def _main(conf):
conf.experiment_name = "{}_{}_{}".format(conf.task_name, conf.dataset,
get_timestamp())
conf.model_dirpath = os.path.join("../results", conf.experiment_name)
try:
print("Create model directory:", conf.model_dirpath)
os.makedirs(conf.model_dirpath)
except OSError as e:
print(e)
# load file paths for training and validation from config file
train_questions = load_questions(conf.training_file_path)
valid_questions = load_questions(conf.validation_file_path)
if conf.max_questions < len(train_questions):
train_questions = random.sample(train_questions, conf.max_questions)
if conf.max_questions < len(valid_questions):
valid_questions = random.sample(valid_questions, conf.max_questions)
print("#Train file instances: " + str(len(train_questions)))
print("#Valid file instances: " + str(len(valid_questions)))
res = load_resources(conf)
print("Word Vocabulary", res.word_embeddings.vocabulary)
print("Char Vocabulary", res.char_vocabulary)
if conf.use_graph_embeddings == 1:
print("Graph Vocabulary", res.graph_embeddings.vocabulary)
print("Entity Vocabulary", res.entity_vocabulary)
print("Relation Vocabulary", res.relation_vocabulary)
print("Configuration:")
print(conf.description())
print("build model ...")
ranking_model = models.simple_joint_qa_predicate_model(conf, res)
ranking_model.summary()
print("Save configuration ...")
conf.save(os.path.join(conf.model_dirpath, "configuration.json"))
train(train_questions, valid_questions, ranking_model, conf, res)
print("Save weights ...")
ranking_model.save_weights(os.path.join(conf.model_dirpath,
"answer_ranking_model.hdf5"),
overwrite=True)
print("Done!")
def stage2(players):
print("Lets begin the second stage!")
first_question = True
current_player = players[0]
questions2 = load_questions(25, stage=2)
question_counter = 0
while len(players) > 3:
if not first_question:
current_player = select_player(players, select=True)
else:
first_question = False
try:
question = questions2[question_counter]
correct = ask_question(current_player, question)
except IndexError:
print("The pool of questions has ended")
correct = ask_question(current_player)
check_current_player_chances(players,
current_player,
correct,
stage="stage2")
question_counter += 1
print("The players going to the FINAL STAGE are:")
for player in players:
print(player)
return players
def main(questions_filepath, output_filepath):
questions = load_questions(questions_filepath)
with open(output_filepath, "w") as fout:
for q in questions:
text = q["text"]
# subject = q["subject"]
predicate = q["predicate"]
# for c in q["candidates"]:
# if c["uri"] == subject:
# ngram = q["ngram"]
fout.write(LABEL_PREFIX + predicate + " " + text + "\n")
def print_questions(docs_json):
qs = load_questions('test.tsv')
for i in range(200):
print(" ")
print(i)
print(qs.questions[i].question)
print(docs_json[str(qs.questions[i].document_id)][str(
qs.questions[i].passages[0])])
print(process_and_tokenize_string(qs.questions[i].question))
print(
process_and_tokenize_string(docs_json[str(
qs.questions[i].document_id)][str(
qs.questions[i].passages[0])]))
def main(questions_filepath, experiment_dir):
conf_filepath = experiment_dir + "/configuration.json"
conf = Configuration.load(conf_filepath)
if conf.use_ner:
print("Using NER questions!")
questions_filepath = "../res/test_after_ner.txt"
# take top 7 candidate entities
conf.test_top_candidates = 7
print("Configuration:")
print(conf.description())
questions = load_questions(questions_filepath)
if conf.max_questions < len(questions):
questions = random.sample(questions, conf.max_questions)
# if conf.max_questions:
# questions = random.sample(questions, conf.max_questions)
print("#Questions:", len(questions))
res = load_resources(conf)
print("Word Vocabulary", res.word_embeddings.vocabulary)
print("Char Vocabulary", res.char_vocabulary)
print("Entity Vocabulary", res.entity_vocabulary)
print("Relation Vocabulary", res.relation_vocabulary)
print("build model ...")
ranking_model = None
ranking_model = models.simple_joint_qa(conf, res)
ranking_model.load_weights(
os.path.join(conf.model_dirpath, "best_answer_ranking_model.hdf5"))
ranking_model.summary()
# ranking_model = None
print('Starting predicting ...\n')
output_filepath = os.path.join(conf.model_dirpath, "predicted_answers.txt")
predicted_data_batches = predict(questions, ranking_model, conf, res)
print('Writing predictions to the model path')
write_prediction(predicted_data_batches, output_filepath, conf)
print('Writing prediction summary to the model path')
recall_output_filepath = os.path.join(conf.model_dirpath,
"prediction_summary.txt")
recall_values = recall_at_n(predicted_data_batches, [1, 3, 5, 10, 20, 50],
conf)
write_prediction_summary(recall_values, recall_output_filepath, conf)
def final_stage(players):
print("Lets begin the final stage!")
prepare_players_for_final(players)
own_question = False
thirty_points_reached = False
max_num_of_questions = 40
questions3 = load_questions(max_num_of_questions, stage=2)
for n in range(max_num_of_questions):
question = questions3[n]
# the stage when 'first come first served'
if not thirty_points_reached:
correct = ask_question(question=question)
current_player = select_player(players, select=True)
check_current_player_chances(players,
current_player,
correct,
stage="final_stage")
if correct:
add_points(current_player, own_question)
thirty_points_reached = check_30_points(players)
else:
# getting previously answering player for the sake of own question
previous_player = current_player
current_player = select_player(players, select=True)
# own question
if current_player.post == previous_player.post and thirty_points_reached:
own_question = True
correct = ask_question(current_player, question=question)
check_current_player_chances(players,
current_player,
correct,
stage="final_stage")
if correct:
add_points(current_player, own_question)
own_question = False
# after the wrong the decision we come back to the stage when 'first come first served'
if current_player == previous_player and not correct:
thirty_points_reached = False
present_scores(players)
if check_winner(players):
exit()
# checking winner after the end of questions
check_winner(players, end=True)
def export_to_file(docs_json, document_indexer, passage_indexer):
if reindex_documents:
print('reindexing_documents is on.. refusing to export to file')
return
data = [] # load_json_from_file('answers')
start = 0
qs = load_questions('test.tsv')
document_indexer.index(docs_json, reindex_documents)
question_count = len(qs.questions)
for i in range(start, start + question_count):
query = qs.questions[i]
answers = []
response = dict()
response['id'] = query.qid
response['answers'] = answers
try:
print('Processing question: ' + str(query.qid) + " (" + str(i) +
")")
top_docs = document_indexer.execute_query(query.question)
sliced_docs = {
top_doc.doc.get_id(): docs_json[str(top_doc.doc.get_id())]
for top_doc in top_docs[0:2]
}
passage_indexer.index(sliced_docs, False)
top_passages = passage_indexer.execute_query(query.question)
for j in range(5):
answer = {
'answer':
str(top_passages[j].passage.get_doc_id()) + ':' +
str(top_passages[j].passage.get_id()),
'score':
str(top_passages[j].get_score())
}
answers.append(answer)
data.append(response)
except Exception as e:
print("error")
with open('data\\answers.json', 'w', encoding='utf-8') as outfile:
json.dump(data, outfile, ensure_ascii=False, indent=2)
def main(questions_filepath, conf):
# conf = Configuration.load(config_filepath)
# conf.padding_position = "pre"
print("Configuration:")
print(conf.description())
questions = load_questions(questions_filepath)
if conf.max_questions < len(questions):
questions = random.sample(questions, conf.max_questions)
# if conf.max_questions:
# questions = random.sample(questions, conf.max_questions)
print("#Questions:", len(questions))
res = load_resources(conf)
print("Word Vocabulary", res.word_embeddings.vocabulary)
print("Char Vocabulary", res.char_vocabulary)
print("Entity Vocabulary", res.entity_vocabulary)
print("Relation Vocabulary", res.relation_vocabulary)
print("build model ...")
ranking_model = None
ranking_model = models.simple_joint_qa_predicate_model(conf, res)
ranking_model.load_weights(os.path.join(conf.model_dirpath, "best_answer_ranking_model.hdf5"))
ranking_model.summary()
# ranking_model = None
print('Starting predicting ...\n')
output_filepath = os.path.join(conf.model_dirpath, "predicate_predicted_answers.txt")
recall_output_filepath = os.path.join(conf.model_dirpath, "predicate_prediction_summary.txt")
predicted_data = predict(questions, ranking_model, conf, res)
recall_ranges = [1, 3, 5, 10, 20, 50]
print('Computing recall values and writing predictions to the model path')
compute_recall_and_write_predictions(predicted_data, output_filepath, recall_output_filepath, recall_ranges, conf,
res)
def score_documents_retrieval(docs_json, document_indexer, passage_indexer):
count = 0
start = 70
question_count = 20
qs = load_questions('test.tsv')
document_indexer.index(None, False)
for i in range(start, start + question_count):
query = qs.questions[i]
top_docs = document_indexer.execute_query(query.question)
sliced_docs = {
top_doc.doc.get_id(): docs_json[str(top_doc.doc.get_id())]
for top_doc in top_docs[0:3]
}
passage_indexer.index(sliced_docs, True)
top_passages = passage_indexer.execute_query(query.question)
print('\nExpected: document: ' + str(qs.questions[i].document_id) +
" Passages: " + str(qs.questions[i].passages))
print('Results:\n' + str(top_passages[0:10]))
def stage1(players):
print("Let's begin the first stage")
questions1 = load_questions(len(players), stage=1)
num_of_players = len(players)
for n in range((num_of_players + 1) * 2):
current_player = select_player(players, num_of_players, n)
if not current_player:
continue
try:
question = questions1[n]
correct = ask_question(current_player, question)
except IndexError:
print("The pool of questions has ended")
correct = ask_question(current_player)
check_current_player_chances(players,
current_player,
correct,
stage="stage1")
check_winner(players)
print("The players going to the SECOND STAGE are:")
for player in players:
print(player)
return players