def _process_attributes(images, vocab, tagger):
"""Processes a list of images.
Args:
images: a list containing BestFirstMetadata objects.
vocab: a Vocabulary object.
tagger: a Tagger object from nltk.
Returns:
a list of AttributeMetadata objects.
"""
attribute_map = get_visual_attributes()
attribute_images = []
for image in images:
caption = image.captions[0]
attributes = attribute_map.sentence_to_attributes(caption)
attribute_images.append(AttributeMetadata(
image_id=image.image_id,
filename=image.filename,
captions=image.captions,
image_features=image.image_features,
object_features=image.object_features,
running_ids=image.running_ids,
running_ids_splits=image.running_ids_splits,
word_ids=image.word_ids,
pointer_ids=image.pointer_ids,
attributes=attributes))
return attribute_images
def __init__(self):
"""Creates handles to the TensorFlow computational graph."""
# TensorFlow ops for JPEG decoding.
self.encoded_jpeg = tf.placeholder(dtype=tf.string)
self.decoded_jpeg = tf.image.decode_jpeg(self.encoded_jpeg, channels=3)
self.decoded_jpeg = tf.image.resize_images(self.decoded_jpeg, [
FLAGS.image_height, FLAGS.image_width])
# Create the model to extract image boxes
self.box_extractor = BoxExtractor(get_faster_rcnn_config(), trainable=False)
self.image_tensor = tf.placeholder(tf.float32, name='image_tensor', shape=[None,
FLAGS.image_height, FLAGS.image_width, 3])
self.boxes, self.scores, self.cropped_images = self.box_extractor(self.image_tensor)
# Create a single TensorFlow Session for all image decoding calls.
self.sess = tf.Session()
self.rcnn_saver = tf.train.Saver(var_list=self.box_extractor.variables)
self.rcnn_saver.restore(self.sess, get_faster_rcnn_checkpoint())
self.lock = threading.Lock()
self.attribute_map = get_visual_attributes()
def __init__(self):
"""Creates handles to the TensorFlow computational graph."""
# TensorFlow ops for JPEG decoding.
self.encoded_jpeg = tf.placeholder(dtype=tf.string)
self.decoded_jpeg = tf.image.decode_jpeg(self.encoded_jpeg, channels=3)
self.decoded_jpeg = tf.image.resize_images(
self.decoded_jpeg, [FLAGS.image_height, FLAGS.image_width])
# Create the model to extract image boxes
self.box_extractor = BoxExtractor(get_faster_rcnn_config(),
trainable=False)
self.image_tensor = tf.placeholder(
tf.float32,
name='image_tensor',
shape=[None, FLAGS.image_height, FLAGS.image_width, 3])
self.boxes, self.scores, self.cropped_images = self.box_extractor(
self.image_tensor)
# Create the model to extract the image features
self.feature_extractor = FeatureExtractor(is_training=False,
global_pool=False)
# Compute the mean ResNet-101 features
self.image_features = tf.reduce_mean(
self.feature_extractor(self.image_tensor), [1, 2])
feature_batch = tf.shape(self.image_features)[0]
feature_depth = tf.shape(self.image_features)[1]
self.object_features = tf.reduce_mean(
self.feature_extractor(self.cropped_images), [1, 2])
self.object_features = tf.reshape(self.object_features,
[feature_batch, 100, feature_depth])
# Create a single TensorFlow Session for all image decoding calls.
self.sess = tf.Session()
rcnn_saver = tf.train.Saver(var_list=self.box_extractor.variables)
resnet_saver = tf.train.Saver(
var_list=self.feature_extractor.variables)
rcnn_saver.restore(self.sess, get_faster_rcnn_checkpoint())
resnet_saver.restore(self.sess, get_resnet_v2_101_checkpoint())
self.lock = threading.Lock()
self.attribute_map = get_visual_attributes()
def main(unused_argv):
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
attribute_map, attribute_embeddings_map = get_visual_attributes(
), np.random.normal(0, 0.1, [1000, 2048])
with tf.Graph().as_default():
image_id, mean_features, object_features, input_seq, target_seq, indicator = import_mscoco(
mode="train",
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
is_mini=FLAGS.is_mini)
up_down_cell = UpDownCell(300, num_image_features=4096)
attribute_image_captioner = AttributeImageCaptioner(
up_down_cell, vocab, pretrained_matrix, attribute_map,
attribute_embeddings_map)
attribute_detector = AttributeDetector(1000)
_, top_k_attributes = attribute_detector(mean_features)
logits, ids = attribute_image_captioner(
top_k_attributes,
lengths=tf.reduce_sum(indicator, axis=1),
mean_image_features=mean_features,
mean_object_features=object_features,
seq_inputs=input_seq)
tf.losses.sparse_softmax_cross_entropy(target_seq,
logits,
weights=indicator)
loss = tf.losses.get_total_loss()
global_step = tf.train.get_or_create_global_step()
optimizer = tf.train.AdamOptimizer()
learning_step = optimizer.minimize(
loss,
var_list=attribute_image_captioner.variables,
global_step=global_step)
captioner_saver = tf.train.Saver(
var_list=attribute_image_captioner.variables + [global_step])
attribute_detector_saver = tf.train.Saver(
var_list=attribute_detector.variables)
captioner_ckpt, captioner_ckpt_name = get_up_down_attribute_checkpoint(
)
attribute_detector_ckpt, attribute_detector_ckpt_name = get_attribute_detector_checkpoint(
)
with tf.Session() as sess:
sess.run(tf.variables_initializer(optimizer.variables()))
if captioner_ckpt is not None:
captioner_saver.restore(sess, captioner_ckpt)
else:
sess.run(
tf.variables_initializer(
attribute_image_captioner.variables + [global_step]))
if attribute_detector_ckpt is not None:
attribute_detector_saver.restore(sess, attribute_detector_ckpt)
else:
sess.run(tf.variables_initializer(
attribute_detector.variables))
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = time.time()
for i in itertools.count():
time_start = time.time()
try:
_target, _ids, _loss, _learning_step = sess.run(
[target_seq, ids, loss, learning_step])
except:
break
iteration = sess.run(global_step)
print(
PRINT_STRING.format(
iteration, _loss,
list_of_ids_to_string(_ids[0, :].tolist(), vocab),
list_of_ids_to_string(_target[0, :].tolist(), vocab),
FLAGS.batch_size / (time.time() - time_start)))
new_save = time.time()
if new_save - last_save > 3600: # save the model every hour
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = new_save
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
print("Finishing training.")
from detailed_captioning.utils import get_val_annotations_file
from detailed_captioning.inputs.spatial_image_features_only import import_mscoco
PRINT_STRING = """({3:.2f} img/sec) iteration: {0:05d}\n caption: {1}\n label: {2}"""
tf.logging.set_verbosity(tf.logging.INFO)
tf.flags.DEFINE_integer("batch_size", 1, "")
tf.flags.DEFINE_integer("beam_size", 3, "")
tf.flags.DEFINE_boolean("is_mini", False, "")
tf.flags.DEFINE_string("mode", "eval", "")
FLAGS = tf.flags.FLAGS
if __name__ == "__main__":
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
attribute_map, attribute_embeddings_map = get_visual_attributes(
), np.random.normal(0, 0.1, [1000, 2048])
with tf.Graph().as_default():
image_id, spatial_features, input_seq, target_seq, indicator = import_mscoco(
mode=FLAGS.mode,
batch_size=FLAGS.batch_size,
num_epochs=1,
is_mini=FLAGS.is_mini)
visual_sentinel_cell = VisualSentinelCell(300, num_image_features=2048)
attribute_image_captioner = AttributeImageCaptioner(
visual_sentinel_cell, vocab, pretrained_matrix, attribute_map,
attribute_embeddings_map)
attribute_detector = AttributeDetector(1000)
_, top_k_attributes = attribute_detector(
tf.reduce_mean(spatial_features, [1, 2]))
logits, ids = attribute_image_captioner(
def main(unused_argv):
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
attribute_map = get_visual_attributes()
attribute_to_word_lookup_table = vocab.word_to_id(
attribute_map.reverse_vocab)
with tf.Graph().as_default():
(image_id, image_features, object_features, input_seq, target_seq,
indicator, attributes) = import_mscoco(mode="train",
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
is_mini=FLAGS.is_mini)
attribute_detector = AttributeDetector(1000)
_, image_attributes, object_attributes = attribute_detector(
image_features, object_features)
grounded_attribute_cell = GroundedAttributeCell(1024)
attribute_captioner = AttributeCaptioner(
grounded_attribute_cell, vocab, pretrained_matrix,
attribute_to_word_lookup_table)
logits, ids = attribute_captioner(lengths=tf.reduce_sum(indicator,
axis=1),
mean_image_features=image_features,
mean_object_features=object_features,
seq_inputs=input_seq,
image_attributes=image_attributes,
object_attributes=object_attributes)
tf.losses.sparse_softmax_cross_entropy(target_seq,
logits,
weights=indicator)
loss = tf.losses.get_total_loss()
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.train.exponential_decay(5e-4,
global_step,
3 * 586363 //
FLAGS.batch_size,
0.8,
staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
learning_step = optimizer.minimize(
loss,
var_list=attribute_captioner.variables,
global_step=global_step)
detector_saver = tf.train.Saver(var_list=attribute_detector.variables +
[global_step])
detector_ckpt, detector_ckpt_name = get_attribute_detector_checkpoint()
captioner_saver = tf.train.Saver(
var_list=attribute_captioner.variables + [global_step])
captioner_ckpt, captioner_ckpt_name = get_grounded_attribute_checkpoint(
)
with tf.Session() as sess:
sess.run(tf.variables_initializer(optimizer.variables()))
if detector_ckpt is not None:
detector_saver.restore(sess, detector_ckpt)
else:
sess.run(
tf.variables_initializer(attribute_detector.variables +
[global_step]))
if captioner_ckpt is not None:
captioner_saver.restore(sess, captioner_ckpt)
else:
sess.run(
tf.variables_initializer(attribute_captioner.variables +
[global_step]))
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = time.time()
for i in itertools.count():
time_start = time.time()
try:
_target, _ids, _loss, _learning_step = sess.run(
[target_seq, ids, loss, learning_step])
except:
break
iteration = sess.run(global_step)
print(
PRINT_STRING.format(
iteration, _loss,
list_of_ids_to_string(_ids[0, :].tolist(), vocab),
list_of_ids_to_string(_target[0, :].tolist(), vocab),
FLAGS.batch_size / (time.time() - time_start)))
new_save = time.time()
if new_save - last_save > 3600: # save the model every hour
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = new_save
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
print("Finishing training.")
def main(unused_argv):
def _is_valid_num_shards(num_shards):
"""Returns True if num_shards is compatible with FLAGS.num_threads."""
return num_shards < FLAGS.num_threads or not num_shards % FLAGS.num_threads
assert _is_valid_num_shards(FLAGS.train_shards), (
"Please make the FLAGS.num_threads commensurate with FLAGS.train_shards"
)
assert _is_valid_num_shards(FLAGS.val_shards), (
"Please make the FLAGS.num_threads commensurate with FLAGS.val_shards")
assert _is_valid_num_shards(FLAGS.test_shards), (
"Please make the FLAGS.num_threads commensurate with FLAGS.test_shards"
)
# Create vocabulary from the glove embeddings.
attribute_map = get_visual_attributes()
if not tf.gfile.IsDirectory(FLAGS.output_dir):
tf.gfile.MakeDirs(FLAGS.output_dir)
# Load image metadata from caption files.
mscoco_train_dataset = _load_and_process_metadata(
FLAGS.train_captions_file, FLAGS.train_image_dir)
mscoco_val_dataset = _load_and_process_metadata(FLAGS.val_captions_file,
FLAGS.val_image_dir)
# Redistribute the MSCOCO data as follows:
# train_dataset = 99% of mscoco_train_dataset
# val_dataset = 1% of mscoco_train_dataset (for validation during training).
# test_dataset = 100% of mscoco_val_dataset (for final evaluation).
train_cutoff = int(0.99 * len(mscoco_train_dataset))
train_dataset = mscoco_train_dataset[:train_cutoff]
val_dataset = mscoco_train_dataset[train_cutoff:]
test_dataset = mscoco_val_dataset
# If needed crop the dataset to make it smaller
max_train_size = len(train_dataset)
if FLAGS.train_dataset_size < max_train_size:
# Shuffle the ordering of images. Make the randomization repeatable.
random.seed(12345)
random.shuffle(train_dataset)
train_dataset = train_dataset[:FLAGS.train_dataset_size]
max_val_size = len(val_dataset)
if FLAGS.val_dataset_size < max_val_size:
# Shuffle the ordering of images. Make the randomization repeatable.
random.seed(12345)
random.shuffle(val_dataset)
val_dataset = val_dataset[:FLAGS.val_dataset_size]
max_test_size = len(test_dataset)
if FLAGS.test_dataset_size < max_test_size:
# Shuffle the ordering of images. Make the randomization repeatable.
random.seed(12345)
random.shuffle(test_dataset)
test_dataset = test_dataset[:FLAGS.test_dataset_size]
# Create the model to extract image features
image_tensor = tf.placeholder(
tf.float32,
name='image_tensor',
shape=[None, FLAGS.image_height, FLAGS.image_width, 3])
feature_extractor = FeatureExtractor(is_training=False, global_pool=False)
# Compute the ResNet-101 features
image_features = tf.reduce_mean(feature_extractor(image_tensor), [1, 2])
with tf.Session() as sess:
resnet_saver = tf.train.Saver(var_list=feature_extractor.variables)
resnet_saver.restore(sess, get_resnet_v2_101_checkpoint())
lock = threading.Lock()
def run_model_fn(images):
lock.acquire()
r = sess.run(image_features, feed_dict={image_tensor: images})
lock.release()
return r
_process_dataset("train", train_dataset, attribute_map,
FLAGS.train_shards, run_model_fn)
_process_dataset("val", val_dataset, attribute_map, FLAGS.val_shards,
run_model_fn)
_process_dataset("test", test_dataset, attribute_map,
FLAGS.test_shards, run_model_fn)
def main(unused_argv):
attribute_map = get_visual_attributes()
with tf.Graph().as_default():
(image_id, image_features, object_features, input_seq, target_seq,
indicator, attributes) = import_mscoco(mode="train",
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
is_mini=FLAGS.is_mini)
attribute_detector = AttributeDetector(1000)
logits, image_detections, object_detections = attribute_detector(
image_features, object_features)
tf.losses.sigmoid_cross_entropy(attributes, logits)
loss = tf.losses.get_total_loss()
global_step = tf.train.get_or_create_global_step()
optimizer = tf.train.AdamOptimizer()
learning_step = optimizer.minimize(
loss,
var_list=attribute_detector.variables,
global_step=global_step)
captioner_saver = tf.train.Saver(
var_list=attribute_detector.variables + [global_step])
captioner_ckpt, captioner_ckpt_name = get_attribute_detector_checkpoint(
)
with tf.Session() as sess:
sess.run(tf.variables_initializer(optimizer.variables()))
if captioner_ckpt is not None:
captioner_saver.restore(sess, captioner_ckpt)
else:
sess.run(
tf.variables_initializer(attribute_detector.variables +
[global_step]))
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = time.time()
for i in itertools.count():
time_start = time.time()
try:
_attributes, _detections, _loss, _ = sess.run(
[attributes, image_detections, loss, learning_step])
except:
break
iteration = sess.run(global_step)
ground_truth_ids = np.where(_attributes[0, :] > 0.5)
print(
PRINT_STRING.format(
FLAGS.batch_size / (time.time() - time_start),
iteration,
_loss,
str(
attribute_map.id_to_word(
_detections[0, :].tolist())),
str(
attribute_map.id_to_word(
ground_truth_ids[0].tolist())),
))
new_save = time.time()
if new_save - last_save > 3600: # save the model every hour
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
last_save = new_save
captioner_saver.save(sess,
captioner_ckpt_name,
global_step=global_step)
print("Finishing training.")
from detailed_captioning.inputs.captions_and_attributes import import_mscoco
PRINT_STRING = """({3:.2f} img/sec) iteration: {0:05d}\n caption: {1}\n label: {2}"""
tf.logging.set_verbosity(tf.logging.INFO)
tf.flags.DEFINE_integer("batch_size", 1, "")
tf.flags.DEFINE_integer("beam_size", 3, "")
tf.flags.DEFINE_boolean("is_mini", False, "")
tf.flags.DEFINE_string("mode", "eval", "")
FLAGS = tf.flags.FLAGS
if __name__ == "__main__":
vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300)
attribute_map = get_visual_attributes()
attribute_to_word_lookup_table = vocab.word_to_id(attribute_map.reverse_vocab)
with tf.Graph().as_default():
(image_id, image_features, object_features, input_seq, target_seq, indicator,
attributes) = import_mscoco(
mode=FLAGS.mode, batch_size=FLAGS.batch_size, num_epochs=1, is_mini=FLAGS.is_mini)
attribute_detector = AttributeDetector(1000)
_, image_attributes, object_attributes = attribute_detector(image_features, object_features)
grounded_attribute_cell = GroundedAttributeCell(1024)
attribute_captioner = AttributeCaptioner(grounded_attribute_cell, vocab, pretrained_matrix,
attribute_to_word_lookup_table,
trainable=False, beam_size=FLAGS.beam_size)
