def run_on_sents(self, session, sents):
input_sents = map(lambda x: ' '.join(x), sents)
self.loader.run_on_sents(input_sents)
feed = {}
predictions_batch = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[
feat]] = self.loader.inputs_test_batch[feat]
feed[self.keep_prob] = 1.0
feed[self.hidden_prob] = 1.0
feed[self.input_keep_prob] = 1.0
feed[self.mlp_prob] = 1.0
weight, arc_outputs, rel_scores, predicted_stags, predicted_jk = session.run(
[
self.weight, self.arc_outputs, self.rel_scores,
self.predicted_stags, self.predicted_jk
],
feed_dict=feed)
weight = weight.astype(bool)
predicted_stags = predicted_stags[weight]
predicted_jk = predicted_jk[weight]
predictions_batch['stags'] = predicted_stags
predictions_batch['jk'] = predicted_jk
non_padding = weight.astype(bool)
non_padding[:, 0] = True ## take the dummy root nodes
predicted_arcs, predicted_rels = predict_arcs_rels(
arc_outputs, rel_scores, non_padding)
predictions_batch['arcs'] = predicted_arcs
predictions_batch['rels'] = predicted_rels
results = self.loader.get_results(predictions_batch)
return results
def run_on_sents(self, session, sents):
input_sents = map(lambda x: ' '.join(x), sents)
self.loader.run_on_sents(input_sents)
feed = {}
predictions_batch = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[feat]] = self.loader.inputs_test_batch[feat]
feed[self.keep_prob] = 1.0
feed[self.hidden_prob] = 1.0
feed[self.input_keep_prob] = 1.0
feed[self.mlp_prob] = 1.0
weight, arc_outputs, rel_scores, predicted_stags, predicted_jk, arc_probs, rel_probs = session.run([self.weight, self.arc_outputs, self.rel_scores, self.predicted_stags, self.predicted_jk, self.arc_probs, self.rel_probs], feed_dict=feed)
weight = weight.astype(bool)
predicted_stags = predicted_stags[weight]
predicted_jk = predicted_jk[weight]
predictions_batch['stags'] = predicted_stags
predictions_batch['jk'] = predicted_jk
non_padding = weight.astype(bool)
non_padding[:, 0] = True ## take the dummy root nodes
predicted_arcs, predicted_rels = predict_arcs_rels(arc_outputs, rel_scores, non_padding)
predictions_batch['arcs'] = predicted_arcs
predictions_batch['rels'] = predicted_rels
results = self.loader.get_results(predictions_batch)
## nbest
true_arc_probs = []
true_rel_probs = []
weight_with_root = weight.copy()
weight_with_root[:,0] = True
for sent_idx in xrange(arc_probs.shape[0]):
true_arc_probs.append(arc_probs[sent_idx][weight[sent_idx]][:,weight_with_root[sent_idx]])
true_rel_probs.append(rel_probs[sent_idx][weight[sent_idx]][:,weight_with_root[sent_idx]])
return results, true_arc_probs, true_rel_probs
def run_batch(self, session, testmode = False):
if not testmode:
feed = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[feat]] = self.loader.inputs_train_batch[feat]
feed[self.keep_prob] = self.opts.dropout_p
feed[self.hidden_prob] = self.opts.hidden_p
feed[self.input_keep_prob] = self.opts.input_dp
feed[self.mlp_prob] = self.opts.mlp_prob
feed[self.word_dropout] = self.opts.word_dropout
if self.opts.word_dropout_alpha > 0:
feed[self.word_dropout_alpha] = self.loader.word_dropout_alpha_vec
if self.opts.word_dropout_jw < 1.0:
feed[self.word_dropout_jw] = self.opts.word_dropout_jw
train_op = self.train_op
_, loss, UAS, rel_acc = session.run([train_op, self.loss, self.UAS, self.rel_acc], feed_dict=feed)
return loss, UAS, rel_acc
else:
feed = {}
predictions_batch = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[feat]] = self.loader.inputs_test_batch[feat]
feed[self.keep_prob] = 1.0
feed[self.hidden_prob] = 1.0
feed[self.input_keep_prob] = 1.0
feed[self.mlp_prob] = 1.0
feed[self.word_dropout] = 1.0
if self.opts.word_dropout_alpha > 0:
feed[self.word_dropout_alpha] = np.ones(self.loader.word_embeddings.shape[0])
if self.opts.word_dropout_jw < 1.0:
feed[self.word_dropout_jw] = 1.0
# loss, accuracy, predictions, weight = session.run([self.loss, self.accuracy, self.predictions, self.weight], feed_dict=feed)
loss, predicted_arcs, predicted_rels, UAS, weight, arc_outputs, rel_scores = session.run([self.loss, self.predicted_arcs, self.predicted_rels, self.UAS, self.weight, self.arc_outputs, self.rel_scores], feed_dict=feed)
if self.test_opts is not None:
if self.opts.embedding_dim > 0:
word_embeddings = session.run(self.embeddings)
with open(os.path.join(self.opts.model_dir, 'word_embeddings.pkl'), 'wb') as fout:
pickle.dump(word_embeddings, fout)
pos_embeddings = session.run(self.pos_embeddings)
with open(os.path.join(self.opts.model_dir, 'pos_embeddings.pkl'), 'wb') as fout:
pickle.dump(pos_embeddings, fout)
weights = session.run(self.input_weights, feed_dict=feed)
with open(os.path.join(self.opts.model_dir, 'pos_weight.pkl'), 'wb') as fout:
pickle.dump(weights, fout)
weight = weight.astype(bool)
predicted_arcs_greedy = predicted_arcs[weight]
predicted_rels_greedy = predicted_rels[weight]
predictions_batch['arcs_greedy'] = predicted_arcs_greedy
predictions_batch['rels_greedy'] = predicted_rels_greedy
non_padding = weight.astype(bool)
non_padding[:, 0] = True ## take the dummy root nodes
predicted_arcs, predicted_rels = predict_arcs_rels(arc_outputs, rel_scores, non_padding)
predictions_batch['arcs'] = predicted_arcs
predictions_batch['rels'] = predicted_rels
# print(predicted_greedy_arcs.shape)
# print(predicted_arcs.shape)
#print(arc_outputs.shape)
return loss, predictions_batch, UAS
def run_batch(self, session, testmode=False):
if not testmode:
feed = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[
feat]] = self.loader.inputs_train_batch[feat]
feed[self.keep_prob] = self.opts.dropout_p
feed[self.hidden_prob] = self.opts.hidden_p
feed[self.input_keep_prob] = self.opts.input_dp
feed[self.mlp_prob] = self.opts.mlp_prob
train_op = self.train_op
_, loss, UAS, rel_acc, stag_acc = session.run(
[train_op, self.loss, self.UAS, self.rel_acc, self.stag_acc],
feed_dict=feed)
return loss, UAS, rel_acc, stag_acc
else:
feed = {}
predictions_batch = {}
for feat in self.inputs_placeholder_dict.keys():
feed[self.inputs_placeholder_dict[
feat]] = self.loader.inputs_test_batch[feat]
feed[self.keep_prob] = 1.0
feed[self.hidden_prob] = 1.0
feed[self.input_keep_prob] = 1.0
feed[self.mlp_prob] = 1.0
# loss, accuracy, predictions, weight = session.run([self.loss, self.accuracy, self.predictions, self.weight], feed_dict=feed)
loss, predicted_arcs, predicted_rels, UAS, weight, arc_outputs, rel_scores, stag_acc, predicted_stags, predicted_jk, probs = session.run(
[
self.loss, self.predicted_arcs, self.predicted_rels,
self.UAS, self.weight, self.arc_outputs, self.rel_scores,
self.stag_acc, self.predicted_stags, self.predicted_jk,
self.probs
],
feed_dict=feed)
weight = weight.astype(bool)
predicted_arcs_greedy = predicted_arcs[weight]
predicted_rels_greedy = predicted_rels[weight]
predicted_stags = predicted_stags[weight]
predicted_jk = predicted_jk[weight]
predictions_batch['arcs_greedy'] = predicted_arcs_greedy
predictions_batch['rels_greedy'] = predicted_rels_greedy
predictions_batch['stags'] = predicted_stags
predictions_batch['jk'] = predicted_jk
non_padding = weight.astype(bool)
non_padding[:, 0] = True ## take the dummy root nodes
predicted_arcs, predicted_rels = predict_arcs_rels(
arc_outputs, rel_scores, non_padding)
predictions_batch['arcs'] = predicted_arcs
predictions_batch['rels'] = predicted_rels
probs = probs[weight]
# print(predicted_greedy_arcs.shape)
# print(predicted_arcs.shape)
#print(arc_outputs.shape)
return loss, predictions_batch, UAS, probs
