def _sg_losses(self, ops={}, suffix=''):
ops = self._frc_losses(ops, suffix)
rel_score = self.get_output('rel_score'+suffix)
cls_score = self.get_output('cls_score' + suffix)
ops['cls_score'] = cls_score
ops['rel_score'] = rel_score
ops['cls_pred'] = self.get_output('cls_pred'+suffix)
ops['predicates'] = self.data['predicates']
ops['rel_rois'] = self.data['rel_rois']
ops['relations'] = self.data['relations']
relations = self.data['relations']
ops['labels'] = self.data['labels']
ops['Xr'] = self.Xr
cls_pred = self.get_output('cls_pred'+suffix)
sbjs = tf.gather(cls_pred, relations[:, 0])
sbjs = tf.expand_dims(sbjs,1)
objs = tf.gather(cls_pred,relations[:,1])
objs = tf.expand_dims(objs,1)
ops['indices'] = tf.cast(tf.concat(concat_dim=1, values=[sbjs,objs]),tf.int32)
ops['Xr'] = tf.gather_nd(params=self.Xr, indices=tf.concat(concat_dim=1, values=ops['indices']))
# point-wise multiplication
ops['Xr'] = tf.nn.softmax(ops['Xr']) # apply softmax
res = 1 - self.mask
# rel_score = tf.multiply(rel_score, ops['Xr'])
rel_score = rel_score * (res + (self.mask*ops['Xr']))
ops['loss_rel'+suffix] = losses.sparse_softmax(rel_score, self.data['predicates'],
name='rel_loss'+suffix, ignore_bg=True)
return ops
def _sg_losses(self, ops={}, suffix=''):
ops = self._frc_losses(ops, suffix)
rel_score = self.get_output('rel_score' + suffix)
ops['loss_rel' + suffix] = losses.sparse_softmax(
rel_score,
self.data['predicates'],
name='rel_loss' + suffix,
ignore_bg=True)
return ops
def _frc_losses(self, ops={}, suffix=''):
# classification loss
cls_score = self.get_output('cls_score'+suffix)
# ops['cls_score' + "_0"] = cls_score
# ops['y_outs']
ops['loss_cls'+suffix] = losses.sparse_softmax(cls_score, self.data['labels'], name='cls_loss'+suffix)
# bounding box regression L1 loss
if cfg.TRAIN.BBOX_REG:
bbox_pred = self.get_output('bbox_pred'+suffix)
ops['loss_box'+suffix] = losses.l1_loss(bbox_pred, self.data['bbox_targets'], 'reg_loss'+suffix,
self.data['bbox_inside_weights'])
else:
print('NO BBOX REGRESSION!!!!!')
return ops
with sess.as_default():
if FLAGS.random_seed is not None:
tf.set_random_seed(FLAGS.random_seed)
# Build execution graph
network = BiDirectional_LSTM(sess, init, next_batch_x, features_size,
FLAGS.attention, FLAGS.attention_size)
eval_predictions, endings, train_logits = network.build_model()
# Compare with next_batch_endings_y
if FLAGS.loss_function == "SIGMOID":
loss = losses.sigmoid(next_batch_endings_y, train_logits)
elif FLAGS.loss_function == "SOFTMAX":
loss = losses.sparse_softmax(next_batch_endings_y, endings)
else:
raise RuntimeError(
f"Loss function {FLAGS.loss_function} not supported.")
accuracy = tf.reduce_mean(
tf.cast(tf.equal(eval_predictions, next_batch_endings_y),
dtype=tf.float32))
"""Initialize iterators"""
train_handle = sess.run(train_iterator.string_handle())
test_handle = sess.run(test_iterator.string_handle())
if FLAGS.use_train_set:
sess.run(train_iterator.initializer,
feed_dict={}
if FLAGS.use_skip_thoughts else {input_x: sentences})
