences_comparison:
training_all_evidences_indices, training_all_evidences_sizes = generate_concat_indices_for_claim(
training_set['data']['b_np'],
training_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
training_set['data'][
'b_concat_indices_for_h'] = training_all_evidences_indices
training_set['data'][
'b_concat_sizes_for_h'] = training_all_evidences_sizes
valid_all_evidences_indices, valid_all_evidences_sizes = generate_concat_indices_for_claim(
valid_set['data']['b_np'],
valid_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
valid_set['data'][
'b_concat_indices_for_h'] = valid_all_evidences_indices
valid_set['data'][
'b_concat_sizes_for_h'] = valid_all_evidences_sizes
X_dict = {
'X_train': training_set['data'],
'X_valid': valid_set['data'],
'y_valid': valid_set['label'],
'embedding': embeddings
}
y_train = training_set['label']
if hasattr(Config, 'training_dump'):
with open(Config.training_dump, 'wb') as f:
pickle.dump((X_dict, y_train),
f,
protocol=pickle.HIGHEST_PROTOCOL)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
estimator.fit(X_dict, y_train)
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
vocab, embeddings = load_whole_glove(Config.glove_path)
vocab = vocab_map(vocab)
test_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.test_set_file,
Config.db_path,
vocab_dict=vocab,
glove_embeddings=embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=is_snopes)
h_sent_sizes = test_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
test_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
test_set['data']['h_sizes'] = h_sizes
test_set['data']['h_np'] = np.expand_dims(test_set['data']['h_np'], 1)
if use_extra_features:
assert hasattr(
Config, 'feature_path'
), "Config should has feature_path if Config.use_feature is True"
test_claim_features, test_evidence_features = load_feature_by_data_set(
Config.test_set_file, Config.feature_path,
Config.max_sentences)
test_set['data']['h_feats'] = test_claim_features
test_set['data']['b_feats'] = test_evidence_features
if use_numeric_feature:
test_num_feat = number_feature(Config.test_set_file,
Config.db_path,
Config.max_sentences, is_snopes)
test_set['data']['num_feat'] = test_num_feat
x_dict = {'X_test': test_set['data'], 'embedding': embeddings}
if use_inter_evidence_comparison:
test_concat_sent_indices, test_concat_sent_sizes = generate_concat_indices_for_inter_evidence(
test_set['data']['b_np'], test_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
test_set['data']['b_concat_indices'] = test_concat_sent_indices
test_set['data']['b_concat_sizes'] = test_concat_sent_sizes
if use_claim_evidences_comparison:
test_all_evidences_indices, test_all_evidences_sizes = generate_concat_indices_for_claim(
test_set['data']['b_np'], test_set['data']['b_sent_sizes'],
Config.max_sentence_size, Config.max_sentences)
test_set['data'][
'b_concat_indices_for_h'] = test_all_evidences_indices
test_set['data']['b_concat_sizes_for_h'] = test_all_evidences_sizes
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
predictions = estimator.predict(
x_dict, restore_param_required=restore_param_required)
generate_submission(predictions, test_set['id'], Config.test_set_file,
Config.submission_file)
if 'label' in test_set:
print_metrics(test_set['label'], predictions, logger)
return estimator
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(
os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("ESIM arguments: " + str(Config.esim_hyper_param))
logger.info("this script is only for FEVER dataset")
if mode == RTERunPhase.train:
# # training mode
if hasattr(Config, 'training_dump') and os.path.exists(
Config.training_dump):
with open(Config.training_dump, 'rb') as f:
(X_dict, y_train) = pickle.load(f)
else:
training_set, vocab, embeddings, _, _ = embed_data_set_with_glove_2(
Config.training_set_file,
Config.db_path,
glove_path=Config.glove_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size)
h_sent_sizes = training_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
training_set['data']['h_sent_sizes'] = np.expand_dims(
h_sent_sizes, 1)
training_set['data']['h_sizes'] = h_sizes
training_set['data']['h_np'] = np.expand_dims(
training_set['data']['h_np'], 1)
training_set['data']['scores'] = load_scores(
Config.training_set_file, Config.max_sentences)
valid_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.dev_set_file,
Config.db_path,
vocab_dict=vocab,
glove_embeddings=embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size)
h_sent_sizes = valid_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
valid_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
valid_set['data']['h_sizes'] = h_sizes
valid_set['data']['h_np'] = np.expand_dims(
valid_set['data']['h_np'], 1)
valid_set['data']['scores'] = load_scores(Config.dev_set_file,
Config.max_sentences)
X_dict = {
'X_train': training_set['data'],
'X_valid': valid_set['data'],
'y_valid': valid_set['label'],
'embedding': embeddings
}
y_train = training_set['label']
if hasattr(Config, 'training_dump'):
with open(Config.training_dump, 'wb') as f:
pickle.dump((X_dict, y_train),
f,
protocol=pickle.HIGHEST_PROTOCOL)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
estimator.fit(X_dict, y_train)
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
vocab, embeddings = load_whole_glove(Config.glove_path)
vocab = vocab_map(vocab)
test_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.test_set_file,
Config.db_path,
vocab_dict=vocab,
glove_embeddings=embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size)
h_sent_sizes = test_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
test_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
test_set['data']['h_sizes'] = h_sizes
test_set['data']['h_np'] = np.expand_dims(test_set['data']['h_np'], 1)
test_set['data']['scores'] = load_scores(Config.test_set_file,
Config.max_sentences)
x_dict = {'X_test': test_set['data'], 'embedding': embeddings}
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
predictions = estimator.predict(
x_dict, restore_param_required=restore_param_required)
generate_submission(predictions, test_set['id'], Config.test_set_file,
Config.submission_file)
if 'label' in test_set:
print_metrics(test_set['label'], predictions, logger)
return estimator
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(
os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("ESIM arguments: " + str(Config.esim_hyper_param))
if hasattr(Config, 'is_snopes'):
is_snopes = Config.is_snopes
else:
is_snopes = False
logger.debug("is_snopes: " + str(is_snopes))
if mode == RTERunPhase.train:
# training mode
if hasattr(Config, 'training_dump') and os.path.exists(
Config.training_dump):
with open(Config.training_dump, 'rb') as f:
dataset_list = pickle.load(f)
else:
# process training JSONL file
training_set, _, _ = embed_data_set_for_elmo(
Config.training_set_file,
Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_h_sent_size=Config.max_claim_size,
threshold_b_sent_size=Config.max_sentence_size,
is_snopes=is_snopes)
h_sent_sizes = training_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
training_set['data']['h_sent_sizes'] = np.expand_dims(
h_sent_sizes, 1)
training_set['data']['h_sizes'] = h_sizes
training_set['data']['h_tokens'] = np.expand_dims(
training_set['data']['h_tokens'], 1)
# training_set['data']['h_ft_np'] = np.expand_dims(training_set['data']['h_ft_np'], 1)
valid_set, _, _ = embed_data_set_for_elmo(
Config.dev_set_file,
Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=is_snopes)
h_sent_sizes = valid_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
valid_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
valid_set['data']['h_sizes'] = h_sizes
valid_set['data']['h_tokens'] = np.expand_dims(
valid_set['data']['h_tokens'], 1)
dataset_list = [training_set, valid_set]
# save processed training data
if hasattr(Config, 'training_dump'):
with open(Config.training_dump, 'wb') as f:
pickle.dump(dataset_list,
f,
protocol=pickle.HIGHEST_PROTOCOL)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
estimator.fit(dataset_list[0]['data'], dataset_list[0]['label'],
dataset_list[1]['data'], dataset_list[1]['label'])
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
test_set, _, _ = embed_data_set_for_elmo(
Config.test_set_file,
Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=is_snopes)
h_sent_sizes = test_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
test_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
test_set['data']['h_sizes'] = h_sizes
test_set['data']['h_tokens'] = np.expand_dims(
test_set['data']['h_tokens'], 1)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
logger.debug("CUDA_VISIBLE_DEVICES: " +
os.environ['CUDA_VISIBLE_DEVICES'])
predictions = estimator.predict(
test_set['data'], restore_param_required=restore_param_required)
generate_submission(predictions, test_set['id'], Config.test_set_file,
Config.submission_file)
if 'label' in test_set:
print_metrics(test_set['label'], predictions, logger)
return estimator
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(
os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
assert hasattr(
Config, 'page_source_file_path'
), "'page_source_file_path' field is needed in config file for this script"
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("ESIM credibility MTL arguments: " +
str(Config.esim_credibility_mtl_hyper_param))
logger.info("this script is only for Snopes dataset")
if mode == RTERunPhase.train:
# training sets
# @formatter:off
claim_training_set, word_vocab, word_embeddings, domain_vocab, domain_embeddings, suffix_vocab, \
suffix_embeddings, protocol_vocab, protocol_embeddings, claim_stance_vocab, claim_stance_embeddings = \
embed_data_set_with_glove_with_credibility(
Config.esim_credibility_mtl_hyper_param['claim_training_set'],
Config.db_path,
Config.page_source_file_path,
glove_path=Config.glove_path,
domain_embedding_size=Config.esim_credibility_mtl_hyper_param['domain_embedding_size'],
suffix_embedding_size=Config.esim_credibility_mtl_hyper_param['suffix_embedding_size'],
protocol_embedding_size=Config.esim_credibility_mtl_hyper_param['protocol_embedding_size'],
stance_embedding_size=Config.esim_credibility_mtl_hyper_param['stance_embedding_size'],
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_sentence_size)
# @formatter:on
claim_h_sent_sizes = claim_training_set['data']['h_sent_sizes']
claim_h_sizes = np.ones(len(claim_h_sent_sizes), np.int32)
claim_training_set['data']['h_sent_sizes'] = np.expand_dims(
claim_h_sent_sizes, 1)
claim_training_set['data']['h_sizes'] = claim_h_sizes
claim_training_set['data']['h_np'] = np.expand_dims(
claim_training_set['data']['h_np'], 1)
logger.info("size of training set: " +
str(claim_training_set['data']['h_np'].shape[0]))
stance_training_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.esim_credibility_mtl_hyper_param['stance_training_set'],
Config.db_path,
Config.glove_path,
vocab_dict=word_vocab,
glove_embeddings=word_embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=True)
stance_h_sent_sizes = stance_training_set['data']['h_sent_sizes']
stance_h_sizes = np.ones(len(stance_h_sent_sizes), np.int32)
stance_training_set['data']['h_sent_sizes'] = np.expand_dims(
stance_h_sent_sizes, 1)
stance_training_set['data']['h_sizes'] = stance_h_sizes
stance_training_set['data']['h_np'] = np.expand_dims(
stance_training_set['data']['h_np'], 1)
# valid sets
claim_valid_set, _, _ = embed_data_set_with_glove_with_credibility(
Config.esim_credibility_mtl_hyper_param['claim_dev_set'],
Config.db_path,
Config.page_source_file_path,
vocab_dict=word_vocab,
glove_embeddings=word_embeddings,
domain_vocab=domain_vocab,
domain_embeddings=domain_embeddings,
suffix_vocab=suffix_vocab,
suffix_embeddings=suffix_embeddings,
protocol_vocab=protocol_vocab,
protocol_embeddings=protocol_embeddings,
stance_vocab=claim_stance_vocab,
stance_embeddings=claim_stance_embeddings,
domain_embedding_size=Config.
esim_credibility_mtl_hyper_param['domain_embedding_size'],
suffix_embedding_size=Config.
esim_credibility_mtl_hyper_param['suffix_embedding_size'],
protocol_embedding_size=Config.
esim_credibility_mtl_hyper_param['protocol_embedding_size'],
stance_embedding_size=Config.
esim_credibility_mtl_hyper_param['stance_embedding_size'],
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_sentence_size)
claim_h_sent_sizes = claim_valid_set['data']['h_sent_sizes']
claim_h_sizes = np.ones(len(claim_h_sent_sizes), np.int32)
claim_valid_set['data']['h_sent_sizes'] = np.expand_dims(
claim_h_sent_sizes, 1)
claim_valid_set['data']['h_sizes'] = claim_h_sizes
claim_valid_set['data']['h_np'] = np.expand_dims(
claim_valid_set['data']['h_np'], 1)
logger.info("size of dev set: " +
str(claim_valid_set['data']['h_np'].shape[0]))
stance_valid_set, _, _, _, _ = embed_data_set_with_glove_2(
Config.esim_credibility_mtl_hyper_param['stance_dev_set'],
Config.db_path,
Config.glove_path,
vocab_dict=word_vocab,
glove_embeddings=word_embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_claim_size,
is_snopes=True)
stance_h_sent_sizes = stance_valid_set['data']['h_sent_sizes']
stance_h_sizes = np.ones(len(stance_h_sent_sizes), np.int32)
stance_valid_set['data']['h_sent_sizes'] = np.expand_dims(
stance_h_sent_sizes, 1)
stance_valid_set['data']['h_sizes'] = stance_h_sizes
stance_valid_set['data']['h_np'] = np.expand_dims(
stance_valid_set['data']['h_np'], 1)
X_dict_claim = {
'train': claim_training_set['data'],
'valid': claim_valid_set['data'],
}
y_claim = {
'train': claim_training_set['label'],
'valid': claim_valid_set['label']
}
X_dict_stance = {
'train': stance_training_set['data'],
'valid': stance_valid_set['data'],
}
y_stance = {
'train': stance_training_set['label'],
'valid': stance_valid_set['label']
}
X_dict = {
'claim': X_dict_claim,
'stance': X_dict_stance,
'word_embedding': word_embeddings,
'domain_embedding': domain_embeddings,
'suffix_embedding': suffix_embeddings,
'protocol_embedding': protocol_embeddings,
'stance_embedding': claim_stance_embeddings
}
y_dict = {'claim': y_claim, 'stance': y_stance}
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
estimator.fit(X_dict, y_dict)
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
dump_source_features_embeddings(
Config.
esim_credibility_mtl_hyper_param['features_embeddings_path'],
domain_vocab, domain_embeddings, suffix_vocab, suffix_embeddings,
protocol_vocab, protocol_embeddings, claim_stance_vocab,
claim_stance_embeddings)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name,
Config.ckpt_folder)
word_vocab, word_embeddings = load_whole_glove(Config.glove_path)
word_vocab = vocab_map(word_vocab)
# @formatter:off
domain_vocab, domain_embeddings, \
suffix_vocab, suffix_embeddings, \
protocol_vocab, protocol_embeddings, \
claim_stance_vocab, claim_stance_embeddings = load_source_features_embeddings(
Config.esim_credibility_mtl_hyper_param['features_embeddings_path'])
# @formatter:on
test_set, _, _ = embed_data_set_with_glove_with_credibility(
Config.esim_credibility_mtl_hyper_param['claim_test_set'],
Config.db_path,
Config.page_source_file_path,
vocab_dict=word_vocab,
glove_embeddings=word_embeddings,
domain_vocab=domain_vocab,
domain_embeddings=domain_embeddings,
suffix_vocab=suffix_vocab,
suffix_embeddings=suffix_embeddings,
protocol_vocab=protocol_vocab,
protocol_embeddings=protocol_embeddings,
stance_vocab=claim_stance_vocab,
stance_embeddings=claim_stance_embeddings,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
threshold_h_sent_size=Config.max_sentence_size)
claim_h_sent_sizes = test_set['data']['h_sent_sizes']
claim_h_sizes = np.ones(len(claim_h_sent_sizes), np.int32)
test_set['data']['h_sent_sizes'] = np.expand_dims(
claim_h_sent_sizes, 1)
test_set['data']['h_sizes'] = claim_h_sizes
test_set['data']['h_np'] = np.expand_dims(test_set['data']['h_np'], 1)
logger.info("size of test set: " +
str(test_set['data']['h_np'].shape[0]))
x_dict = {
'X_test': test_set['data'],
'word_embedding': word_embeddings,
'domain_embedding': domain_embeddings,
'suffix_embedding': suffix_embeddings,
'protocol_embedding': protocol_embeddings,
'stance_embedding': claim_stance_embeddings
}
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(
os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0,
maxMemory=1.0 -
Config.max_gpu_memory)[0])
predictions = estimator.predict(x_dict, restore_param_required)
generate_submission(predictions, test_set['id'], Config.test_set_file,
Config.submission_file)
if 'label' in test_set:
print_metrics(test_set['label'], predictions, logger)
return estimator
Created on Fri May 17 10:56:28 2019
@author: lukasmalik
"""
# ===========================================================================
import numpy as np
import pandas as pd
from nltk.tokenize import TweetTokenizer
from scripts.load_data import *
from scripts.load_model import *
from tqdm import tqdm
df = load_data()
model = load_model() # loads pretrained word2vec model with 400 dimensions
# ===========================================================================
# calculate the document vector as average of all the words
def avg_feature_vector(tweet, model, num_features,index2word_set,tokenizer):
'''
calculates the average vector
'''
feature_vec = np.zeros((num_features, ), dtype='float32')
n_words = 0
words = tokenizer.tokenize(tweet)
for word in words:
#print(word) # sanity check
if word in index2word_set:
n_words += 1
feature_vec = np.add(feature_vec, model[word])
if (n_words > 0):
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("ESIM arguments: " + str(Config.esim_hyper_param))
if hasattr(Config, 'is_snopes'):
is_snopes = Config.is_snopes
else:
is_snopes = False
logger.debug("is_snopes: " + str(is_snopes))
if mode == RTERunPhase.train:
# training mode
training_set = embed_data_set_with_bert(Config.training_set_file, Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
is_snopes=is_snopes,
port=Config.bert_port,
port_out=Config.bert_port_out)
h_sent_sizes = training_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
training_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
training_set['data']['h_sizes'] = h_sizes
training_set['data']['h_bert_np'] = np.expand_dims(training_set['data']['h_bert_np'], 1)
valid_set = embed_data_set_with_bert(Config.dev_set_file, Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
is_snopes=is_snopes,
port=Config.bert_port,
port_out=Config.bert_port_out)
h_sent_sizes = valid_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
valid_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
valid_set['data']['h_sizes'] = h_sizes
valid_set['data']['h_bert_np'] = np.expand_dims(valid_set['data']['h_bert_np'], 1)
X_dict = {
'X_train': training_set['data'],
'X_valid': valid_set['data'],
'y_valid': valid_set['label']
}
if estimator is None:
estimator = get_estimator(Config.estimator_name, Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0, maxMemory=1.0 - Config.max_gpu_memory)[0])
estimator.fit(X_dict, training_set['label'])
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name, Config.ckpt_folder)
test_set = embed_data_set_with_bert(Config.test_set_file, Config.db_path,
threshold_b_sent_num=Config.max_sentences,
threshold_b_sent_size=Config.max_sentence_size,
is_snopes=is_snopes,
port=Config.bert_port,
port_out=Config.bert_port_out)
h_sent_sizes = test_set['data']['h_sent_sizes']
h_sizes = np.ones(len(h_sent_sizes), np.int32)
test_set['data']['h_sent_sizes'] = np.expand_dims(h_sent_sizes, 1)
test_set['data']['h_sizes'] = h_sizes
test_set['data']['h_bert_np'] = np.expand_dims(test_set['data']['h_bert_np'], 1)
x_dict = {
'X_test': test_set['data']
}
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0, maxMemory=1.0 - Config.max_gpu_memory)[0])
predictions = estimator.predict(x_dict, restore_param_required)
generate_submission(predictions, test_set['id'], Config.test_set_file, Config.submission_file)
if 'label' in test_set:
print_metrics(test_set['label'], predictions, logger)
return estimator
def main(mode: RTERunPhase, config=None, estimator=None):
LogHelper.setup()
logger = LogHelper.get_logger(os.path.splitext(os.path.basename(__file__))[0] + "_" + str(mode))
if config is not None and isinstance(config, str):
logger.info("model: " + str(mode) + ", config: " + str(config))
Config.load_config(config)
if hasattr(Config, 'is_snopes'):
is_snopes = Config.is_snopes
else:
is_snopes = False
logger.debug("is_snopes: " + str(is_snopes))
logger.info("scorer type: " + Config.estimator_name)
logger.info("random seed: " + str(Config.seed))
logger.info("BERT sentence embedding arguments: " + str(Config.bert_sent_hyper_parameter))
if mode == RTERunPhase.train:
# training mode
if hasattr(Config, 'training_dump') and os.path.exists(Config.training_dump):
with open(Config.training_dump, 'rb') as f:
(X_train, Y_labels_train, X_valid, Y_labels_valid) = pickle.load(f)
else:
# process training JSONL file
X_train, Y_labels_train = read_data_set_from_jsonl(Config.training_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
X_valid, Y_labels_valid = read_data_set_from_jsonl(Config.dev_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
X_train['b_sizes'] = get_num_sents_of_bodies(X_train['b'])
X_valid['b_sizes'] = get_num_sents_of_bodies(X_valid['b'])
b_train = X_train['b']
b_encoded_train = encode_multi_sentence_set_with_bert(b_train, Config.max_sentences, port=Config.bert_port,
port_out=Config.bert_port_out)
X_train['b'] = b_encoded_train
logger.debug("b_encoded_train.shape: " + str(b_encoded_train.shape))
h_train = X_train['h']
h_encoded_train = encode_single_sentence_set_with_bert(h_train, port=Config.bert_port,
port_out=Config.bert_port_out)
X_train['h'] = h_encoded_train
logger.debug("h_encoded_train.shape: " + str(h_encoded_train.shape))
b_valid = X_valid['b']
b_encoded_valid = encode_multi_sentence_set_with_bert(b_valid, Config.max_sentences, port=Config.bert_port,
port_out=Config.bert_port_out)
X_valid['b'] = b_encoded_valid
logger.debug("b_encoded_valid.shape: " + str(b_encoded_valid.shape))
h_valid = X_valid['h']
h_encoded_valid = encode_single_sentence_set_with_bert(h_valid, port=Config.bert_port,
port_out=Config.bert_port_out)
X_valid['h'] = h_encoded_valid
logger.debug("h_encoded_valid.shape: " + str(h_encoded_valid.shape))
if hasattr(Config, 'training_dump'):
with open(Config.training_dump, 'wb') as f:
pickle.dump((X_train, Y_labels_train, X_valid, Y_labels_valid), f, protocol=pickle.HIGHEST_PROTOCOL)
if estimator is None:
estimator = get_estimator(Config.estimator_name, Config.ckpt_folder)
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0, maxMemory=1.0 - Config.max_gpu_memory)[0])
estimator.fit(X_train, Y_labels_train, X_valid, Y_labels_valid)
save_model(estimator, Config.model_folder, Config.pickle_name, logger)
else:
# testing mode
restore_param_required = estimator is None
if estimator is None:
estimator = load_model(Config.model_folder, Config.pickle_name)
if estimator is None:
estimator = get_estimator(Config.estimator_name, Config.ckpt_folder)
X_test, Y_labels_test = read_data_set_from_jsonl(Config.test_set_file,
Config.db_path,
num_sentences=Config.max_sentences,
is_snopes=is_snopes)
X_test['b_sizes'] = get_num_sents_of_bodies(X_test['b'])
b_test = X_test['b']
b_encoded_test = encode_multi_sentence_set_with_bert(b_test, Config.max_sentences, port=Config.bert_port,
port_out=Config.bert_port_out)
X_test['b'] = b_encoded_test
logger.debug("b_encoded_test.shape: " + str(b_encoded_test.shape))
h_test = X_test['h']
h_encoded_test = encode_single_sentence_set_with_bert(h_test, port=Config.bert_port,
port_out=Config.bert_port_out)
X_test['h'] = h_encoded_test
logger.debug("h_encoded_test.shape: " + str(h_encoded_test.shape))
if 'CUDA_VISIBLE_DEVICES' not in os.environ or not str(os.environ['CUDA_VISIBLE_DEVICES']).strip():
os.environ['CUDA_VISIBLE_DEVICES'] = str(
GPUtil.getFirstAvailable(maxLoad=1.0, maxMemory=1.0 - Config.max_gpu_memory)[0])
predictions = estimator.predict(X_test, restore_param_required)
generate_submission(predictions, X_test['id'], Config.test_set_file, Config.submission_file)
if Y_labels_test is not None:
print_metrics(Y_labels_test, predictions, logger)
return estimator