def __init__(self, name, inputs, tower_setup, n_convs=2, n_features=None, dilations=None, strides=None,
filter_size=None, activation="relu", batch_norm_decay=BATCH_NORM_DECAY_DEFAULT, l2=L2_DEFAULT):
super(ResidualUnit2, self).__init__()
# TODO: dropout
curr, n_features_inp = prepare_input(inputs)
res = curr
assert n_convs >= 1, n_convs
if dilations is not None:
assert strides is None
elif strides is None:
strides = [[1, 1]] * n_convs
if filter_size is None:
filter_size = [[3, 3]] * n_convs
if n_features is None:
n_features = n_features_inp
if not isinstance(n_features, list):
n_features = [n_features] * n_convs
with tf.variable_scope(name):
curr = self.create_and_apply_batch_norm(curr, n_features_inp, batch_norm_decay, tower_setup, "bn0")
curr = get_activation(activation)(curr)
if strides is None:
strides_res = [1, 1]
else:
strides_res = numpy.prod(strides, axis=0).tolist()
if (n_features[-1] != n_features_inp) or (strides_res != [1, 1]):
W0 = self.create_weight_variable("W0", [1, 1] + [n_features_inp, n_features[-1]], l2, tower_setup)
if dilations is None:
res = conv2d(curr, W0, strides_res)
else:
res = conv2d(curr, W0)
W1 = self.create_weight_variable("W1", filter_size[0] + [n_features_inp, n_features[0]], l2, tower_setup)
if dilations is None:
curr = conv2d(curr, W1, strides[0])
else:
curr = conv2d_dilated(curr, W1, dilations[0])
for idx in range(1, n_convs):
curr = self.create_and_apply_batch_norm(curr, n_features[idx - 1], batch_norm_decay,
tower_setup, "bn" + str(idx + 1))
curr = get_activation(activation)(curr)
Wi = self.create_weight_variable("W" + str(idx + 1),
filter_size[idx] + [n_features[idx - 1], n_features[idx]],
l2, tower_setup)
if dilations is None:
curr = conv2d(curr, Wi, strides[idx])
else:
curr = conv2d_dilated(curr, Wi, dilations[idx])
curr += res
self.outputs = [curr]
def __init__(self, name, inputs, n_features, tower_setup, old_order=False, filter_size=(3, 3),
strides=(1, 1), dilation=None, pool_size=(1, 1), pool_strides=None, activation="relu", dropout=0.0,
batch_norm=False, bias=False, batch_norm_decay=BATCH_NORM_DECAY_DEFAULT, l2=L2_DEFAULT):
super(Conv, self).__init__()
# mind the order of dropout, conv, activation and batchnorm!
# default: batchnorm -> activation -> dropout -> conv -> pool
# if old_order: dropout -> conv -> batchnorm -> activation -> pool
curr, n_features_inp = prepare_input(inputs)
filter_size = list(filter_size)
strides = list(strides)
pool_size = list(pool_size)
if pool_strides is None:
pool_strides = pool_size
with tf.variable_scope(name):
W = self.create_weight_variable("W", filter_size + [n_features_inp, n_features], l2, tower_setup)
b = None
if bias:
b = self.create_bias_variable("b", [n_features], tower_setup)
if old_order:
curr = apply_dropout(curr, dropout)
if dilation is None:
curr = conv2d(curr, W, strides)
else:
curr = conv2d_dilated(curr, W, dilation)
if bias:
curr += b
if batch_norm:
curr = self.create_and_apply_batch_norm(curr, n_features, batch_norm_decay, tower_setup)
curr = get_activation(activation)(curr)
else:
if batch_norm:
curr = self.create_and_apply_batch_norm(curr, n_features_inp, batch_norm_decay, tower_setup)
curr = get_activation(activation)(curr)
curr = apply_dropout(curr, dropout)
if dilation is None:
curr = conv2d(curr, W, strides)
else:
curr = conv2d_dilated(curr, W, dilation)
if bias:
curr += b
if pool_size != [1, 1]:
curr = max_pool(curr, pool_size, pool_strides)
self.outputs = [curr]
def __init__(self,
name,
inputs,
n_features,
tower_setup,
activation="relu",
dropout=0.0,
batch_norm=False,
batch_norm_decay=BATCH_NORM_DECAY_DEFAULT,
l2=L2_DEFAULT):
super(FullyConnected, self).__init__()
inp, n_features_inp = prepare_collapsed_input_and_dropout(
inputs, dropout)
with tf.variable_scope(name):
if batch_norm:
inp = tf.expand_dims(inp, axis=0)
inp = tf.expand_dims(inp, axis=0)
inp = self.create_and_apply_batch_norm(inp, n_features_inp,
batch_norm_decay,
tower_setup)
inp = tf.squeeze(inp, axis=[0, 1])
W = self.create_weight_variable("W", [n_features_inp, n_features],
l2, tower_setup)
b = self.create_bias_variable("b", [n_features], tower_setup)
z = tf.matmul(inp, W) + b
h = get_activation(activation)(z)
self.outputs = [h]
def __init__(self,
name,
inputs,
tower_setup,
n_features,
concat,
activation="relu",
filter_size=(3, 3),
batch_norm_decay=BATCH_NORM_DECAY_DEFAULT,
l2=L2_DEFAULT):
super(Upsampling, self).__init__()
filter_size = list(filter_size)
assert isinstance(concat, list)
assert len(concat) > 0
curr, n_features_inp = prepare_input(inputs)
concat_inp, n_features_concat = prepare_input(concat)
curr = tf.image.resize_nearest_neighbor(curr,
tf.shape(concat_inp)[1:3])
curr = tf.concat([curr, concat_inp], axis=3)
n_features_curr = n_features_inp + n_features_concat
with tf.variable_scope(name):
W = self.create_weight_variable(
"W", filter_size + [n_features_curr, n_features], l2,
tower_setup)
b = self.create_bias_variable("b", [n_features], tower_setup)
curr = conv2d(curr, W) + b
curr = get_activation(activation)(curr)
self.outputs = [curr]
def __init__(self, name, inputs, tower_setup, activation="relu", batch_norm_decay=BATCH_NORM_DECAY_DEFAULT):
super(Collapse, self).__init__()
curr, n_features_inp = prepare_input(inputs)
with tf.variable_scope(name):
inp = self.create_and_apply_batch_norm(curr, n_features_inp, batch_norm_decay, tower_setup)
h_act = get_activation(activation)(inp)
out = global_avg_pool(h_act)
self.outputs = [out]
def __init__(self, name, inputs, targets, n_classes, void_label, tower_setup, filter_size=(1, 1),
input_activation=None, dilation=None, resize_targets=False, resize_logits=False, loss="ce",
fraction=None, l2=L2_DEFAULT, dropout=0.0):
super(SegmentationSoftmax, self).__init__()
assert targets.get_shape().ndims == 4, targets.get_shape()
assert not (resize_targets and resize_logits)
inp, n_features_inp = prepare_input(inputs)
filter_size = list(filter_size)
with tf.variable_scope(name):
if input_activation is not None:
inp = get_activation(input_activation)(inp)
inp = apply_dropout(inp, dropout)
self.W, self.b, self.W_adjustable, self.b_adjustable, self.n_classes_current, W, b = self.create_weights(
n_classes, filter_size, n_features_inp, l2, tower_setup)
self.adjustable_output_assign_data = self._create_adjustable_output_assign_data(tower_setup)
if self.n_classes_current is None:
self.n_classes_current = n_classes
if dilation is None:
y_pred = conv2d(inp, W) + b
else:
y_pred = conv2d_dilated(inp, W, dilation) + b
self.outputs = [tf.nn.softmax(y_pred, -1, 'softmax')]
if resize_targets:
targets = tf.image.resize_nearest_neighbor(targets, tf.shape(y_pred)[1:3])
if resize_logits:
y_pred = tf.image.resize_images(y_pred, tf.shape(targets)[1:3])
pred = tf.argmax(y_pred, axis=3)
targets = tf.cast(targets, tf.int64)
targets = tf.squeeze(targets, axis=3)
# TODO: Void label is not considered in the iou calculation.
if void_label is not None:
# avoid nan by replacing void label by 0
# note: the loss for these cases is multiplied by 0 below
void_label_mask = tf.equal(targets, void_label)
no_void_label_mask = tf.logical_not(void_label_mask)
targets = tf.where(void_label_mask, tf.zeros_like(targets), targets)
else:
no_void_label_mask = None
self.measures = self.create_measures(n_classes, pred, targets)
self.loss = self.create_loss(loss, fraction, no_void_label_mask, targets, tower_setup, void_label, y_pred)
self.add_scalar_summary(self.loss, "loss")
def __init__(self, name, inputs, targets, n_classes, n_features, tower_setup, activation="linear", dropout=0.0,
batch_norm=False,
batch_norm_decay=BATCH_NORM_DECAY_DEFAULT, l2=L2_DEFAULT):
super(FullyConnectedWithTripletLoss, self).__init__()
self.measures = {}
inp, n_features_inp = prepare_collapsed_input_and_dropout(inputs, dropout)
with tf.variable_scope(name):
if batch_norm:
inp = tf.expand_dims(inp, axis=0)
inp = tf.expand_dims(inp, axis=0)
inp = self.create_and_apply_batch_norm(inp, n_features_inp, batch_norm_decay, tower_setup)
inp = tf.squeeze(inp, axis=[0, 1])
W = self.create_weight_variable("W", [n_features_inp, n_features], l2, tower_setup)
b = self.create_bias_variable("b", [n_features], tower_setup)
z = tf.matmul(inp, W) + b
h = get_activation(activation)(z)
self.outputs = [h]
size = smart_shape(h)[0]
def get_loss(idx):
anchor = h[idx, :]
anchor_class = targets[idx]
class_division = tf.cast(tf.equal(targets, anchor_class), tf.int32)
partitioned_output = tf.dynamic_partition(h, class_division, 2)
# Positives
positive_distances = tf.abs(anchor - partitioned_output[1])
pos_dis_val = tf.norm(positive_distances,axis=1)
hardest_positive_idx = tf.arg_max(pos_dis_val,0)
pos_div_size = smart_shape(partitioned_output[1])[0]
pos_divider = tf.one_hot(hardest_positive_idx,pos_div_size,dtype=tf.int32)
hardest_positive = tf.dynamic_partition(positive_distances,pos_divider,2)[1]
hardest_positive_class = tf.gather(targets, hardest_positive_idx)
hardest_positive = tf.norm(hardest_positive, axis=1)
# Negatives
negative_distances = tf.abs(anchor - partitioned_output[0])
neg_dis_val = tf.norm(negative_distances, axis=1)
hardest_negative_idx = tf.arg_min(neg_dis_val,0)
neg_div_size = smart_shape(partitioned_output[0])[0]
neg_divider = tf.one_hot(hardest_negative_idx,neg_div_size,dtype=tf.int32)
hardest_negative = tf.dynamic_partition(negative_distances,neg_divider,2)[1]
hardest_negative_class = tf.gather(targets,hardest_negative_idx)
hardest_negative = tf.norm(hardest_negative, axis=1)
# margin = 1
# loss = tf.maximum(0., margin + hardest_positive - hardest_negative)
loss = tf.log1p(tf.exp(hardest_positive - hardest_negative))
return loss
loss = tf.map_fn(get_loss, tf.range(0, size), dtype=tf.float32)
self.loss = tf.reduce_sum(loss)
self.add_scalar_summary(self.loss, "loss")
self.n_features = n_features
## Print - debug example code
# def test(a):
# print(a)
# return numpy.array([5], dtype="int32")
#
# t, = tf.py_func(test, [WHATEVER YOU WANT TO PRINT], [tf.int32])
# with tf.control_dependencies([t]):
# loss = tf.identity(loss)