def main(unused_argv):
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
with tf.Graph().as_default():
image_id, spatial_features, input_seq, target_seq, indicator = (
import_mscoco(mode="train",
batch_size=BATCH_SIZE,
num_epochs=100,
is_mini=True))
visual_sentinel_cell = VisualSentinelCell(300)
image_captioner = ImageCaptioner(visual_sentinel_cell, vocab,
pretrained_matrix)
logits, ids = image_captioner(lengths=tf.reduce_sum(indicator, axis=1),
spatial_image_features=spatial_features,
seq_inputs=input_seq)
tf.losses.sparse_softmax_cross_entropy(target_seq,
logits,
weights=indicator)
loss = tf.losses.get_total_loss()
learning_step = tf.train.GradientDescentOptimizer(1.0).minimize(
loss, var_list=image_captioner.variables)
captioner_saver = tf.train.Saver(var_list=image_captioner.variables)
captioner_ckpt, captioner_ckpt_name = get_visual_sentinel_checkpoint()
with tf.Session() as sess:
sess.run(tf.variables_initializer(image_captioner.variables))
if captioner_ckpt is not None:
captioner_saver.restore(sess, captioner_ckpt)
captioner_saver.save(sess, captioner_ckpt_name)
last_save = time.time()
for i in itertools.count():
time_start = time.time()
try:
_ids, _loss, _learning_step = sess.run(
[ids, loss, learning_step])
except:
break
print(
PRINT_STRING.format(
i, _loss,
list_of_ids_to_string(_ids[0, :].tolist(), vocab),
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)
last_save = new_save
captioner_saver.save(sess, captioner_ckpt_name)
print("Finishing training.")
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)
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)
image_captioner = ImageCaptioner(ShowAttendAndTellCell(300),
vocab,
pretrained_matrix,
trainable=False,
beam_size=FLAGS.beam_size)
logits, ids = image_captioner(spatial_image_features=spatial_features)
captioner_saver = tf.train.Saver(
var_list=remap_decoder_name_scope(image_captioner.variables))
captioner_ckpt, captioner_ckpt_name = get_show_attend_and_tell_checkpoint(
)
with tf.Session() as sess:
def main(unused_argv):
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
with tf.Graph().as_default():
image_id, spatial_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)
image_captioner = ImageCaptioner(SpatialAttentionCell(300), vocab,
pretrained_matrix)
logits, ids = image_captioner(lengths=tf.reduce_sum(indicator, axis=1),
spatial_image_features=spatial_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=image_captioner.variables,
global_step=global_step)
captioner_saver = tf.train.Saver(var_list=image_captioner.variables +
[global_step])
captioner_ckpt, captioner_ckpt_name = get_spatial_attention_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(image_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:
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_train_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}"""
BATCH_SIZE = 10
BEAM_SIZE = 16
if __name__ == "__main__":
vocab, pretrained_matrix = load_glove(vocab_size=100000,
embedding_size=300)
with tf.Graph().as_default():
image_id, spatial_features, input_seq, target_seq, indicator = (
import_mscoco(mode="train",
batch_size=BATCH_SIZE,
num_epochs=1,
is_mini=True))
image_captioner = ImageCaptioner(ShowAttendAndTellCell(300),
vocab,
pretrained_matrix,
trainable=False,
beam_size=BEAM_SIZE)
logits, ids = image_captioner(spatial_image_features=spatial_features)
captioner_saver = tf.train.Saver(
var_list=remap_decoder_name_scope(image_captioner.variables))
captioner_ckpt, captioner_ckpt_name = get_show_attend_and_tell_checkpoint(
)
with tf.Session() as sess:
assert (captioner_ckpt is not None)
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, spatial_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)
show_attend_and_tell_cell = ShowAttendAndTellCell(300, num_image_features=2048)
attribute_image_captioner = AttributeImageCaptioner(
show_attend_and_tell_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(top_k_attributes,
lengths=tf.reduce_sum(indicator, axis=1),
spatial_image_features=spatial_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_show_attend_and_tell_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.")