export_gen = data_gen_coco(files,
'export',
batch_size=config['export_batch_size'],
norm=config['normalize'],
**config)
data = export_gen.make_one_shot_iterator().get_next()
H, W, margin = tf.constant(config['preprocessing']['resize'][0]), tf.constant(config['preprocessing']['resize'][1]),\
tf.constant(config['homography_adaptation']['valid_border_margin'])
outputs = homography_adaptation(data['image'], model,
config['homography_adaptation'])
prob = tf.map_fn(
lambda p: box_nms(p,
config['nms'],
min_prob=config['detection_threshold'],
keep_top_k=config['top_k']), outputs['prob'])
prob = tf.image.crop_to_bounding_box(prob[..., tf.newaxis], margin, margin,
H - 2 * margin, W - 2 * margin)
prob = tf.squeeze(tf.image.pad_to_bounding_box(prob, margin, margin, H, W),
axis=-1)
prediction = tf.cast(tf.greater_equal(prob, config['detection_threshold']),
dtype=tf.int32)
tf.keras.backend.get_session().graph.finalize()
while True:
try:
image, name, pred = tf.keras.backend.get_session().run(
[data['image'], data['name'], prediction])
probability1 = tf.image.crop_to_bounding_box(probability1[..., tf.newaxis],
margin, margin, H1 - 2 * margin,
W1 - 2 * margin)
probability1 = tf.squeeze(tf.image.pad_to_bounding_box(probability1, margin,
margin, H1, W1),
axis=-1)
probability2 = tf.image.crop_to_bounding_box(probability2[..., tf.newaxis],
margin, margin, H2 - 2 * margin,
W2 - 2 * margin)
probability2 = tf.squeeze(tf.image.pad_to_bounding_box(probability2, margin,
margin, H2, W2),
axis=-1)
probability1 = tf.map_fn(
lambda p: box_nms(p,
config['nms'],
min_prob=config['detection_threshold'],
keep_top_k=config['top_k']), probability1)
probability2 = tf.map_fn(
lambda p: box_nms(p,
config['nms'],
min_prob=config['detection_threshold'],
keep_top_k=config['top_k']), probability2)
descriptor1 = tf.image.resize_bicubic(
outputs[..., 130:386],
config['grid_size'] * tf.shape(outputs[..., 130:386])[1:3])
descriptor1 = tf.nn.l2_normalize(descriptor1, axis=-1)
descriptor2 = tf.image.resize_bicubic(
outputs[..., 386:],
config['grid_size'] * tf.shape(outputs[..., 386:])[1:3])
descriptor2 = tf.nn.l2_normalize(descriptor2, axis=-1)
export_gen = data_gen_hpatches(files, norm=config['normalize'], **config)
dataset = export_gen.make_one_shot_iterator().get_next()
with tf.device('/cpu:0'):
probability1 = tf.map_fn(
lambda i: tf.py_func(lambda x: classical_detector(x, **config), [i], tf
.float32), dataset['image'])
probability2 = tf.map_fn(
lambda i: tf.py_func(lambda x: classical_detector(x, **config), [i], tf
.float32), dataset['warped_image'])
if config['events']:
probability2 = probability2 * dataset['homography']['mask']
if config['nms']:
probability1 = tf.map_fn(
lambda p: box_nms(
p, config['nms'], min_prob=0., keep_top_k=config['top_k']),
probability1)
probability2 = tf.map_fn(
lambda p: box_nms(
p, config['nms'], min_prob=0., keep_top_k=config['top_k']),
probability2)
tf.keras.backend.get_session().graph.finalize()
i = 0
while True:
try:
if not config['events']:
prob1, prob2, image, warped_image, homography = tf.keras.backend.get_session(
).run([
probability1, probability2, dataset['image'],
pbar = tqdm(total=config['eval_iter'] if config['eval_iter'] > 0 else None)
export_gen = data_gen_hpatches(files, norm=config['normalize'], **config)
dataset = export_gen.make_one_shot_iterator().get_next()
with tf.device('/cpu:0'):
probability1, descriptor1 = tf.map_fn(lambda i: tf.py_func(lambda x: classical_detector_descriptor(x, **config),
[i], (tf.float32, tf.float32)), dataset['image'], [tf.float32, tf.float32])
probability2, descriptor2 = tf.map_fn(lambda i: tf.py_func(lambda x: classical_detector_descriptor(x, **config),
[i], (tf.float32, tf.float32)), dataset['warped_image'], [tf.float32, tf.float32])
if config['events']:
probability2 = probability2 * dataset['homography']['mask']
if config['nms']:
probability1 = tf.map_fn(lambda p: box_nms(p, config['nms'], min_prob=0., keep_top_k=config['top_k']), probability1)
probability2 = tf.map_fn(lambda p: box_nms(p, config['nms'], min_prob=0., keep_top_k=config['top_k']), probability2)
tf.keras.backend.get_session().graph.finalize()
i = 0
while True:
try:
if not config['events']:
prob1, prob2, desc1, desc2, image, warped_image, homography =\
tf.keras.backend.get_session().run([probability1, probability2, descriptor1, descriptor2,
dataset['image'], dataset['warped_image'], dataset['homography']])
else:
prob1, prob2, desc1, desc2, image, warped_image, homography = \
tf.keras.backend.get_session().run([probability1, probability2, descriptor1, descriptor2,
dataset['image'], dataset['warped_image'],