if cost_avg < decay_cost:
decay_cost = cost_avg
decay_epoch = epoch
rnn.save(args.model_dir, "decay_")
#XXX: estimate training model
if epoch % valid_freq == 0:
y_relations = []
for i, (x, y, words) in enumerate(zip(x_train, y_train, train_words_list)):
y_pred = rnn.predict(x)
# extract relations from current document
y_relations.extend(extract_relations(y_pred, tag_to_j, words))
# evaluate all relations
precision, recall, f1 = scorer.evaluate_relation(train_relations_list, y_relations)
log.info(" train precision {:7.4f} recall {:7.4f}, f1 {:7.4f}".format(precision, recall, f1))
# validate model
if epoch % valid_freq == 0:
y_relations = []
for i, (x, y, words) in enumerate(zip(x_valid, y_valid, valid_words_list)):
y_pred = rnn.predict(x)
# extract relations from current document
y_relations.extend(extract_relations(y_pred, tag_to_j, words))
# evaluate all relations
precision, recall, f1 = scorer.evaluate_relation(valid_relations_list, y_relations)
log.info(" valid precision {:7.4f} recall {:7.4f}, f1 {:7.4f} (best {:10.8f}) {}".format(precision, recall, f1, (f1 if f1 > best_f1 else best_f1), ("++" if f1 > best_f1 else " ")))
if f1 > best_f1: # save best model
decay_epoch = epoch
rnn.save(args.model_dir, "decay_")
# validate model
if epoch % valid_freq == 0:
y_relations = []
for i, (x, y,
words) in enumerate(zip(x_valid, y_valid,
valid_words_list)):
y_pred = rnn.predict(x)
# extract relations from current document
y_relations.extend(extract_relations(y_pred, tag_to_j, words))
# evaluate all relations
precision, recall, f1 = scorer.evaluate_relation(
valid_relations_list, y_relations)
log.info(
"valid set precision {:.2f}, recall {:.2f}, f1 {:.2f}{} (best {})"
.format(precision, recall, f1,
(" +" if f1 > best_f1 else " "), best_f1))
if f1 > best_f1: # save best model
best_f1 = f1
best_epoch = epoch
#best_rnn = copy.deepcopy(rnn)
rnn.save(args.model_dir, "f1_")
if f1 >= 1.: # perfect
print " WOOHOO!!!"
break
# learning rate decay if no improvement after some epochs
if epoch - decay_epoch >= decay_after:
#y = rnn.predict(train_x)
y_relations = []
for relation in extract_relations(doc_id, doc, y, tag_to_j,
all_tsenses):
y_relations.append(relation)
# check relations
validator.check_type(relation)
validator.check_sense(relation)
validator.check_args(relation)
validator.check_connective(relation)
# evaluate relations
try:
precision, recall, f1 = scorer.evaluate_relation(
train_y_data[doc_id], y_relations)
print "evaluate train set precision {0:.2f}, recall {1:.2f}, f1 {2:.2f}".format(
precision, recall, f1)
except ZeroDivisionError:
precision, recall, f1 = 0.0, 0.0, 0.0
if f1 == 1.0:
print "WOOHOO!!!"
break
# learning rate decay if no improvement after some epochs
if epoch - best_epoch >= decay_after:
learn_rate *= decay_rate
if epoch > epochs or learn_rate < decay_min:
break
epoch += 1
# extract relations
#y = rnn.predict(train_x)
y_relations = []
for relation in extract_relations(doc_id, doc, y, tag_to_j, all_tsenses):
y_relations.append(relation)
# check relations
validator.check_type(relation)
validator.check_sense(relation)
validator.check_args(relation)
validator.check_connective(relation)
# evaluate relations
try:
precision, recall, f1 = scorer.evaluate_relation(train_y_data[doc_id], y_relations)
print "evaluate train set precision {0:.2f}, recall {1:.2f}, f1 {2:.2f}".format(precision, recall, f1)
except ZeroDivisionError:
precision, recall, f1 = 0.0, 0.0, 0.0
if f1 == 1.0:
print "WOOHOO!!!"
break
# learning rate decay if no improvement after some epochs
if epoch - best_epoch >= decay_after:
learn_rate *= decay_rate
if epoch > epochs or learn_rate < decay_min:
break
epoch += 1
