def featuremap_select(input, percentage, name='FeaturemapSelect', layer_collector=None):
l = []
for input_i in input:
total_dim = input_i.get_shape().as_list()[-1]
target_dim = int(total_dim * percentage)
target_idxs = [i for i in range(total_dim)]
target_idxs = np.random.choice(target_idxs, target_dim, replace=False)
target_idxs = [[i] for i in target_idxs]
input_i_t = tf.transpose(input_i, [3,1,2,0])
input_i_s = tf.gather_nd(input_i_t, target_idxs)
l.append(tf.transpose(input_i_s, [3,1,2,0]))
safe_append(layer_collector, l, name)
return l
def combine_score_layer(input,
shape,
name='CombineScore',
layer_collector=None,
param_collector=None):
input_shape = tf.shape(input)
l = tf.transpose(tf.reshape(tf.transpose(input, [0, 3, 1, 2]), [
input_shape[0],
int(shape),
tf.cast(
tf.cast(input_shape[1], tf.float32) / tf.cast(shape, tf.float32) *
tf.cast(input_shape[3], tf.float32), tf.int32), input_shape[2]
]), [0, 2, 3, 1],
name=name)
safe_append(layer_collector, l, name)
return l
def proposal_layer(
input,
feature_stride,
anchor_scales,
name='ProposalLayer',
layer_collector=None,
):
with tf.variable_scope(name) as scope:
l = tf.py_func(_proposal_layer, [
input[0],
input[1],
input[2],
input[3],
feature_stride,
anchor_scales,
], tf.float32)
safe_append(layer_collector, l, name)
return l
def anchor_target_layer(
input,
feature_stride,
anchor_scales,
name='AnchorLayer',
layer_collector=None,
):
with tf.variable_scope(name) as scope:
layers = tf.py_func(_anchor_target_layer, [
input[0],
input[1],
input[2],
input[3],
feature_stride,
anchor_scales,
], [tf.float32, tf.float32, tf.float32, tf.float32])
for layer in layers:
safe_append(layer_collector, layer, name)
return layers