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 test_masked_maximum(self):
tf.reset_default_graph()
data = tf.placeholder(tf.float32, shape=[None, None])
mask = tf.placeholder(tf.float32, shape=[None, None])
masked_maximums = utils.masked_maximum(data, mask)
with self.test_session() as sess:
result = sess.run(masked_maximums,
feed_dict={
data: [[-2.0, 1.0, 2.0, -1.0, 0.0],
[-2.0, -1.0, -3.0, -5.0, -4.0]],
mask: [[1, 1, 1, 1, 1], [1, 1, 1, 1, 1]]
})
self.assertAllClose(result, [[2.0], [-1.0]])
result = sess.run(masked_maximums,
feed_dict={
data: [[-2.0, 1.0, 2.0, -1.0, 0.0],
[-2.0, -1.0, -3.0, -5.0, -4.0]],
mask: [[1, 1, 0, 1, 1], [0, 0, 1, 1, 1]]
})
self.assertAllClose(result, [[1.0], [-3.0]])
result = sess.run(masked_maximums,
feed_dict={
data: [[-2.0, 1.0, 2.0, -1.0, 0.0],
[-2.0, -1.0, -3.0, -5.0, -4.0]],
mask: [[0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]
})
self.assertAllClose(result, [[-2.0], [-5.0]])
def build_prediction(self, examples, **kwargs):
"""Builds tf graph for prediction.
Args:
examples: dict of input tensors keyed by name.
prediction_task: the specific prediction task.
Returns:
predictions: dict of prediction results keyed by name.
"""
options = self._model_proto
is_training = self._is_training
# Text Global-Maximum-Pooling features.
(caption_string,
caption_length) = (examples[InputDataFields.concat_caption_string],
examples[InputDataFields.concat_caption_length])
(caption_token_ids, caption_features) = self._extract_text_feature(
caption_string,
caption_length,
vocabulary_list=self._open_vocabulary_list,
initial_embedding=self._open_vocabulary_initial_embedding,
embedding_dims=options.embedding_dims,
trainable=options.train_word_embedding,
max_norm=None)
with slim.arg_scope(
build_hyperparams(options.text_fc_hyperparams, is_training)):
caption_features = slim.fully_connected(
caption_features,
num_outputs=self._num_classes,
activation_fn=None,
scope='caption')
oov = len(self._open_vocabulary_list)
caption_masks = tf.to_float(
tf.logical_not(tf.equal(caption_token_ids, oov)))
# logits shape = [batch, num_classes].
logits = utils.masked_maximum(data=caption_features,
mask=tf.expand_dims(caption_masks,
axis=-1),
dim=1)
logits = tf.squeeze(logits, axis=1)
predictions = {
TextClassificationPredictions.vocab:
tf.constant(self._vocabulary_list),
TextClassificationPredictions.logits: logits,
}
return predictions
def _predict(self,
text_strings,
text_lengths,
vocabulary_list,
vocabulary_word_embedding,
hidden_units,
output_units,
dropout_keep_proba=1.0,
regularizer=1e-5,
is_training=False):
"""Predicts labels using the texts.
Args:
text_strings: A [batch, num_tokens] string tensor.
text_lengths: A [batch] int tensor.
vocabulary_list: A list of string of length vocab_size.
vocabulary_word_embedding: A [vocab_size, embedding_dims] numpy array.
"""
# Initial embeddings.
init_width = 0.03
oov_emb = init_width * (
np.random.rand(1, vocabulary_word_embedding.shape[-1]) * 2 - 1)
embedding_array_data = np.concatenate([vocabulary_word_embedding, oov_emb],
axis=0)
# Word embedding process.
with tf.name_scope('word_embedding'):
table = tf.contrib.lookup.index_table_from_tensor(
vocabulary_list, num_oov_buckets=1)
embedding_weights = tf.get_variable(
name='weights',
initializer=embedding_array_data.astype(np.float32),
trainable=False) # Freeze the word embedding.
token_ids = table.lookup(text_strings)
token_embs = tf.nn.embedding_lookup(
embedding_weights, token_ids, max_norm=None)
# Multiplayer perceptron.
with tf.variable_scope('text_classifier'):
oov = len(vocabulary_list)
masks = tf.to_float(tf.logical_not(tf.equal(token_ids, oov)))
hiddens = slim.fully_connected(
token_embs,
num_outputs=hidden_units,
activation_fn=None,
trainable=is_training,
weights_regularizer=tf.contrib.layers.l2_regularizer(regularizer),
scope='layer1')
hiddens = utils.masked_maximum(
data=hiddens, mask=tf.expand_dims(masks, axis=-1), dim=1)
hiddens = tf.squeeze(hiddens, axis=1)
hiddens = tf.nn.relu(hiddens)
hiddens = slim.dropout(
hiddens, dropout_keep_proba, is_training=is_training)
logits = slim.fully_connected(
hiddens,
num_outputs=output_units,
activation_fn=None,
trainable=is_training,
weights_regularizer=tf.contrib.layers.l2_regularizer(regularizer),
scope='layer2')
return logits
def extract_labels(self, examples):
"""Extracts the pseudo labels.
Args:
examples: A dictionary involving image-level annotations.
Returns:
labels: A [batch, num_classes] tensor denoting the presence of classes.
"""
init_width = 0.03
embedding_dims = self._open_vocabulary_word_embedding.shape[-1]
classes_to_match = _replace_class_names(self._classes)
# Check if all classes appear in the open-vocabulary.
for class_name in classes_to_match:
if not class_name in self._open_vocabulary_list:
raise ValueError('Class %s has no vector representation.' % class_name)
with tf.name_scope('word_vector_match_extractor'):
# Create hash table and word embedding weights.
table = tf.contrib.lookup.index_table_from_tensor(
self._open_vocabulary_list, num_oov_buckets=1)
oov_emb = init_width * (np.random.rand(1, embedding_dims) * 2 - 1)
embedding_array_data = np.concatenate(
[self._open_vocabulary_word_embedding, oov_emb], axis=0)
embedding_weights = tf.get_variable(
name='weights',
initializer=embedding_array_data.astype(np.float32),
trainable=False) # Freeze the word embedding.
# Lookup to get the class/token embeddings.
class_embs = tf.nn.embedding_lookup(
embedding_weights,
table.lookup(tf.constant(classes_to_match)),
max_norm=None)
token_ids = table.lookup(examples[InputDataFields.concat_caption_string])
token_embs = tf.nn.embedding_lookup(
embedding_weights, token_ids, max_norm=None)
# Compute token-to-class similarity and apply max-pooling.
# Max-pooling: i.e., treat the top-1 as a match.
# similarity shape = [batch, max_num_tokens, num_classes].
# similarity_pooled shape = [batch, num_classes]
batch, num_tokens = utils.get_tensor_shape(
examples[InputDataFields.concat_caption_string])
similarity = self._cosine_similarity(class_embs, token_embs)
oov = len(self._open_vocabulary_list)
mask = tf.not_equal(token_ids, oov)
similarity_pooled = utils.masked_maximum(
data=similarity,
mask=tf.expand_dims(tf.to_float(mask), axis=-1),
dim=1)
similarity_pooled = tf.squeeze(similarity_pooled, axis=1)
labels_most_similar = tf.one_hot(
indices=tf.argmax(similarity_pooled, axis=-1),
depth=self.num_classes,
dtype=tf.float32)
labels_most_similar = tf.where(
tf.reduce_any(mask, axis=-1),
x=labels_most_similar,
y=tf.zeros(shape=[batch, self.num_classes]))
# Consider the exact match.
labels_exact_match = _match_labels(
class_texts=examples[InputDataFields.concat_caption_string],
vocabulary_list=classes_to_match)
return tf.where(
tf.reduce_any(labels_exact_match > 0, axis=-1),
x=labels_exact_match,
y=labels_most_similar)