test_preds = np.zeros((len(test), siamesetransquest_config["n_fold"]))
for i in range(siamesetransquest_config["n_fold"]):
if os.path.exists(siamesetransquest_config['best_model_dir']) and os.path.isdir(
siamesetransquest_config['best_model_dir']):
shutil.rmtree(siamesetransquest_config['best_model_dir'])
if os.path.exists(siamesetransquest_config['cache_dir']) and os.path.isdir(
siamesetransquest_config['cache_dir']):
shutil.rmtree(siamesetransquest_config['cache_dir'])
train_df, eval_df = train_test_split(train, test_size=0.1, random_state=SEED * i)
model = SiameseTransQuestModel(MODEL_NAME, args=siamesetransquest_config)
model.train_model(train_df, eval_df)
model = SiameseTransQuestModel(siamesetransquest_config['best_model_dir'])
dev_preds[:, i] = model.predict(dev_sentence_pairs)
test_preds[:, i] = model.predict(test_sentence_pairs)
dev['predictions'] = dev_preds.mean(axis=1)
test['predictions'] = test_preds.mean(axis=1)
dev = un_fit(dev, 'labels')
dev = un_fit(dev, 'predictions')
test = un_fit(test, 'predictions')
dev.to_csv(os.path.join(TEMP_DIRECTORY, RESULT_FILE), header=True, sep='\t', index=False, encoding='utf-8')
draw_scatterplot(dev, 'labels', 'predictions', os.path.join(TEMP_DIRECTORY, RESULT_IMAGE), "English-Chinese")
print_stat(dev, 'labels', 'predictions')
format_submission(df=test, index=index, language_pair="en-zh", method="SiameseTransQuest",
path=os.path.join(TEMP_DIRECTORY, SUBMISSION_FILE))
test_size=0.1,
random_state=SEED * i)
model = SiameseTransQuestModel(MODEL_NAME,
args=siamesetransquest_config)
model.train_model(train_df, eval_df)
model = SiameseTransQuestModel(
siamesetransquest_config['best_model_dir'])
dev_preds[:, i] = model.predict(dev_sentence_pairs)
test_preds[:, i] = model.predict(test_sentence_pairs)
dev['predictions'] = dev_preds.mean(axis=1)
test['predictions'] = test_preds.mean(axis=1)
dev = un_fit(dev, 'labels')
dev = un_fit(dev, 'predictions')
test = un_fit(test, 'predictions')
dev.to_csv(os.path.join(TEMP_DIRECTORY, RESULT_FILE),
header=True,
sep='\t',
index=False,
encoding='utf-8')
draw_scatterplot(dev, 'labels', 'predictions',
os.path.join(TEMP_DIRECTORY, RESULT_IMAGE), "Sinhala-English")
print_stat(dev, 'labels', 'predictions')
format_submission(df=test,
index=index,
language_pair="si-en",
method="SiameseTransQuest",
path=os.path.join(TEMP_DIRECTORY, SUBMISSION_FILE))
for dev, test, index, language in zip(dev_list, test_list, index_list,
[*languages]):
dev = un_fit(dev, 'labels')
dev = un_fit(dev, 'predictions')
test = un_fit(test, 'predictions')
dev.to_csv(os.path.join(
TEMP_DIRECTORY,
RESULT_FILE.split(".")[0] + "_" + language + "." +
RESULT_FILE.split(".")[1]),
header=True,
sep='\t',
index=False,
encoding='utf-8')
draw_scatterplot(
dev, 'labels', 'predictions',
os.path.join(
TEMP_DIRECTORY,
RESULT_IMAGE.split(".")[0] + "_" + language + "." +
RESULT_IMAGE.split(".")[1]), language)
print_stat(dev, 'labels', 'predictions')
if language == "RU-EN":
format_submission(df=test,
index=index,
language_pair=language.lower(),
method="TransQuest",
path=os.path.join(
TEMP_DIRECTORY,
SUBMISSION_FILE.split(".")[0] + "_" + language +
"." + SUBMISSION_FILE.split(".")[1]),
index_type="Auto")
