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
train_op = self.train_op
_, loss, accuracy = session.run(
[train_op, self.loss, self.accuracy], feed_dict=feed)
return loss, accuracy
else:
feed = {}
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
if self.beam_size == 0:
loss, accuracy, predictions, weight = session.run(
[self.loss, self.accuracy, self.predictions, self.weight],
feed_dict=feed)
weight = weight.astype(bool)
predictions = predictions[weight]
return loss, accuracy, predictions
else:
predictions, scores, weight_beam, weight, back_pointers = session.run(
[
self.predictions, self.scores, self.weight_beam,
self.weight, self.back_pointers
],
feed_dict=feed)
weight = weight.astype(bool)
weight_beam = weight_beam.astype(bool)
predictions, scores, indices = back_track(
predictions, scores, back_pointers, weight_beam)
## predictions [batch_size, seq_len], scores [batch_size, seq_en], back_pointer [batch_size, seq_len-1]
b = predictions.shape[0] / self.beam_size
n = predictions.shape[1]
predictions = predictions.reshape([b, -1])[:, :n]
## [bB, n] => [b, Bn]
predictions = predictions[weight]
scores = scores[weight_beam]
return predictions, scores
def run_batch(self, session, testmode=False):
if not testmode:
feed = {}
for placeholder, data in zip(self.inputs_placeholder_list,
self.loader.inputs_train_batch):
feed[placeholder] = data
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
train_op = self.train_op
#train_op = tf.no_op()
_, loss, accuracy, predictions, global_loss, scores = session.run(
[
train_op, self.loss, self.accuracy, self.predictions,
self.global_loss, self.scores
],
feed_dict=feed)
#scores = scores.T
#for i in xrange(scores.shape[0]):
# print(i)
# print(scores[i, :])
#for i in xrange(global_loss.shape[0]):
# print(i)
# print(global_loss[i, :])
#for i in xrange(17):
# print(i)
# print(global_loss_test[i, :])
#predictions = predictions.reshape([3, 4, 18])
#for i in xrange(18):
# print(i)
# print(predictions[:, :, i])
return loss, accuracy
else:
feed = {}
for placeholder, data in zip(self.inputs_placeholder_list,
self.loader.inputs_test_batch):
feed[placeholder] = data
feed[self.keep_prob] = 1.0
feed[self.hidden_prob] = 1.0
feed[self.input_keep_prob] = 1.0
if self.beam_size == 0:
loss, accuracy, predictions, weight = session.run(
[self.loss, self.accuracy, self.predictions, self.weight],
feed_dict=feed)
weight = weight.astype(bool)
predictions = predictions[weight]
return loss, accuracy, predictions
else:
predictions, scores, weight_beam, weight, back_pointers, correct_so_far = session.run(
[
self.predictions, self.scores, self.weight_beam,
self.weight, self.back_pointers, self.correct_so_far
],
feed_dict=feed)
weight = weight.astype(bool)
weight_beam = weight_beam.astype(bool)
#predictions_test = predictions.reshape([3, 4, 18])
#for i in xrange(17):
# print(i)
# print(np.sum(correct_so_far[i, :, :], axis=-1))
# print(predictions_test[:, :, i])
#print(predictions_test.shape)
#for i in xrange(18):
# print(i)
# print(predictions_test[:, :, i])
#print(predictions_test.shape)
predictions, scores, indices = back_track(
predictions, scores, back_pointers, weight_beam)
## predictions [batch_size, seq_len], scores [batch_size, seq_en], back_pointer [batch_size, seq_len-1]
b = predictions.shape[0] / self.beam_size
n = predictions.shape[1]
#predictions = predictions.reshape([3, 4, 18])
#for i in xrange(18):
# print(i)
# print(predictions[:, :, i])
#print(predictions.shape)
predictions = predictions.reshape([b, -1])[:, :n]
## [bB, n] => [b, Bn]
predictions = predictions[weight]
scores = scores[weight_beam]
return predictions, scores