def testExtractKeypointDescriptor(self):
image = tf.constant(
[[[0, 255, 255], [128, 64, 196]], [[0, 0, 32], [32, 128, 16]]],
dtype=tf.uint8)
# Arbitrary model function used to test ExtractKeypointDescriptor. The
# generated feature_map is a replicated version of the image, concatenated
# with zeros to achieve the required dimensionality. The attention is simply
# the norm of the input image pixels.
def _test_model_fn(image, normalized_image, reuse):
del normalized_image, reuse # Unused variables in the test.
image_shape = tf.shape(image)
attention = tf.squeeze(tf.norm(image, axis=3))
feature_map = tf.concat(
[
tf.tile(image, [1, 1, 1, 341]),
tf.zeros([1, image_shape[1], image_shape[2], 1])
],
axis=3)
return attention, feature_map
boxes, feature_scales, features, scores = (
feature_extractor.ExtractKeypointDescriptor(
image,
layer_name='resnet_v1_50/block3',
image_scales=tf.constant([1.0]),
iou=1.0,
max_feature_num=10,
abs_thres=1.5,
model_fn=_test_model_fn))
exp_boxes = [[-145.0, -145.0, 145.0, 145.0], [-113.0, -145.0, 177.0, 145.0]]
exp_feature_scales = [1.0, 1.0]
exp_features = np.array(
np.concatenate(
(np.tile([[-1.0, 127.0 / 128.0, 127.0 / 128.0], [-1.0, -1.0, -0.75]
], [1, 341]), np.zeros([2, 1])),
axis=1))
exp_scores = [[1.723042], [1.600781]]
with self.test_session() as sess:
boxes_out, feature_scales_out, features_out, scores_out = sess.run(
[boxes, feature_scales, features, scores])
self.assertAllEqual(exp_boxes, boxes_out)
self.assertAllEqual(exp_feature_scales, feature_scales_out)
self.assertAllClose(exp_features, features_out)
self.assertAllClose(exp_scores, scores_out)
def __init__(self, config_path):
tf.logging.set_verbosity(tf.logging.INFO)
# Parse DelfConfig proto.
config = delf_config_pb2.DelfConfig()
self.config = config
with tf.gfile.FastGFile(config_path, 'r') as f:
text_format.Merge(f.read(), config)
# Tell TensorFlow that the model will be built into the default Graph.
with tf.Graph().as_default():
self.byte_value = tf.placeholder(tf.string)
self.image_tf = tf.image.decode_jpeg(self.byte_value, channels=3)
self.input_image = tf.placeholder(tf.uint8, [None, None, 3])
self.input_score_threshold = tf.placeholder(tf.float32)
self.input_image_scales = tf.placeholder(tf.float32, [None])
self.input_max_feature_num = tf.placeholder(tf.int32)
model_fn = feature_extractor.BuildModel(
'resnet_v1_50/block3', 'softplus', 'use_l2_normalized_feature',
1)
boxes, self.feature_scales, features, scores, self.attention = feature_extractor.ExtractKeypointDescriptor(
self.input_image,
layer_name='resnet_v1_50/block3',
image_scales=self.input_image_scales,
iou=1.0,
max_feature_num=self.input_max_feature_num,
abs_thres=self.input_score_threshold,
model_fn=model_fn)
## Optimistic restore.
latest_checkpoint = config.model_path + 'variables/variables'
variables_to_restore = tf.global_variables()
reader = tf.train.NewCheckpointReader(latest_checkpoint)
saved_shapes = reader.get_variable_to_shape_map()
variable_names_to_restore = [
var.name.split(':')[0] for var in variables_to_restore
]
for shape in saved_shapes:
if not shape in variable_names_to_restore:
print(shape)
for var_name in variable_names_to_restore:
if not var_name in saved_shapes:
print(
"WARNING. Saved weight not exists in checkpoint. Init var:",
var_name)
var_names = sorted([(var.name, var.name.split(':')[0])
for var in variables_to_restore
if var.name.split(':')[0] in saved_shapes])
restore_vars = []
with tf.variable_scope('', reuse=True):
for var_name, saved_var_name in var_names:
try:
curr_var = tf.get_variable(saved_var_name)
var_shape = curr_var.get_shape().as_list()
if var_shape == saved_shapes[saved_var_name]:
# print("restore var:", saved_var_name)
restore_vars.append(curr_var)
except ValueError:
print("Ignore due to ValueError on getting var:",
saved_var_name)
saver = tf.train.Saver(restore_vars)
sess = tf.Session()
self.sess = sess
# Initialize variables.
init_op = tf.global_variables_initializer()
sess.run(init_op)
saver.restore(sess, latest_checkpoint)
# # Loading model that will be used.
# tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING],
# config.model_path)
graph = tf.get_default_graph()
self.attention_score = tf.reshape(
scores, [tf.shape(scores)[0]]) # remove extra dim.
self.locations, self.descriptors = feature_extractor.DelfFeaturePostProcessing(
boxes, features, config)