def _train(self, ts, dropout, batchsz, model, sess, word_keep):
start_time = time.time()
steps = int(math.floor(len(ts) / float(batchsz)))
shuffle = np.random.permutation(np.arange(steps))
total_loss = total_err = total_sum = 0
for i in range(steps):
si = shuffle[i]
ts_i = batch(ts, si, batchsz)
feed_dict = model.ex2dict(ts_i, 1.0 - dropout, True, word_keep)
_, step, summary_str, lossv = sess.run(
[self.train_op, self.global_step, self.summary_op, self.loss],
feed_dict=feed_dict)
self.train_writer.add_summary(summary_str, step)
total_loss += lossv
duration = time.time() - start_time
print('Train (Loss %.4f) (%.3f sec)' %
(float(total_loss) / len(ts), duration))
def _train(self, ts, dropout, batchsz, model, sess, word_keep):
start_time = time.time()
steps = int(math.floor(len(ts) / float(batchsz)))
shuffle = np.random.permutation(np.arange(steps))
total_loss = total_err = total_sum = 0
for i in range(steps):
si = shuffle[i]
ts_i = batch(ts, si, batchsz)
if self.drop_words > 0.0:
word_probs = np.random.random(ts_i['x'].shape)
drop_indices = np.where((word_probs > self.drop_words)
& (ts_i['x'] != self.vocab['<PAD>']))
ts_i['x'][drop_indices[0],
drop_indices[1]] = self.vocab['<OOV>']
feed_dict = model.ex2dict(ts_i, 1.0 - dropout, True, word_keep)
_, step, summary_str, lossv = sess.run(
[self.train_op, self.global_step, self.summary_op, model.loss],
feed_dict=feed_dict)
self.train_writer.add_summary(summary_str, step)
total_loss += lossv
duration = time.time() - start_time
print('Train (Loss %.4f) (%.3f sec)' %
(float(total_loss) / len(ts), duration))
def test(self, ts, batchsz=1, phase='Test', conll_file=None, txts=None):
total_correct = total_sum = fscore = 0
total_gold_count = total_guess_count = total_overlap_count = 0
start_time = time.time()
steps = int(math.floor(len(ts) / float(batchsz)))
# Only if they provide a file and the raw txts, we can write CONLL file
handle = None
if conll_file is not None and txts is not None:
handle = open(conll_file, "w")
# total_cm = np.ndarray((len(self.idx2label), len(self.idx2label)))
for i in range(steps):
ts_i = batch(ts, i, batchsz)
correct, count, overlaps, golds, guesses = self._batch(
ts_i, handle, txts)
total_correct += correct
total_sum += count
total_gold_count += golds
total_guess_count += guesses
total_overlap_count += overlaps
# total_cm += cm
duration = time.time() - start_time
total_acc = total_correct / float(total_sum)
# Only show the fscore if requested
if self.fscore > 0:
fscore = fScore(total_overlap_count, total_gold_count,
total_guess_count, self.fscore)
print('%s (F%d = %.4f) (Acc %d/%d = %.4f) (%.3f sec)' %
(phase, self.fscore, float(fscore), total_correct, total_sum,
total_acc, duration))
# #show label specific metrics
# for i in len(self.idx2label.keys()):
# label = self.idx2label[i]
# not_i = [x for x in self.idx2label.keys() if x != i]
# tp = total_cm[i,i]
# fp = np.sum(total_cm[i, not_i])
# fn = np.sum(total_cm[not_i, i])
# fscore_l = fScore(tp, tp+fn, tp+fp, 1)
# print ('%s -- %s (F%d = %.4f)' % (
# phase, label, self.fscore, fscore_l))
else:
print('%s (Acc %d/%d = %.4f) (%.3f sec)' %
(phase, total_correct, total_sum, total_acc, duration))
if handle is not None:
handle.close()
return total_acc, fscore
def test(self, ts, batchsz=1, phase='Test', conll_file=None, txts=None):
total_correct = total_sum = fscore = 0
total_gold_count = total_guess_count = total_overlap_count = 0
start_time = time.time()
steps = int(math.floor(len(ts)/float(batchsz)))
# Only if they provide a file and the raw txts, we can write CONLL file
handle = None
if conll_file is not None and txts is not None:
handle = open(conll_file, "w")
for i in range(steps):
ts_i = batch(ts, i, batchsz)
correct, count, overlaps, golds, guesses = self._batch(ts_i, handle, txts)
total_correct += correct
total_sum += count
total_gold_count += golds
total_guess_count += guesses
total_overlap_count += overlaps
duration = time.time() - start_time
total_acc = total_correct / float(total_sum)
# Only show the fscore if requested
if self.fscore > 0:
fscore = fScore(total_overlap_count,
total_gold_count,
total_guess_count,
self.fscore)
print('%s (F%d = %.4f) (Acc %d/%d = %.4f) (%.3f sec)' %
(phase,
self.fscore,
float(fscore),
total_correct,
total_sum,
total_acc,
duration))
else:
print('%s (Acc %d/%d = %.4f) (%.3f sec)' %
(phase,
total_correct,
total_sum, total_acc,
duration))
if handle is not None:
handle.close()
return total_acc, fscore