def f(input):
conv = Conv2D(
filter,
k,
strides=s,
padding=pad,
kernel_regularizer=l2(w_decay),
kernel_initializer=initial(seed),
)(input)
return _bn_relu()(conv)
def f(input):
activation = _bn_relu()(input)
conv = Conv2D(
filter,
k,
strides=s,
padding=pad,
kernel_regularizer=l2(w_decay),
kernel_initializer=initial(seed),
)(activation)
return conv
def f(input):
conv = _conv_bn_relu(filter, (3, 3), s=strides, w_decay=w_decay)(input)
residual = Conv2D(
filter,
(3, 3),
strides=1,
padding='same',
kernel_regularizer=l2(w_decay),
kernel_initializer=initial(seed),
)(conv)
residual = BatchNormalization(axis=3)(residual)
sum = _shortcut(input, residual, w_decay)
return Activation('relu')(sum)
def f(input):
if fst_layer_fst_stage:
conv = Conv2D(
filter,
(3, 3),
strides=1,
padding='same',
kernel_regularizer=l2(w_decay),
kernel_initializer=initial(seed),
)(input)
else:
conv = _bn_relu_conv(filter, 3, s=strides, w_decay=w_decay)(input)
residual = _bn_relu_conv(filter, 3, 1, w_decay=w_decay)(conv)
return _shortcut(input, residual, w_decay)
def _shortcut(input, residual, w_decay):
input_shape = K.int_shape(input)
res_shape = K.int_shape(residual)
size_equality = int(round(input_shape[1] / res_shape[1]))
channel_equality = input_shape[3] == res_shape[3]
x = input
if size_equality > 1 or not channel_equality:
x = Conv2D(
res_shape[3],
(1, 1),
strides=(size_equality, size_equality),
padding='same',
kernel_regularizer=l2(w_decay),
kernel_initializer=initial(seed),
)(input)
return Add()([x, residual])
def nonlocal_resnet(input_shape=(
32,
32,
3,
),
num_classes=10,
w_decay=1e-4,
depth=34,
stage=None):
if depth >= 50:
block_function = _bottleneck_block
else:
block_function = _normal_block
if stage is None:
stage = [3, 4, 6, 3]
# network begins
filter = filter_num
input = Input(input_shape)
conv1 = _conv_bn_relu(filter=filter, k=3, s=1, w_decay=w_decay)(input)
# repetition to build res stage
block = conv1
for i, r in enumerate(stage):
block = _resblock(block_function, filter, r, i, w_decay, i == 0)(block)
filter *= 2
if i == 3:
block = non_local(block)
# classifier
avg = GlobalAvgPool2D()(block)
output = Dense(
num_classes,
activation='softmax',
kernel_initializer=initial(seed),
)(avg)
model = Model(input, output)
model.summary()
return model
def non_local(input,
bottle_dim=None,
compression=2,
mode='embedded',
residual=True):
if mode not in ['gaussian', 'embedded', 'dot', 'concatenate']:
raise ValueError('mode is not right.')
shape = K.int_shape(input)
length, dim1, dim2, channel = shape
# define bottle_dim
if bottle_dim is None:
bottle_dim = ceil(shape[3]) // 2
else:
bottle_dim = int(bottle_dim)
if bottle_dim < 1:
raise ValueError('bottle_dim must > 1')
# compute f
if mode == 'gaussian':
xi = Reshape((-1, bottle_dim))(input)
xj = Reshape((-1, bottle_dim))(input)
f = Dot(axes=2)([xi, xj])
f = Activation('softmax')(f)
if mode == 'embedded':
theta = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
theta = Reshape((-1, bottle_dim))(theta)
phi = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
phi = Reshape((-1, bottle_dim))(phi)
if compression > 1:
phi = MaxPool1D(compression)(phi)
f = Dot(axes=2)([theta, phi])
f = Activation('softmax')(f)
if mode == 'dot':
theta = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
theta = Reshape((-1, bottle_dim))(theta)
phi = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
phi = Reshape((-1, bottle_dim))(phi)
f = Dot(axes=2)([theta, phi])
size = K.int_shape(f)
f = Lambda(lambda out: (1. / float(size[-1]) * out))(f)
if mode == 'concatenate':
theta = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
phi = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
new_x = Concatenate(axis=3)(theta, phi)
f = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(new_x)
f = Reshape((-1, bottle_dim))(f)
f = Activation('relu')(f)
# compute g
g = Conv2D(bottle_dim, (1, 1),
padding='same',
kernel_initializer=initial())(input)
g = Reshape((-1, bottle_dim))(g)
if compression > 1:
g = MaxPool1D(compression)(g)
# compute y
y = Dot(axes=[2, 1])([f, g])
y = Reshape((dim1, dim2, bottle_dim))(y)
y = Conv2D(channel, (1, 1), padding='same',
kernel_initializer=initial())(y)
# residual
if residual == True:
y = Add()([y, input])
return y
