def _cnn_lstm(self, inputs, state, feature_prev, sum_kernel, config, reuse_variables, is_training = True, lstm=True):
def make_cell():
cell = self._get_lstm_cell(config, is_training)
if is_training and config.keep_prob < 1:
cell = tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=config.keep_prob)
return cell
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
# instead of using the last feature layer, we want to bring recurrency on other layers
#
_, _, feature = modelEAST.model(inputs, is_training=False)
if feature_prev is not None:
# add correlation computation to state, so the state is updated;
cvt = cost_volume(feature_prev, feature, 5)
feature_
if lstm:
cell = tf.contrib.rnn.MultiRNNCell(
[make_cell() for _ in range(config.num_layers)], state_is_tuple=True)
self._initial_state = cell.zero_state(config.batch_size, tf.float32)
if state is None:
state = self._initial_state
output, state = cell(feature, state)
# F_heatmap = slim.conv2d(output, 1, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None)
with tf.variable_scope('pred_module', reuse=reuse_variables):
F_score = slim.conv2d(output, 1, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None)
# 4 channel of axis aligned bbox and 1 channel rotation angle
geo_map = slim.conv2d(output, 4, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None) * 512
angle_map = (slim.conv2d(output, 1, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None) - 0.5) * np.pi/2 # angle is between [-45, 45]
F_geometry = tf.concat([geo_map, angle_map], axis=-1)
feature_prev = feature
return F_score, F_geometry, state, feature_prev
else:
pass
def _cnn_lstm(self,
inputs,
state,
config,
reuse_variables,
is_training=True,
lstm=True):
def make_cell():
cell = self._get_lstm_cell(config, is_training)
if is_training and config.keep_prob < 1:
cell = tf.contrib.rnn.DropoutWrapper(
cell, output_keep_prob=config.keep_prob)
return cell
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
_, _, feature = modelEAST.model(inputs, is_training=False)
# by xiaolong Sep. 13th
# feature_stoped = tf.stop_gradient(feature)
if lstm:
cell = tf.contrib.rnn.MultiRNNCell(
[make_cell() for _ in range(config.num_layers)],
state_is_tuple=True)
self._initial_state = cell.zero_state(config.batch_size,
tf.float32)
if state is None:
state = self._initial_state
output, state = cell(feature, state)
# F_heatmap = slim.conv2d(output, 1, 1, activation_fn=tf.nn.sigmoid, normalizer_fn=None)
with tf.variable_scope('pred_module', reuse=reuse_variables):
F_score = slim.conv2d(output,
1,
1,
activation_fn=tf.nn.sigmoid,
normalizer_fn=None)
# 4 channel of axis aligned bbox and 1 channel rotation angle
geo_map = slim.conv2d(output,
4,
1,
activation_fn=tf.nn.sigmoid,
normalizer_fn=None) * 512
angle_map = (slim.conv2d(output,
1,
1,
activation_fn=tf.nn.sigmoid,
normalizer_fn=None) -
0.5) * np.pi / 2 # angle is between [-45, 45]
F_geometry = tf.concat([geo_map, angle_map], axis=-1)
return F_score, F_geometry, state
else:
pass