def __init__(self, name, inputs, n_features, tower_setup, filter_size=(3, 3), old_order=False,
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=Layer.BATCH_NORM_DECAY_DEFAULT, l2=Layer.L2_DEFAULT,
padding="SAME", W_initializer=None, b_initializer=None, freeze_batchnorm_override=None):
super(Conv, self).__init__()
# mind the order of dropout, conv, activation and batchnorm!
# batchnorm -> activation -> dropout -> conv -> pool
# if old_order: dropout -> conv -> batchnorm -> activation -> pool (used for example in tensorpack)
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):
if isinstance(W_initializer, list):
W_initializer = tf.constant_initializer(W_initializer)
W = self.create_weight_variable("W", filter_size + [n_features_inp, n_features], l2, tower_setup,
initializer=W_initializer)
b = None
if bias:
if b_initializer is not None:
b_initializer = tf.constant_initializer(b_initializer)
b = self.create_bias_variable("b", [n_features], tower_setup, initializer=b_initializer)
else:
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, padding=padding)
else:
curr = conv2d_dilated(curr, W, dilation, padding=padding)
if bias:
curr += b
if batch_norm:
curr = self.create_and_apply_batch_norm(curr, n_features, batch_norm_decay, tower_setup, freeze_batchnorm_override=freeze_batchnorm_override)
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, freeze_batchnorm_override=freeze_batchnorm_override)
curr = get_activation(activation)(curr)
curr = apply_dropout(curr, dropout)
if dilation is None:
curr = conv2d(curr, W, strides, padding=padding)
else:
curr = conv2d_dilated(curr, W, dilation, padding=padding)
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,
dataset,
n_features,
tower_setup,
filter_size=(1, 1),
input_activation=None,
dilation=None,
l2=Layer.L2_DEFAULT,
dropout=0.0):
super().__init__()
if n_features == -1:
n_features = dataset.num_classes()
filter_size = list(filter_size)
inp, n_features_inp = prepare_input(inputs)
if input_activation is not None:
inp = get_activation(input_activation)(inp)
inp = apply_dropout(inp, dropout)
with tf.variable_scope(name):
W = self.create_weight_variable(
"W", filter_size + [n_features_inp, n_features], l2,
tower_setup)
b = self.create_bias_variable("b", [n_features], tower_setup)
if dilation is None:
output = conv2d(inp, W) + b
else:
output = conv2d_dilated(inp, W, dilation) + b
self.outputs = [output]
def __init__(self,
name,
inputs,
n_features,
tower_setup,
activation="relu",
dropout=0.0,
batch_norm=False,
batch_norm_decay=Layer.BATCH_NORM_DECAY_DEFAULT,
l2=Layer.L2_DEFAULT):
super(FullyConnected, self).__init__()
inp, n_features_inp = prepare_input(inputs)
with tf.variable_scope(name):
inp = apply_dropout(inp, dropout)
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]
self.n_features = n_features
def __init__(self, name, inputs, tower_setup, n_convs=2, n_features=None, dilations=None, strides=None,
filter_size=None, activation="relu", dropout=0.0, batch_norm_decay=Layer.BATCH_NORM_DECAY_DEFAULT,
l2=Layer.L2_DEFAULT):
super().__init__()
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 tower_setup.is_training:
curr = apply_dropout(curr, dropout)
if strides is None:
strides_res = [1, 1]
else:
strides_res = np.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, filter_size=(3, 3, 3), old_order=False,
strides=(1, 1, 1), pool_size=(1, 1, 1), pool_strides=None, activation="relu", dropout=0.0,
batch_norm=False, bias=False, batch_norm_decay=Layer.BATCH_NORM_DECAY_DEFAULT, l2=Layer.L2_DEFAULT,
padding="SAME", init_type="random"):
super(Conv3DOverBatch, self).__init__()
# mind the order of dropout, conv, activation and batchnorm!
# batchnorm -> activation -> dropout -> conv -> pool
# if old_order: dropout -> conv -> batchnorm -> activation -> pool (used for example in tensorpack)
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
# Introduce a fake "batch dimension", previous dimension becomes time
curr = tf.expand_dims(curr, axis=0)
with tf.variable_scope(name):
if init_type == "identity":
assert n_features == n_features_inp
identity_filter = np.zeros(filter_size, dtype=np.float32)
identity_filter[int(filter_size[0]/2.0), int(filter_size[1]/2.0), int(filter_size[2]/2.0)] = 1.0
init_weights = np.zeros(list(filter_size) + [n_features_inp, n_features], dtype=np.float32)
for i in range(n_features):
init_weights[:, :, :, i, i] = identity_filter
W = self.create_weight_variable("W", filter_size + [n_features_inp, n_features], l2, tower_setup, initializer=init_weights)
elif init_type == "avgtime":
assert n_features == n_features_inp
avgtime_filter = np.zeros(filter_size, dtype=np.float32)
avgtime_filter[:, int(filter_size[1]/2.0), int(filter_size[2]/2.0)] = 1.0/filter_size[0]
init_weights = np.zeros(list(filter_size) + [n_features_inp, n_features], dtype=np.float32)
for i in range(n_features):
init_weights[:, :, :, i, i] = avgtime_filter # Maybe we'ld like sth else than per-channel averaging?
W = self.create_weight_variable("W", filter_size + [n_features_inp, n_features], l2, tower_setup, initializer=init_weights)
else: # random
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)
curr = conv3d(curr, W, strides, padding=padding)
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)
curr = conv3d(curr, W, strides, padding=padding)
if bias:
curr += b
if pool_size != [1, 1, 1]:
curr = max_pool3d(curr, pool_size, pool_strides)
# Remove the fake dimension again
curr = tf.squeeze(curr, axis=0)
self.outputs = [curr]