train = fit(train, 'labels')
test = fit(test, 'labels')
# print(train)
if transformer_config["evaluate_during_training"]:
if transformer_config["n_fold"] > 1:
test_preds = np.zeros((len(test), transformer_config["n_fold"]))
for i in range(transformer_config["n_fold"]):
if os.path.exists(
transformer_config['output_dir']) and os.path.isdir(
transformer_config['output_dir']):
shutil.rmtree(transformer_config['output_dir'])
model = QuestModel(MODEL_TYPE,
MODEL_NAME,
num_labels=NUM_LABELS,
use_cuda=torch.cuda.is_available(),
args=transformer_config)
train, eval_df = train_test_split(train,
test_size=0.11,
random_state=SEED * i)
# model.train_model(train, eval_df=eval_df)
if NUM_LABELS == 1:
model.train_model(train,
eval_df=eval_df,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
else:
model.train_model(train,
eval_df=eval_df,
train = fit(train, 'labels')
test = fit(test, 'labels')
if transformer_config["evaluate_during_training"]:
if transformer_config["n_fold"] > 1:
test_preds = np.zeros((len(test), transformer_config["n_fold"]))
for i in range(transformer_config["n_fold"]):
if os.path.exists(
transformer_config['output_dir']) and os.path.isdir(
transformer_config['output_dir']):
shutil.rmtree(transformer_config['output_dir'])
model = QuestModel(MODEL_TYPE,
MODEL_NAME,
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=transformer_config)
train, eval_df = train_test_split(train,
test_size=0.1,
random_state=SEED * i)
model.train_model(train,
eval_df=eval_df,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
model = QuestModel(MODEL_TYPE,
transformer_config["best_model_dir"],
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=transformer_config)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--results_fname')
parser.add_argument('--train_path')
parser.add_argument('--test_path')
parser.add_argument('--inject_features')
parser.add_argument('--output_dir')
args = parser.parse_args()
train_config.update({
'output_dir':
os.path.join(args.output_dir, 'outputs'),
'best_model_dir':
os.path.join(args.output_dir, 'best_model'),
'cache_dir':
os.path.join(args.output_dir, 'cache_dir')
})
train, test = read_data_files(args.train_path,
args.test_path,
inject_features=args.inject_features)
if train_config['evaluate_during_training']:
if train_config['n_fold'] > 1:
test_preds = np.zeros((len(test), train_config['n_fold']))
for i in range(train_config['n_fold']):
if os.path.exists(
train_config['output_dir']) and os.path.isdir(
train_config['output_dir']):
shutil.rmtree(train_config['output_dir'])
model = QuestModel(MODEL_TYPE,
MODEL_NAME,
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=train_config)
train, eval_df = train_test_split(train,
test_size=0.1,
random_state=SEED * i)
model.train_model(train,
eval_df=eval_df,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
model = QuestModel(MODEL_TYPE,
train_config['best_model_dir'],
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=train_config)
result, model_outputs, wrong_predictions = model.eval_model(
test,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
test_preds[:, i] = model_outputs
test['predictions'] = test_preds.mean(axis=1)
else:
model = QuestModel(MODEL_TYPE,
MODEL_NAME,
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=train_config)
train, eval_df = train_test_split(train,
test_size=0.1,
random_state=SEED)
model.train_model(train,
eval_df=eval_df,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
model = QuestModel(MODEL_TYPE,
train_config['best_model_dir'],
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=train_config)
result, model_outputs, wrong_predictions = model.eval_model(
test,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
test['predictions'] = model_outputs
else:
model = QuestModel(MODEL_TYPE,
MODEL_NAME,
num_labels=1,
use_cuda=torch.cuda.is_available(),
args=train_config)
model.train_model(train,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error,
inject_features=args.inject_features)
result, model_outputs, wrong_predictions = model.eval_model(
test,
pearson_corr=pearson_corr,
spearman_corr=spearman_corr,
mae=mean_absolute_error)
test['predictions'] = model_outputs
test = un_fit(test, 'labels')
test = un_fit(test, 'predictions')
test.to_csv(os.path.join(args.output_dir,
'{}.tsv'.format(args.results_fname)),
header=True,
sep='\t',
index=False,
encoding='utf-8')
draw_scatterplot(
test, 'labels', 'predictions',
os.path.join(args.output_dir, '{}.png'.format(args.results_fname)),
MODEL_TYPE + ' ' + MODEL_NAME)