def test(args):
model_dir = args.model_dir
with open(os.path.join(model_dir, 'args.pkl'), 'rb') as f:
model_args = pickle.load(f)
if not hasattr(model_args, 'language'):
model_args.language = 'eng'
#model_args.stags_dir = 'pred'
fn_txt_valid = 'data/{}/conll09/{}.txt'.format(model_args.language,
args.data)
#fn_preds_valid = 'data/{}/conll09/pred/{}_predicates.txt'.format(model_args.language, args.data)
#fn_preds_valid = 'data/{}/conll09/pred/_{}_predicates_mt.txt'.format(model_args.language, args.data)
fn_preds_valid = 'data/{}/conll09/pred/{}_predicates.txt'.format(
model_args.language, args.data)
if args.stags is None:
fn_stags_valid = 'data/{}/conll09/{}/{}_stags_{}.txt'.format(
model_args.language, model_args.stags_dir, args.data,
model_args.stag_type)
else:
fn_stags_valid = args.stags
fn_sys = '{}.txt'.format(args.data)
fn_sys = os.path.join(args.model_dir, fn_sys)
vocabs = vocab.get_vocabs(model_args.language, model_args.stag_type)
with tf.Graph().as_default():
tf.set_random_seed(model_args.seed)
np.random.seed(model_args.seed)
print("Building model...")
model = SRL_Model(vocabs, model_args)
saver = tf.train.Saver()
with tf.Session() as session:
print('Restoring model...')
saver.restore(session, tf.train.latest_checkpoint(model_dir))
print('-' * 78)
print('Validating...')
valid_loss = model.run_testing_epoch(session, vocabs, fn_txt_valid,
fn_preds_valid,
fn_stags_valid, fn_sys,
model_args.language)
print('Validation loss: {}'.format(valid_loss))
print('-' * 78)
print('Running evaluation script...')
labeled_f1, unlabeled_f1 = run_evaluation_script(
fn_txt_valid, fn_sys)
print('Labeled F1: {0:.2f}'.format(labeled_f1))
print('Unlabeled F1: {0:.2f}'.format(unlabeled_f1))
def train(args):
# Set the filepaths for training and validation
fn_txt_train = 'data/{}/conll09/train.txt'.format(args.language)
fn_preds_train = 'data/{}/conll09/{}/train_predicates.txt'.format(
args.language, args.training_split)
if args.use_stags:
fn_stags_train = 'data/{}/conll09/{}/train_stags_{}.txt'.format(
args.language, args.stags_dir, args.stag_type)
else:
fn_stags_train = fn_preds_train
fn_txt_valid = 'data/{}/conll09/dev.txt'.format(args.language)
if args.language == 'eng':
fn_preds_valid = 'data/{}/conll09/pred/_dev_predicates_mt.txt'.format(
args.language)
else:
fn_preds_valid = 'data/{}/conll09/gold/dev_predicates.txt'.format(
args.language)
if args.use_stags:
#fn_stags_valid = 'data/{}/conll09/{}/dev_stags_{}.txt'.format(
# args.language, args.stags_dir, args.stag_type)
fn_stags_valid = 'data/{}/conll09/{}/dev_stags_{}.txt'.format(
args.language, args.stags_dir, args.stag_type)
else:
fn_stags_valid = fn_preds_valid
# Come up with a model name based on the hyperparameters
model_suffix = '_'
if args.training_split == 'gold':
model_suffix += 'g'
else:
model_suffix += 'p'
if args.testing_split == 'gold':
model_suffix += 'g'
else:
model_suffix += 'p'
if args.language != 'eng':
model_suffix += '_' + args.language
if args.restrict_labels:
model_suffix += '_rl'
if args.use_stags:
model_suffix += '_st{}_{}'.format(args.stag_embed_size, args.stag_type)
if args.stags_dir == 'gold':
model_suffix += 'g'
elif args.stags_dir == 'pred':
model_suffix += 'p'
elif args.stags_dir == 'malt':
model_suffix += 'm'
elif args.stags_dir == 'pred_pos':
model_suffix += 'o'
elif args.stags_dir == 'predicted_stag_0.5':
model_suffix += '5'
elif args.stags_dir == 'pred_new':
model_suffix += 'n'
elif args.stags_dir == 'pred_elmo':
model_suffix += 'e'
if args.use_stag_features:
model_suffix += 'f{}'.format(args.stag_feature_embed_size)
if args.dropout < 1.0:
model_suffix += '_dr{}'.format(args.dropout)
if args.recurrent_dropout < 1.0:
model_suffix += '_rdr{}'.format(args.recurrent_dropout)
if args.use_word_dropout:
model_suffix += '_wdr'
if args.use_basic_classifier:
model_suffix += '_bc'
if args.use_highway_lstm:
model_suffix += '_hw'
if args.optimizer != 'adam':
model_suffix += '_' + args.optimizer
if args.seed != 89:
model_suffix += '_s{}'.format(args.seed)
fn_sys = 'output/predictions/dev{}.txt'.format(model_suffix)
# Prepare for saving the model
model_dir = 'output/models/srl' + model_suffix + '/'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
print('Saving args to', model_dir + 'args.pkl')
with open(model_dir + 'args.pkl', 'wb') as f:
pickle.dump(args, f)
vocabs = vocab.get_vocabs(args.language, args.stag_type)
with tf.Graph().as_default():
tf.set_random_seed(args.seed)
np.random.seed(args.seed)
print("Building model...")
model = SRL_Model(vocabs, args)
saver = tf.train.Saver(max_to_keep=1)
with tf.Session() as session:
best_f1 = 0
bad_streak = 0
session.run(tf.global_variables_initializer())
for i in range(args.max_epochs):
print('-' * 78)
print('Epoch {}'.format(i))
start = timer()
train_loss = model.run_training_epoch(session, vocabs,
fn_txt_train,
fn_preds_train,
fn_stags_train,
args.language)
end = timer()
print('Done with epoch {}'.format(i))
print('Avg loss: {}, total time: {}'.format(
train_loss, end - start))
print('-' * 78)
print('Validating...')
valid_loss = model.run_testing_epoch(session, vocabs,
fn_txt_valid,
fn_preds_valid,
fn_stags_valid, fn_sys,
args.language)
print('Validation loss: {}'.format(valid_loss))
print('-' * 78)
print('Running evaluation script...')
labeled_f1, unlabeled_f1 = run_evaluation_script(
fn_txt_valid, fn_sys)
print('Labeled F1: {0:.2f}'.format(labeled_f1))
print('Unlabeled F1: {0:.2f}'.format(unlabeled_f1))
if labeled_f1 > best_f1:
best_f1 = labeled_f1
bad_streak = 0
print('Saving model to', model_dir + 'model')
saver.save(session, model_dir + 'model')
else:
print('F1 deteriorated (best score: {})'.format(best_f1))
bad_streak += 1
if bad_streak >= args.early_stopping:
print(
'No F1 improvement for %d epochs, stopping early' %
args.early_stopping)
print('Best F1 score: {0:.2f}'.format(best_f1))
break