def test_masked_avg(self):
tf.reset_default_graph()
data = tf.placeholder(tf.float32, shape=[None, None])
mask = tf.placeholder(tf.float32, shape=[None, None])
masked_avgs = utils.masked_avg(data, mask)
with self.test_session() as sess:
result = sess.run(masked_avgs,
feed_dict={
data: [[1, 2, 3], [4, 5, 6]],
mask: [[1, 0, 1], [0, 1, 0]]
})
self.assertAllClose(result, [[2], [5]])
result = sess.run(masked_avgs,
feed_dict={
data: [[1, 2, 3], [4, 5, 6]],
mask: [[0, 1, 0], [1, 0, 1]]
})
self.assertAllClose(result, [[2], [5]])
result = sess.run(masked_avgs,
feed_dict={
data: [[1, 2, 3], [4, 5, 6]],
mask: [[0, 0, 0], [0, 0, 0]]
})
self.assertAllClose(result, [[0], [0]])
def build_loss(self, predictions, **kwargs):
"""Build tf graph to compute loss.
Args:
predictions: dict of prediction results keyed by name.
Returns:
loss_dict: dict of loss tensors keyed by name.
"""
options = self._model_proto
(image_id, image_ids_gathered,
similarity) = (predictions[_FIELD_IMAGE_ID],
predictions[_FIELD_IMAGE_IDS_GATHERED],
predictions[_FIELD_SIMILARITY])
distance = 1.0 - similarity
pos_mask = tf.cast(
tf.equal(tf.expand_dims(image_id, axis=1),
tf.expand_dims(image_ids_gathered, axis=0)), tf.float32)
neg_mask = 1.0 - pos_mask
if options.triplet_ap_use_avg:
distance_ap = utils.masked_avg(distance, pos_mask)
else:
distance_ap = utils.masked_maximum(distance, pos_mask)
# negatives_outside: smallest D_an where D_an > D_ap.
mask = tf.cast(tf.greater(distance, distance_ap), tf.float32)
mask = mask * neg_mask
negatives_outside = utils.masked_minimum(distance, mask)
# negatives_inside: largest D_an.
negatives_inside = utils.masked_maximum(distance, neg_mask)
# distance_an: the semihard negatives.
mask_condition = tf.greater(tf.reduce_sum(mask, axis=1, keepdims=True),
0.0)
distance_an = tf.where(mask_condition, negatives_outside,
negatives_inside)
# Triplet loss.
losses = tf.maximum(distance_ap - distance_an + options.triplet_margin,
0)
return {
'triplet_loss': tf.reduce_mean(losses),
}
def _create_lv0_edge_scores(self, proposal_repr, slogan_repr, label_repr,
dbpedia_repr, proposal_mask, slogan_mask,
label_mask, dbpedia_mask):
"""Creates adjacency matrix. Each elem denotes an edge weight.
Args:
proposal_repr: A [batch, max_proposal_num, dims] float tensor.
slogan_repr: A [batch, max_slogan_num, dims] float tensor.
Returns:
proposal_scores: A [batch, max_proposal_num] float tensor denoting
weights of different proposals.
slogan_scores: A [batch, max_slogan_num] float tensor denoting weights
of different slogans.
label_to_proposal_scores: A [batch, max_proposal_num, max_label_num] tensor.
dbpedia_to_slogan_scores: A [batch, max_slogan_num, max_dbpedia_num] tensor.
"""
options = self._options
is_training = self._is_training
(batch_i, max_proposal_num, max_slogan_num, max_label_num,
max_dbpedia_num) = (proposal_repr.get_shape()[0].value,
utils.get_tensor_shape(proposal_repr)[1],
utils.get_tensor_shape(slogan_repr)[1],
utils.get_tensor_shape(label_repr)[1],
utils.get_tensor_shape(dbpedia_repr)[1])
with tf.name_scope('create_lv0_attention_weights'):
# Predictions for updating slogan.
# slogan_dbpedia_to_proposal_scores shape =
# [batch, max_proposal_num, max_slogan_num + max_dbpedia_num].
# slogan_dbpedia_to_slogan_scores shape =
# [batch, max_slogan_num, max_slogan_num + max_dbpedia_num].
slogan_dbpedia_repr = tf.concat([slogan_repr, dbpedia_repr],
axis=1)
slogan_dbpedia_to_proposal_scores = self._create_edge_weights_helper(
proposal_repr,
slogan_dbpedia_repr,
scope='slogan_dbpedia_to_proposal_scores')
slogan_dbpedia_to_slogan_scores = self._create_edge_weights_helper(
slogan_repr,
slogan_dbpedia_repr,
scope='slogan_dbpedia_to_slogan_scores')
# Predictions for updating proposal.
# proposal_label_to_proposal_scores shape =
# [batch, max_proposal_num, max_proposal_num + max_label_num].
# proposal_label_to_slogan_scores shape =
# [batch, max_slogan_num, max_proposal_num + max_label_num]
proposal_label_repr = tf.concat([proposal_repr, label_repr],
axis=1)
proposal_label_to_proposal_scores = self._create_edge_weights_helper(
proposal_repr,
proposal_label_repr,
scope='proposal_label_to_proposal_scores')
proposal_label_to_slogan_scores = self._create_edge_weights_helper(
slogan_repr,
proposal_label_repr,
scope='proposal_label_to_slogan_scores')
# Compute slogan_dbpedia_to_slogan_scores.
# slogan_dbpedia_context_scores shape = [batch, 1, max_slogan_num + max_dbpedia_num]
# slogan_dbpedia_to_slogan_scores shape = [batch, max_slogan_num, max_slogan_num + max_dbpedia_num]
slogan_dbpedia_context_scores = utils.masked_avg(
slogan_dbpedia_to_proposal_scores,
mask=tf.expand_dims(proposal_mask, 2),
dim=1)
slogan_dbpedia_to_slogan_scores = tf.add(
slogan_dbpedia_context_scores, slogan_dbpedia_to_slogan_scores)
slogan_scores = tf.linalg.diag_part(
tf.slice(slogan_dbpedia_to_slogan_scores,
begin=[0, 0, 0],
size=[batch_i, max_slogan_num, max_slogan_num]))
dbpedia_to_slogan_scores = tf.slice(
slogan_dbpedia_to_slogan_scores,
begin=[0, 0, max_slogan_num],
size=[batch_i, max_slogan_num, max_dbpedia_num])
# Compute proposal_label_to_proposal_scores.
# proposal_label_context_scores shape = [batch, 1, max_proposal_num + max_label_num]
# proposal_label_to_proposal_scores shape = [batch, max_proposal_num, max_proposal_num + max_label_num]
proposal_label_context_scores = utils.masked_avg(
proposal_label_to_slogan_scores,
mask=tf.expand_dims(slogan_mask, 2),
dim=1)
proposal_label_to_proposal_scores = tf.add(
proposal_label_context_scores,
proposal_label_to_proposal_scores)
proposal_scores = tf.linalg.diag_part(
tf.slice(proposal_label_to_proposal_scores,
begin=[0, 0, 0],
size=[batch_i, max_proposal_num, max_proposal_num]))
label_to_proposal_scores = tf.slice(
proposal_label_to_proposal_scores,
begin=[0, 0, max_proposal_num],
size=[batch_i, max_proposal_num, max_label_num])
return proposal_scores, slogan_scores, label_to_proposal_scores, dbpedia_to_slogan_scores
def _create_lv1_edge_scores(self, proposal_repr, slogan_repr,
proposal_mask, slogan_mask):
"""Creates adjacency matrix. Each elem denotes an edge weight.
Args:
proposal_repr: A [batch, max_proposal_num, dims] float tensor.
slogan_repr: A [batch, max_slogan_num, dims] float tensor.
Returns:
proposal_scores: A [batch, max_proposal_num] float tensor denoting
weights of different proposals.
slogan_scores: A [batch, max_slogan_num] float tensor denoting weights
of different slogans.
"""
options = self._options
is_training = self._is_training
with tf.name_scope('create_attention_weights'):
if options.attention_type == graph_creator_pb2.ConvGraphCreator.CO_ATTENTION:
# Use co-attention to determine edge importance.
# slogan_to_proposal_scores shape = [batch, max_proposal_num, max_slogan_num].
# proposal_scores shape = [batch, max_proposal_num]
# slogan_scores shape = [batch, max_slogan_num]
slogan_to_proposal_scores = self._create_edge_weights_helper(
proposal_repr,
slogan_repr,
scope='slogan_to_proposal_scores')
proposal_scores = utils.masked_avg(slogan_to_proposal_scores,
mask=tf.expand_dims(
slogan_mask, 1),
dim=2)
proposal_scores = tf.squeeze(proposal_scores, axis=-1)
slogan_scores = utils.masked_avg(slogan_to_proposal_scores,
mask=tf.expand_dims(
proposal_mask, 2),
dim=1)
slogan_scores = tf.squeeze(slogan_scores, axis=1)
elif options.attention_type == graph_creator_pb2.ConvGraphCreator.SELF_ATTENTION:
# Use self-attention to determine edge importance.
# similarity_proposal_proposal shape = [batch, max_proposal_num, max_proposal_num]
# similarity_slogan_slogan shape = [batch, max_slogan_num, max_slogan_num]
# proposal_scores shape = [batch, 1, max_proposal_num]
# slogan_scores shape = [batch, 1, max_slogan_num]
similarity_proposal_proposal = self._create_edge_weights_helper(
proposal_repr,
proposal_repr,
scope='similarity_proposal_proposal')
similarity_slogan_slogan = self._create_edge_weights_helper(
slogan_repr, slogan_repr, scope='similarity_slogan_slogan')
proposal_scores = utils.masked_avg(
similarity_proposal_proposal,
tf.expand_dims(proposal_mask, 2),
dim=1)
proposal_scores = tf.squeeze(proposal_scores, axis=1)
slogan_scores = utils.masked_avg(similarity_slogan_slogan,
tf.expand_dims(
slogan_mask, 2),
dim=1)
slogan_scores = tf.squeeze(slogan_scores, axis=1)
else:
raise ValueError('Invalid attention type %s' %
options.attention_type)
return proposal_scores, slogan_scores
def _calc_oicr_loss(self,
labels,
num_proposals,
proposals,
scores_0,
scores_1,
scope,
iou_threshold=0.5):
"""Calculates the OICR loss at refinement stage `i`.
Args:
labels: A [batch, num_classes] float tensor.
num_proposals: A [batch] int tensor.
proposals: A [batch, max_num_proposals, 4] float tensor.
scores_0: A [batch, max_num_proposal, 1 + num_classes] float tensor,
representing the proposal score at `k-th` refinement.
scores_1: A [batch, max_num_proposal, 1 + num_classes] float tensor,
representing the proposal score at `(k+1)-th` refinement.
Returns:
oicr_cross_entropy_loss: a scalar float tensor.
"""
with tf.name_scope(scope):
(batch, max_num_proposals,
num_classes_plus_one) = utils.get_tensor_shape(scores_0)
num_classes = num_classes_plus_one - 1
# For each class, look for the most confident proposal.
# proposal_ind shape = [batch, num_classes].
proposal_mask = tf.sequence_mask(num_proposals,
maxlen=max_num_proposals,
dtype=tf.float32)
proposal_ind = utils.masked_argmax(tf.nn.softmax(scores_0,
axis=-1)[:, :,
1:],
tf.expand_dims(proposal_mask,
axis=-1),
dim=1)
# Deal with the most confident proposal per each class.
# Unstack the `proposal_ind`, `labels`.
# proposal_labels shape = [batch, max_num_proposals, num_classes].
proposal_labels = []
indices_0 = tf.range(batch, dtype=tf.int64)
for indices_1, label_per_class in zip(
tf.unstack(proposal_ind, axis=-1),
tf.unstack(labels, axis=-1)):
# Gather the most confident proposal for the class.
# confident_proosal shape = [batch, 4].
indices = tf.stack([indices_0, indices_1], axis=-1)
confident_proposal = tf.gather_nd(proposals, indices)
# Get the Iou from all the proposals to the most confident proposal.
# iou shape = [batch, max_num_proposals].
confident_proposal_tiled = tf.tile(
tf.expand_dims(confident_proposal, axis=1),
[1, max_num_proposals, 1])
iou = box_utils.iou(
tf.reshape(proposals, [-1, 4]),
tf.reshape(confident_proposal_tiled, [-1, 4]))
iou = tf.reshape(iou, [batch, max_num_proposals])
# Filter out irrelevant predictions using image-level label.
target = tf.to_float(tf.greater_equal(iou, iou_threshold))
target = tf.where(label_per_class > 0,
x=target,
y=tf.zeros_like(target))
proposal_labels.append(target)
proposal_labels = tf.stack(proposal_labels, axis=-1)
# Add background targets, and normalize the sum value to 1.0.
# proposal_labels shape = [batch, max_num_proposals, 1 + num_classes].
bkg = tf.logical_not(tf.reduce_sum(proposal_labels, axis=-1) > 0)
proposal_labels = tf.concat(
[tf.expand_dims(tf.to_float(bkg), axis=-1), proposal_labels],
axis=-1)
proposal_labels = tf.div(
proposal_labels,
tf.reduce_sum(proposal_labels, axis=-1, keepdims=True))
assert_op = tf.Assert(
tf.reduce_all(
tf.abs(tf.reduce_sum(proposal_labels, axis=-1) -
1) < 1e-6),
["Probabilities not sum to ONE", proposal_labels])
# Compute the loss.
with tf.control_dependencies([assert_op]):
losses = tf.nn.softmax_cross_entropy_with_logits(
labels=tf.stop_gradient(proposal_labels), logits=scores_1)
oicr_cross_entropy_loss = tf.reduce_mean(
utils.masked_avg(data=losses, mask=proposal_mask, dim=1))
return oicr_cross_entropy_loss
def build_loss(self, predictions, examples, **kwargs):
"""Build tf graph to compute loss.
Args:
predictions: dict of prediction results keyed by name.
examples: dict of inputs keyed by name.
Returns:
loss_dict: dict of loss tensors keyed by name.
"""
options = self._model_proto
loss_dict = {}
with tf.name_scope('losses'):
# Extract image-level labels.
if not options.caption_as_label:
labels = self._extract_class_label(
class_texts=examples[InputDataFields.object_texts],
vocabulary_list=self._vocabulary_list)
else:
labels = self._extract_class_label(
class_texts=slim.flatten(
examples[InputDataFields.caption_strings]),
vocabulary_list=self._vocabulary_list)
# A prediction model from caption to class
# Loss of the multi-instance detection network.
midn_class_logits = predictions[NOD2Predictions.midn_class_logits]
losses = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels, logits=midn_class_logits)
# Hard-negative mining.
if options.midn_loss_negative_mining == nod2_model_pb2.NOD2Model.NONE:
if options.classification_loss_use_sum:
assert False
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(tf.reduce_sum(losses, axis=-1)),
options.midn_loss_weight)
else:
if options.caption_as_label:
loss_masks = tf.to_float(
tf.reduce_any(labels > 0, axis=-1))
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.squeeze(
utils.masked_avg(tf.reduce_mean(losses,
axis=-1),
mask=loss_masks,
dim=0)),
options.midn_loss_weight)
else:
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(losses), options.midn_loss_weight)
elif options.midn_loss_negative_mining == nod2_model_pb2.NOD2Model.HARDEST:
assert False
loss_masks = self._midn_loss_mine_hardest_negative(
labels, losses)
loss_dict['midn_cross_entropy_loss'] = tf.reduce_mean(
utils.masked_avg(data=losses, mask=loss_masks, dim=1))
else:
raise ValueError('Invalid negative mining method.')
# Losses of the online instance classifier refinement network.
(num_proposals,
proposals) = (predictions[DetectionResultFields.num_proposals],
predictions[DetectionResultFields.proposal_boxes])
batch, max_num_proposals, _ = utils.get_tensor_shape(proposals)
proposal_scores_0 = predictions[
NOD2Predictions.oicr_proposal_scores + '_at_0']
if options.oicr_use_proba_r_given_c:
proposal_scores_0 = predictions[
NOD2Predictions.midn_proba_r_given_c]
proposal_scores_0 = tf.concat([
tf.fill([batch, max_num_proposals, 1], 0.0), proposal_scores_0
],
axis=-1)
global_step = tf.train.get_or_create_global_step()
oicr_loss_mask = tf.cast(global_step > options.oicr_start_step,
tf.float32)
for i in range(options.oicr_iterations):
proposal_scores_1 = predictions[
NOD2Predictions.oicr_proposal_scores +
'_at_{}'.format(i + 1)]
oicr_cross_entropy_loss_at_i = model_utils.calc_oicr_loss(
labels,
num_proposals,
proposals,
tf.stop_gradient(proposal_scores_0),
proposal_scores_1,
scope='oicr_{}'.format(i + 1),
iou_threshold=options.oicr_iou_threshold)
loss_dict['oicr_cross_entropy_loss_at_{}'.format(
i + 1)] = tf.multiply(
oicr_loss_mask * oicr_cross_entropy_loss_at_i,
options.oicr_loss_weight)
proposal_scores_0 = tf.nn.softmax(proposal_scores_1, axis=-1)
# Min-entropy loss.
mask = tf.sequence_mask(num_proposals,
maxlen=max_num_proposals,
dtype=tf.float32)
proba_r_given_c = predictions[NOD2Predictions.midn_proba_r_given_c]
losses = tf.log(proba_r_given_c + _EPSILON)
losses = tf.squeeze(utils.masked_sum_nd(data=losses,
mask=mask,
dim=1),
axis=1)
min_entropy_loss = tf.reduce_mean(
tf.reduce_sum(losses * labels, axis=1))
min_entropy_loss = tf.multiply(min_entropy_loss,
options.min_entropy_loss_weight)
max_proba = tf.reduce_mean(
utils.masked_maximum(data=proba_r_given_c,
mask=tf.expand_dims(mask, -1),
dim=1))
tf.losses.add_loss(min_entropy_loss)
tf.summary.scalar('loss/min_entropy_loss', min_entropy_loss)
tf.summary.scalar('loss/max_proba', max_proba)
return loss_dict
def build_loss(self, predictions, examples, **kwargs):
"""Build tf graph to compute loss.
Args:
predictions: dict of prediction results keyed by name.
examples: dict of inputs keyed by name.
Returns:
loss_dict: dict of loss tensors keyed by name.
"""
options = self._model_proto
loss_dict = {}
with tf.name_scope('losses'):
# Extract image-level labels.
labels = self._extract_class_label(
class_texts=slim.flatten(predictions[
NOD3Predictions.training_only_caption_strings]),
vocabulary_list=self._vocabulary_list)
# A prediction model from caption to class
# Loss of the multi-instance detection network.
midn_class_logits = predictions[NOD3Predictions.midn_class_logits]
losses = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels, logits=midn_class_logits)
# Hard-negative mining.
if options.midn_loss_negative_mining == nod3_model_pb2.NOD3Model.NONE:
if options.classification_loss_use_sum:
assert False
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(tf.reduce_sum(losses, axis=-1)),
options.midn_loss_weight)
else:
if options.caption_as_label:
loss_masks = tf.to_float(
tf.reduce_any(labels > 0, axis=-1))
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.squeeze(
utils.masked_avg(tf.reduce_mean(losses,
axis=-1),
mask=loss_masks,
dim=0)),
options.midn_loss_weight)
else:
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(losses), options.midn_loss_weight)
elif options.midn_loss_negative_mining == nod3_model_pb2.NOD3Model.HARDEST:
assert False
loss_masks = self._midn_loss_mine_hardest_negative(
labels, losses)
loss_dict['midn_cross_entropy_loss'] = tf.reduce_mean(
utils.masked_avg(data=losses, mask=loss_masks, dim=1))
else:
raise ValueError('Invalid negative mining method.')
# Triplet loss
if options.triplet_loss_weight > 0:
(image_id, image_ids_gathered,
similarity) = (predictions[NOD3Predictions.image_id],
predictions[NOD3Predictions.image_id],
predictions[NOD3Predictions.similarity])
distance = 1.0 - similarity
pos_mask = tf.cast(
tf.equal(tf.expand_dims(image_id, axis=1),
tf.expand_dims(image_ids_gathered, axis=0)),
tf.float32)
neg_mask = 1.0 - pos_mask
distance_ap = utils.masked_maximum(distance, pos_mask)
if options.triplet_loss_use_semihard:
# Use the semihard.
# negatives_outside: smallest D_an where D_an > D_ap.
mask = tf.cast(tf.greater(distance, distance_ap),
tf.float32)
mask = mask * neg_mask
negatives_outside = utils.masked_minimum(distance, mask)
# negatives_inside: largest D_an.
negatives_inside = utils.masked_maximum(distance, neg_mask)
# distance_an: the semihard negatives.
mask_condition = tf.greater(
tf.reduce_sum(mask, axis=1, keepdims=True), 0.0)
distance_an = tf.where(mask_condition, negatives_outside,
negatives_inside)
else:
# Use the hardest.
distance_an = utils.masked_minimum(distance, neg_mask)
losses = tf.maximum(
distance_ap - distance_an + options.triplet_loss_margin, 0)
num_loss_examples = tf.count_nonzero(losses, dtype=tf.float32)
triplet_loss = tf.reduce_mean(losses)
loss_dict['triplet_loss'] = tf.multiply(
triplet_loss, options.triplet_loss_weight)
# Losses of the online instance classifier refinement network.
(num_proposals,
proposals) = (predictions[DetectionResultFields.num_proposals],
predictions[DetectionResultFields.proposal_boxes])
batch, max_num_proposals, _ = utils.get_tensor_shape(proposals)
proposal_scores_0 = predictions[
NOD3Predictions.oicr_proposal_scores + '_at_0']
if options.oicr_use_proba_r_given_c:
proposal_scores_0 = predictions[
NOD3Predictions.midn_proba_r_given_c]
proposal_scores_0 = tf.concat([
tf.fill([batch, max_num_proposals, 1], 0.0), proposal_scores_0
],
axis=-1)
global_step = tf.train.get_or_create_global_step()
oicr_loss_mask = tf.cast(global_step > options.oicr_start_step,
tf.float32)
for i in range(options.oicr_iterations):
proposal_scores_1 = predictions[
NOD3Predictions.oicr_proposal_scores +
'_at_{}'.format(i + 1)]
oicr_cross_entropy_loss_at_i = model_utils.calc_oicr_loss(
labels,
num_proposals,
proposals,
tf.stop_gradient(proposal_scores_0),
proposal_scores_1,
scope='oicr_{}'.format(i + 1),
iou_threshold=options.oicr_iou_threshold)
loss_dict['oicr_cross_entropy_loss_at_{}'.format(
i + 1)] = tf.multiply(
oicr_loss_mask * oicr_cross_entropy_loss_at_i,
options.oicr_loss_weight)
proposal_scores_0 = tf.nn.softmax(proposal_scores_1, axis=-1)
# Min-entropy loss.
mask = tf.sequence_mask(num_proposals,
maxlen=max_num_proposals,
dtype=tf.float32)
proba_r_given_c = predictions[NOD3Predictions.midn_proba_r_given_c]
losses = tf.log(proba_r_given_c + _EPSILON)
losses = tf.squeeze(utils.masked_sum_nd(data=losses,
mask=mask,
dim=1),
axis=1)
min_entropy_loss = tf.reduce_mean(
tf.reduce_sum(losses * labels, axis=1))
min_entropy_loss = tf.multiply(min_entropy_loss,
options.min_entropy_loss_weight)
max_proba = tf.reduce_mean(
utils.masked_maximum(data=proba_r_given_c,
mask=tf.expand_dims(mask, -1),
dim=1))
tf.losses.add_loss(min_entropy_loss)
if options.triplet_loss_weight > 0:
tf.summary.scalar('loss/num_loss_examples', num_loss_examples)
tf.summary.scalar('loss/min_entropy_loss', min_entropy_loss)
tf.summary.scalar('loss/max_proba', max_proba)
return loss_dict
def build_loss(self, predictions, examples, **kwargs):
"""Build tf graph to compute loss.
Args:
predictions: dict of prediction results keyed by name.
examples: dict of inputs keyed by name.
Returns:
loss_dict: dict of loss tensors keyed by name.
"""
options = self._model_proto
loss_dict = {}
with tf.name_scope('losses'):
# Extract image-level labels.
assert options.caption_as_label
vocabulary_list = self._vocabulary_list
mapping = {
'traffic light': 'stoplight',
'fire hydrant': 'hydrant',
'stop sign': 'sign',
'parking meter': 'meter',
'sports ball': 'ball',
'baseball bat': 'bat',
'baseball glove': 'glove',
'tennis racket': 'racket',
'wine glass': 'wineglass',
'hot dog': 'hotdog',
'potted plant': 'plant',
'dining table': 'table',
'cell phone': 'cellphone',
'teddy bear': 'teddy',
'hair drier': 'hairdryer',
}
vocabulary_list = [
mapping.get(cls, cls) for cls in vocabulary_list
]
labels_gt = self._extract_class_label(
class_texts=slim.flatten(
examples[InputDataFields.caption_strings]),
vocabulary_list=vocabulary_list)
examples[NOD4Predictions.debug_groundtruth_labels] = labels_gt
if options.label_strategem == nod4_model_pb2.NOD4Model.EXACTLY_MATCH:
labels = labels_gt
elif options.label_strategem == nod4_model_pb2.NOD4Model.W2V_SYNONYM_MATCH:
labels_ps = self._extract_pseudo_label(
texts=slim.flatten(
examples[InputDataFields.caption_strings]),
vocabulary_list=vocabulary_list,
open_vocabulary_list=self._open_vocabulary_list,
embedding_dims=options.embedding_dims)
select_op = tf.reduce_any(labels_gt > 0, axis=-1)
labels = tf.where(select_op, labels_gt, labels_ps)
labels_ps = tf.where(select_op, tf.zeros_like(labels_ps),
labels_ps)
examples[NOD4Predictions.debug_pseudo_labels] = labels_ps
else:
raise ValueError('Invalid label strategy')
# Loss of the multi-instance detection network.
midn_class_logits = predictions[NOD4Predictions.midn_class_logits]
losses = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels, logits=midn_class_logits)
# Hard-negative mining.
if options.midn_loss_negative_mining == nod4_model_pb2.NOD4Model.NONE:
if options.classification_loss_use_sum:
assert False
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(tf.reduce_sum(losses, axis=-1)),
options.midn_loss_weight)
else:
if options.caption_as_label:
loss_masks = tf.to_float(
tf.reduce_any(labels > 0, axis=-1))
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.squeeze(
utils.masked_avg(tf.reduce_mean(losses,
axis=-1),
mask=loss_masks,
dim=0)),
options.midn_loss_weight)
else:
loss_dict['midn_cross_entropy_loss'] = tf.multiply(
tf.reduce_mean(losses), options.midn_loss_weight)
elif options.midn_loss_negative_mining == nod4_model_pb2.NOD4Model.HARDEST:
assert False
loss_masks = self._midn_loss_mine_hardest_negative(
labels, losses)
loss_dict['midn_cross_entropy_loss'] = tf.reduce_mean(
utils.masked_avg(data=losses, mask=loss_masks, dim=1))
else:
raise ValueError('Invalid negative mining method.')
# Losses of the online instance classifier refinement network.
(num_proposals,
proposals) = (predictions[DetectionResultFields.num_proposals],
predictions[DetectionResultFields.proposal_boxes])
batch, max_num_proposals, _ = utils.get_tensor_shape(proposals)
proposal_scores_0 = predictions[
NOD4Predictions.oicr_proposal_scores + '_at_0']
if options.oicr_use_proba_r_given_c:
proposal_scores_0 = predictions[
NOD4Predictions.midn_proba_r_given_c]
proposal_scores_0 = tf.concat([
tf.fill([batch, max_num_proposals, 1], 0.0), proposal_scores_0
],
axis=-1)
global_step = tf.train.get_or_create_global_step()
oicr_loss_mask = tf.cast(global_step > options.oicr_start_step,
tf.float32)
for i in range(options.oicr_iterations):
proposal_scores_1 = predictions[
NOD4Predictions.oicr_proposal_scores +
'_at_{}'.format(i + 1)]
oicr_cross_entropy_loss_at_i = model_utils.calc_oicr_loss(
labels,
num_proposals,
proposals,
tf.stop_gradient(proposal_scores_0),
proposal_scores_1,
scope='oicr_{}'.format(i + 1),
iou_threshold=options.oicr_iou_threshold)
loss_dict['oicr_cross_entropy_loss_at_{}'.format(
i + 1)] = tf.multiply(
oicr_loss_mask * oicr_cross_entropy_loss_at_i,
options.oicr_loss_weight)
proposal_scores_0 = tf.nn.softmax(proposal_scores_1, axis=-1)
# Min-entropy loss.
mask = tf.sequence_mask(num_proposals,
maxlen=max_num_proposals,
dtype=tf.float32)
proba_r_given_c = predictions[NOD4Predictions.midn_proba_r_given_c]
losses = tf.log(proba_r_given_c + _EPSILON)
losses = tf.squeeze(utils.masked_sum_nd(data=losses,
mask=mask,
dim=1),
axis=1)
min_entropy_loss = tf.reduce_mean(
tf.reduce_sum(losses * labels, axis=1))
min_entropy_loss = tf.multiply(min_entropy_loss,
options.min_entropy_loss_weight)
max_proba = tf.reduce_mean(
utils.masked_maximum(data=proba_r_given_c,
mask=tf.expand_dims(mask, -1),
dim=1))
tf.losses.add_loss(min_entropy_loss)
tf.summary.scalar('loss/min_entropy_loss', min_entropy_loss)
tf.summary.scalar('loss/max_proba', max_proba)
return loss_dict