def call(self, inputs):
depthwise_conv_on_filters = []
sliced_inputs = [sliced for sliced in tf.split(inputs, self.input_dim, self.channel_axis)]
sliced_kernels = [sliced for sliced in tf.split(self.depthwise_kernel, self.input_dim, 3)]
# See https://www.tensorflow.org/versions/r0.12/api_docs/python/array_ops/slicing_and_joining
for i in range(self.input_dim): #在python3中xrange合并为range
depthwise_conv_on_filters.append(K.conv3d(sliced_inputs[i],
sliced_kernels[i],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate))
depthwise_conv = K.concatenate(depthwise_conv_on_filters)
pointwise_conv = K.conv3d(depthwise_conv, self.pointwise_kernel,
strides=(1, 1, 1), padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs = pointwise_conv
if self.bias:
outputs = K.bias_add(
outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, inputs):
if self.rank == 1:
if self.Masked == False:
outputs = K.conv1d(inputs,
self.kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
else:
outputs = K.conv1d(inputs,
self.kernel * self.kernel_mask,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
if self.Masked == False:
outputs = K.conv2d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
else:
outputs = K.conv2d(inputs,
self.kernel * self.kernel_mask,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
if self.Masked == False:
outputs = K.conv3d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
else:
outputs = K.conv3d(inputs,
self.kernel * self.kernel_mask,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, input_tensor, training=None):
input_transposed = tf.transpose(input_tensor, [4, 0, 1, 2, 3, 5])
input_shape = K.shape(input_transposed)
input_tensor_reshaped = K.reshape(input_transposed, [
input_shape[1] * input_shape[0], self.input_height, self.input_width, self.input_depth, self.input_num_atoms])
input_tensor_reshaped.set_shape((None, self.input_height, self.input_width, self.input_depth, self.input_num_atoms))
if self.upsamp_type == 'resize':
# added 1 more self.scaling
upsamp = K.resize_images(input_tensor_reshaped, self.scaling, self.scaling, self.scaling, 'channels_last')
outputs = K.conv3d(upsamp, kernel=self.W, strides=(1, 1, 1), padding=self.padding, data_format='channels_last')
elif self.upsamp_type == 'subpix':
conv = K.conv3d(input_tensor_reshaped, kernel=self.W, strides=(1, 1, 1), padding='same',
data_format='channels_last')
outputs = tf.depth_to_space(conv, self.scaling)
else:
batch_size = input_shape[1] * input_shape[0]
# Infer the dynamic output shape:
out_height = deconv_length(self.input_height, self.scaling, self.kernel_size, self.padding)
out_width = deconv_length(self.input_width, self.scaling, self.kernel_size, self.padding)
out_depth = deconv_length(self.input_depth, self.scaling, self.kernel_size, self.padding)
output_shape = (batch_size, out_height, out_width, out_depth, self.num_capsule * self.num_atoms)
outputs = K.conv3d_transpose(input_tensor_reshaped, self.W, output_shape, (self.scaling, self.scaling, self.scaling),
padding=self.padding, data_format='channels_last')
votes_shape = K.shape(outputs)
_, conv_height, conv_width, conv_depth, _ = outputs.get_shape()
votes = K.reshape(outputs, [input_shape[2], input_shape[1], input_shape[0], votes_shape[1], votes_shape[2],
self.num_capsule, self.num_atoms])
votes.set_shape((None, self.input_num_capsule, conv_height.value, conv_width.value, conv_depth.value,
self.num_capsule, self.num_atoms))
logit_shape = K.stack([
input_shape[1], input_shape[0], votes_shape[1], votes_shape[2], votes_shape[3], self.num_capsule])
biases_replicated = K.tile(self.b, [votes_shape[1], votes_shape[2], votes_shape[3], 1, 1])
activations = update_routing(
votes=votes,
biases=biases_replicated,
logit_shape=logit_shape,
num_dims=7,
input_dim=self.input_num_capsule,
output_dim=self.num_capsule,
num_routing=self.routings)
return activations
def call(self, input_tensor, training=None):
input_transposed = tf.transpose(input_tensor, [0, 3, 4, 1, 2])
input_shape = K.shape(input_transposed)
input_tensor_reshaped = K.reshape(input_tensor, [input_shape[0], 1, self.input_num_capsule * self.input_num_atoms, self.input_height, self.input_width])
input_tensor_reshaped.set_shape((None, 1, self.input_num_capsule * self.input_num_atoms, self.input_height, self.input_width))
# conv = Conv3D(input_tensor_reshaped, self.W, (self.strides, self.strides),
# padding=self.padding, data_format='channels_first')
conv = K.conv3d(input_tensor_reshaped, self.W, strides=(self.input_num_atoms, self.strides, self.strides), padding=self.padding, data_format='channels_first')
votes_shape = K.shape(conv)
_, _, _, conv_height, conv_width = conv.get_shape()
conv = tf.transpose(conv, [0, 2, 1, 3, 4])
votes = K.reshape(conv, [input_shape[0], self.input_num_capsule, self.num_capsule, self.num_atoms, votes_shape[3], votes_shape[4]])
votes.set_shape((None, self.input_num_capsule, self.num_capsule, self.num_atoms, conv_height.value, conv_width.value))
logit_shape = K.stack([input_shape[0], self.input_num_capsule, self.num_capsule, votes_shape[3], votes_shape[4]])
biases_replicated = K.tile(self.b, [1, 1, conv_height.value, conv_width.value])
activations = update_routing(
votes=votes,
biases=biases_replicated,
logit_shape=logit_shape,
num_dims=6,
input_dim=self.input_num_capsule,
output_dim=self.num_capsule,
num_routing=self.routings)
a2 = tf.transpose(activations, [0, 3, 4, 1, 2])
return a2
def call(self, inputs):
if self.rank == 1:
outputs = K.conv1d(
inputs,
self.kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
outputs = K.conv2d(
inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
outputs = K.conv3d(
inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(
outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def cal_conv(self, inputs, kernel):
if self.rank == 1:
outputs = K.conv1d(
inputs,
kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
outputs = K.conv2d(
inputs,
kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
outputs = K.conv3d(
inputs,
kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
return outputs
def call(self, inputs):
def _l2normalize(v, eps=1e-12):
return v / (K.sum(v**2)**0.5 + eps)
def power_iteration(W, u):
_u = u
_v = _l2normalize(K.dot(_u, K.transpose(W)))
_u = _l2normalize(K.dot(_v, W))
return _u, _v
if self.spectral_normalization:
W_shape = self.kernel.shape.as_list()
#Flatten the Tensor
W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]])
_u, _v = power_iteration(W_reshaped, self.u)
#Calculate Sigma
sigma = K.dot(_v, W_reshaped)
sigma = K.dot(sigma, K.transpose(_u))
#normalize it
W_bar = W_reshaped / sigma
#reshape weight tensor
if training in {0, False}:
W_bar = K.reshape(W_bar, W_shape)
else:
with tf.control_dependencies([self.u.assign(_u)]):
W_bar = K.reshape(W_bar, W_shape)
#update weitht
self.kernel = W_bar
if self.rank == 1:
outputs = K.conv1d(inputs,
self.kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
outputs = K.conv2d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
outputs = K.conv3d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, x, mask=None):
# compute the candidate hidden state
input_shape = self.input_spec[0].shape
conv_out = K.conv3d(x,
self.W,
strides=self.subsample,
border_mode=self.border_mode,
dim_ordering=self.dim_ordering,
volume_shape=input_shape,
filter_shape=self.W_shape)
if self.dim_ordering == 'th':
transform = conv_out + K.reshape(self.b,
(1, self.nb_filter, 1, 1, 1))
elif self.dim_ordering == 'tf':
transform = conv_out + K.reshape(self.b,
(1, 1, 1, 1, self.nb_filter))
else:
raise Exception('Invalid dim_ordering: ' + self.dim_ordering)
transform = self.activation(transform)
transform_gate = K.conv3d(x,
self.W_gate,
strides=self.subsample,
border_mode=self.border_mode,
dim_ordering=self.dim_ordering,
volume_shape=input_shape,
filter_shape=self.W_shape)
if self.bias:
if self.dim_ordering == 'th':
transform_gate += K.reshape(self.b_gate,
(1, self.nb_filter, 1, 1, 1))
elif self.dim_ordering == 'tf':
transform_gate += K.reshape(self.b_gate,
(1, 1, 1, 1, self.nb_filter))
else:
raise Exception('Invalid dim_ordering: ' + self.dim_ordering)
transform_gate = K.sigmoid(transform_gate)
#transform_gate = self.activation(transform_gate)
carry_gate = 1.0 - transform_gate
return transform * transform_gate + x * carry_gate
def sf_conv3d_base(x):
sf2, expand_tensor = x
return K.conv3d(sf2,
expand_tensor,
strides=[1, 1, 1, 1, 1],
padding="valid",
data_format="channels_first")
def call(self, inputs, mask=None):
# compute channels_firste candidate hidden state
# Arguments
'''
x: Tensor or variable.
kernel: kernel tensor.
strides: strides tuple.
padding: string, `"same"` or `"valid"`.
data_format: string, `"channels_last"` or `"channels_first"`.
Whether to use Theano or TensorFlow/CNTK data format
for inputs/kernels/outputs.
dilation_rate: tuple of 3 integers.
'''
transform = K.conv3d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
transform = K.bias_add(transform,
self.bias,
data_format=self.data_format)
if self.activation is not None:
transform = self.activation(transform)
transform_gate = K.conv3d(inputs,
self.kernel_gate,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
transform = K.bias_add(transform,
self.bias_gate,
data_format=self.data_format)
transform_gate = K.sigmoid(transform_gate)
carry_gate = 1.0 - transform_gate
return transform * transform_gate + inputs * carry_gate
def wavelet_transform(img, filters=None, levels=None):
if levels is None:
vimg = tf.pad(img, [(0, 0),
(0, 2**int(np.ceil(np.log2(K.int_shape(img)[1]))) -
K.int_shape(img)[1]),
(0, 2**int(np.ceil(np.log2(K.int_shape(img)[2]))) -
K.int_shape(img)[2]),
(0, 2**int(np.ceil(np.log2(K.int_shape(img)[3]))) -
K.int_shape(img)[3]), (0, 0)])
else:
vimg = img
if filters is None:
w = pywt.Wavelet('db4')
dec_hi = np.array(w.dec_hi[::-1])
dec_lo = np.array(w.dec_lo[::-1])
filters = np.stack([
dec_lo[None, None, :] * dec_lo[None, :, None] *
dec_lo[:, None, None], dec_lo[None, None, :] *
dec_lo[None, :, None] * dec_hi[:, None, None],
dec_lo[None, None, :] * dec_hi[None, :, None] *
dec_lo[:, None, None], dec_lo[None, None, :] *
dec_hi[None, :, None] * dec_hi[:, None, None],
dec_hi[None, None, :] * dec_lo[None, :, None] *
dec_lo[:, None, None], dec_hi[None, None, :] *
dec_lo[None, :, None] * dec_hi[:, None, None],
dec_hi[None, None, :] * dec_hi[None, :, None] *
dec_lo[:, None, None], dec_hi[None, None, :] *
dec_hi[None, :, None] * dec_hi[:, None, None]
]).transpose((1, 2, 3, 0))[:, :, :, None, :]
filters = K.constant(filters)
if levels is None:
print(K.int_shape(vimg)[1:4])
levels = pywt.dwtn_max_level(K.int_shape(vimg)[1:4], 'db4')
print(levels)
t = vimg.shape[1]
h = vimg.shape[2]
w = vimg.shape[3]
res = K.conv3d(vimg, filters, strides=(2, 2, 2), padding='same')
if levels > 1:
res = K.concatenate([
wavelet_transform(res[:, :, :, :, :1],
filters,
levels=(levels - 1)), res[:, :, :, :, 1:]
],
axis=-1)
'''
res = K.permute_dimensions(res, (0, 4, 1, 2, 3))
res = K.reshape(res, (-1, 2, t // 2, h // 2, w // 2))
res = K.permute_dimensions(res, (0, 2, 1, 3, 4))
res = K.reshape(res, (-1, 1, t, h, w))
res = K.permute_dimensions(res, (0, 2, 3, 4, 1))
'''
res = K.reshape(res, (-1, t, h, w, 1))
#print('wt', levels, K.int_shape(img), K.int_shape(vimg), K.int_shape(filters), K.int_shape(res))
return res
def _alphabeta_dtd(layer, R, beta, parameter2):
print('_convolutional3d_alphabeta_dtd')
alpha = 1 + beta
X = layer.input + 1e-12
if not alpha == 0:
Wp = K.maximum(layer.kernel, 1e-12)
Zp = K.conv3d(X,
Wp,
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
Salpha = alpha * (R / Zp)
Calpha = K.conv3d_transpose(Salpha,
Wp,
K.shape(layer.input),
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
else:
Calpha = 0
if not beta == 0:
Wn = K.minimum(layer.kernel, -1e-12)
Zn = K.conv3d(X,
Wn,
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
Sbeta = -beta * (R / Zn)
Cbeta = K.conv3d_transpose(Sbeta,
Wn,
K.shape(layer.input),
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
else:
Cbeta = 0
return X * (Calpha + Cbeta)
def call(self, inputs):
if self.rank == 1:
expert_outputs = K.conv1d(inputs,
self.expert_kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
expert_outputs = K.conv2d(inputs,
self.expert_kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
expert_outputs = K.conv3d(inputs,
self.expert_kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
expert_outputs = K.reshape(expert_outputs,
(-1, ) + self.o_shape[1:-1] +
(self.n_filters, self.n_experts_per_filter))
if self.use_expert_bias:
expert_outputs = K.bias_add(expert_outputs,
self.expert_bias,
data_format=self.data_format)
if self.expert_activation is not None:
expert_outputs = self.expert_activation(expert_outputs)
gating_outputs = tf.tensordot(
inputs, self.gating_kernel,
axes=self.rank + 1) # samples x n_filters x n_experts_per_filter
if self.use_gating_bias:
gating_outputs = K.bias_add(gating_outputs,
self.gating_bias,
data_format=self.data_format)
if self.gating_activation is not None:
gating_outputs = self.gating_activation(gating_outputs)
gating_outputs = K.reshape(gating_outputs,
self.new_gating_outputs_shape)
outputs = K.sum(expert_outputs * gating_outputs,
axis=-1,
keepdims=False)
return outputs
def get_grad_tensor_3d(img_tensor,apply_gauss=True):
grad_x = K.conv3d(img_tensor, SOBEL_X_3D, padding='same')
grad_y = K.conv3d(img_tensor, SOBEL_Y_3D, padding='same')
grad_z= K.conv3d(img_tensor, SOBEL_Z_3D, padding='same')
grad_tensor = K.sqrt(grad_x * grad_x + grad_y * grad_y + grad_z*grad_z)
grad_tensor = K.greater(grad_tensor, 100.0 * K.epsilon())
grad_tensor = K.cast(grad_tensor, K.floatx())
grad_tensor = K.clip(grad_tensor, K.epsilon(), 1.0)
grad_map = K.sum(grad_tensor, axis=-1, keepdims=True)
for i in range(n_classes):
if i ==0:
grad_tensor=grad_map[:]
else:
grad_tensor = K.concatenate([grad_tensor,grad_map], axis=-1)
# del grad_map
# grad_tensor = K.concatenate([grad_tensor, grad_tensor], axis=CHANNEL_AXIS)
grad_tensor = K.greater(grad_tensor, 100.0 * K.epsilon())
grad_tensor = K.cast(grad_tensor, K.floatx())
if apply_gauss:
grad_tensor = K.conv3d(grad_tensor, GAUSS_KERNEL_3D, padding='same')
return grad_tensor
def var_rad_layer(x):
in_img, in_vox = x
tf_weights = tf.expand_dims(
tf.stack([
gkern_tf(d=3,
kernlen=k_dim,
nsigs=[
c_wid / vox_size[0, 0], c_wid / vox_size[0, 1],
c_wid / vox_size[0, 2]
])
for c_wid in np.linspace(min_width, max_width, gk_count)
], -1), -2)
return K.conv3d(in_img, kernel=tf_weights, padding='same')
def do_3d_convolution(feature_matrix,
kernel_matrix,
pad_edges=False,
stride_length_px=1):
"""Convolves 3-D feature maps with 3-D kernel.
m = number of rows in kernel
n = number of columns in kernel
h = number of height in kernel
c = number of output feature maps (channels)
:param feature_matrix: Input feature maps (numpy array). Dimensions must be
M x N x H x C or 1 x M x N x H x C.
:param kernel_matrix: Kernel as numpy array. Dimensions must be
m x n x h x C x c.
:param pad_edges: See doc for `do_2d_convolution`.
:param stride_length_px: See doc for `do_2d_convolution`.
:return: feature_matrix: Output feature maps (numpy array). Dimensions will
be 1 x M x N x H x c or
1 x (M - m + 1) x (N - n + 1) x (H - h + 1) x c, depending on
whether or not edges are padded.
"""
error_checking.assert_is_numpy_array_without_nan(feature_matrix)
error_checking.assert_is_numpy_array_without_nan(kernel_matrix)
error_checking.assert_is_numpy_array(kernel_matrix, num_dimensions=5)
error_checking.assert_is_boolean(pad_edges)
error_checking.assert_is_integer(stride_length_px)
error_checking.assert_is_geq(stride_length_px, 1)
if len(feature_matrix.shape) == 4:
feature_matrix = numpy.expand_dims(feature_matrix, axis=0)
error_checking.assert_is_numpy_array(feature_matrix, num_dimensions=5)
if pad_edges:
padding_string = 'same'
else:
padding_string = 'valid'
feature_tensor = K.conv3d(x=K.variable(feature_matrix),
kernel=K.variable(kernel_matrix),
strides=(stride_length_px, stride_length_px,
stride_length_px),
padding=padding_string,
data_format='channels_last')
return feature_tensor.eval(session=K.get_session())
def _compute_mask_output(self, mask_tensor):
if self.layer.rank == 1:
mask_output = K.conv1d(mask_tensor, self.mask_kernel,
self.layer.strides[0], self.layer.padding,
self.layer.data_format,
self.layer.dilation_rate[0])
if self.layer.rank == 2:
mask_output = K.conv2d(mask_tensor, self.mask_kernel,
self.layer.strides, self.layer.padding,
self.layer.data_format,
self.layer.dilation_rate)
if self.layer.rank == 3:
mask_output = K.conv3d(mask_tensor, self.mask_kernel,
self.layer.strides, self.layer.padding,
self.layer.data_format,
self.layer.dilation_rate)
return mask_output
def _z_dtd(layer, R, parameter1, parameter2):
print('_convolutional3d_z_dtd')
X = layer.input + 1e-12
Z = K.conv3d(X,
layer.kernel,
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
S = R / Z
C = K.conv3d_transpose(S,
layer.kernel,
K.shape(layer.input),
strides=layer.strides,
padding=layer.padding,
data_format=layer.data_format)
return X * C
def call(self, x, **kwargs):
x = K.spatial_3d_padding(x, padding=self.padding)
# we imitate depthwise_conv3d actually
channels = x.shape[-1]
x = K.concatenate(
[
K.conv3d(
x=x[:, :, :, :, i:i + 1],
kernel=self.blur_kernel[..., i:i + 1, :],
strides=self.pool_size,
padding='valid',
) for i in range(0, channels)
],
axis=-1,
)
return x
def call(self, inputs):
self.kernel = K.reverse(K.permute_dimensions(
self.tied_to.get_weights()[0], (0, 1, 2, 4, 3)),
axes=(0, 1, 2))
outputs = K.conv3d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, inputs, training=None):
def _l2normalize(v, eps=1e-12):
return v / (K.sum(v ** 2) ** 0.5 + eps)
def power_iteration(W, u):
#Accroding the paper, we only need to do power iteration one time.
_u = u
_v = _l2normalize(K.dot(_u, K.transpose(W)))
_u = _l2normalize(K.dot(_v, W))
return _u, _v
#Spectral Normalization
W_shape = self.kernel.shape.as_list()
#Flatten the Tensor
W_reshaped = K.reshape(self.kernel, [-1, W_shape[-1]])
_u, _v = power_iteration(W_reshaped, self.u)
#Calculate Sigma
sigma=K.dot(_v, W_reshaped)
sigma=K.dot(sigma, K.transpose(_u))
#normalize it
W_bar = W_reshaped / sigma
#reshape weight tensor
if training:
W_bar = K.reshape(W_bar, W_shape)
else:
with tf.control_dependencies([self.u.assign(_u)]):
W_bar = K.reshape(W_bar, W_shape)
outputs = K.conv3d(
inputs,
W_bar,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(
outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def depthwiseConvGPU(cube, k):
'''
Convolves a kernel depthwise (each z-plane independently)
by using K.conv3d with an (x,y,1) 3d kernel
Parameters
----------
cube : ndarray.
3d ndarray, (x, y, z).
k : ndarray.
2d ndarray (x, y) to be convolved depthwise.
Returns
-------
cubeC : ndarray.
3d ndarray (x, y, z) convolved with kernel depthwise.
'''
kT = K.variable(k.reshape(1, 1, k.shape[0], k.shape[1], 1))
xT = K.variable(
cube.reshape(1, 1, cube.shape[0], cube.shape[1], cube.shape[2]))
out = K.conv3d(xT, kernel=kT, border_mode='same')
cubeC = out.eval()
return np.squeeze(cubeC)
def call(self, input_tensor, training=None):
input_transposed = tf.transpose(input_tensor, [4, 0, 1, 2, 3, 5])
input_shape = K.shape(input_transposed)
input_tensor_reshaped = K.reshape(input_transposed, [
input_shape[0] * input_shape[1], self.input_height,
self.input_width, self.input_depth, self.input_num_atoms])
input_tensor_reshaped.set_shape((None, self.input_height, self.input_width, self.input_depth, self.input_num_atoms))
conv = K.conv3d(input_tensor_reshaped, self.W, (self.strides, self.strides, self.strides),
padding=self.padding, data_format='channels_last', dilation_rate=(1, 1, 1))
votes_shape = K.shape(conv)
_, conv_height, conv_width, conv_depth, _ = conv.get_shape()
votes = K.reshape(conv, [input_shape[1], input_shape[0], votes_shape[1], votes_shape[2], votes_shape[3],
self.num_capsule, self.num_atoms])
votes.set_shape((None, self.input_num_capsule, conv_height.value, conv_width.value, conv_depth.value,
self.num_capsule, self.num_atoms))
logit_shape = K.stack([
input_shape[1], input_shape[0], votes_shape[1], votes_shape[2], votes_shape[3], self.num_capsule])
biases_replicated = K.tile(self.b, [conv_height.value, conv_width.value, conv_depth.value, 1, 1])
activations = update_routing(
votes=votes,
biases=biases_replicated,
logit_shape=logit_shape,
num_dims=7,
input_dim=self.input_num_capsule,
output_dim=self.num_capsule,
num_routing=self.routings)
return activations
def call(self, inputs, **kwargs):
t1, t2, t3, t4, t5 = inputs
# concatenation = keras.layers.Concatenate()
layer_list = []
for i in range(t1.shape[3]):
layer_list.append(t1[:, :, :, i:i+1])
layer_list.append(t2[:, :, :, i:i+1])
layer_list.append(t3[:, :, :, i:i+1])
layer_list.append(t4[:, :, :, i:i+1])
layer_list.append(t5[:, :, :, i:i+1])
tensor = keras.backend.stack(layer_list, axis=-1)
# print(self.kernel.shape)
tensor = K.conv3d(tensor, self.kernel, padding='same', strides=(1, 1, 5))
# print(tensor.shape)
# tensor = tf.layers.conv3d(tensor, t1.shape[3], (1, 1, 5), activation='relu', strides=(1, 1, 5), padding='same')
# conv3D_layer = keras.layers.Conv3D(int(t1.shape[3]), (1, 1, 5), strides=(1, 1, 5), activation='relu', padding='same')
# conv3D_layer.set_weights = self.add_weight(name='kernel', shape=(1, 1, 5, int(t1.shape[3])), initializer='uniform', trainable=True)
# tensor = conv3D_layer(tensor)
tensor = keras.backend.squeeze(tensor, axis=-2)
return tensor
def call(self, inputs):
if self.rank == 1:
_conv = lambda x, k: K.conv1d( # noqa: E731
x,
k,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0],
)
if self.rank == 2:
_conv = lambda x, k: K.conv2d( # noqa: E731
x,
k,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate,
)
if self.rank == 3:
_conv = lambda x, k: K.conv3d( # noqa: E731
x,
k,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate,
)
def _both_call(inputs, convfunc):
x, lo, hi = inputs
output = _single_call(x, convfunc)
out_lo, out_hi = _interval_call([lo, hi], convfunc)
output = [output, out_lo, out_hi]
return output
def _interval_call(inputs, convfunc):
lo, hi = inputs
lo_neg = K.minimum(lo, 0.0)
lo_pos = K.maximum(lo, 0.0)
hi_neg = K.minimum(hi, 0.0)
hi_pos = K.maximum(hi, 0.0)
min_kernel_pos = K.maximum(self.kernel - self.min_kernel, 0.0)
min_kernel_neg = K.minimum(self.kernel - self.min_kernel, 0.0)
max_kernel_pos = K.maximum(self.kernel + self.max_kernel, 0.0)
max_kernel_neg = K.minimum(self.kernel + self.max_kernel, 0.0)
lo_out = (convfunc(lo_pos, min_kernel_pos) +
convfunc(hi_pos, min_kernel_neg) +
convfunc(lo_neg, max_kernel_pos) +
convfunc(hi_neg, max_kernel_neg))
hi_out = (convfunc(lo_pos, max_kernel_neg) +
convfunc(hi_pos, max_kernel_pos) +
convfunc(lo_neg, min_kernel_neg) +
convfunc(hi_neg, min_kernel_pos))
if self.use_bias:
lo_out = K.bias_add(
lo_out,
self.bias - self.min_bias,
data_format=self.data_format,
)
hi_out = K.bias_add(
hi_out,
self.bias + self.max_bias,
data_format=self.data_format,
)
if self.activation is not None:
lo_out = self.activation(lo_out)
hi_out = self.activation(hi_out)
output = [lo_out, hi_out]
return output
def _single_call(inputs, convfunc):
x = inputs
output = convfunc(x, self.kernel)
if self.use_bias:
output = K.bias_add(output,
self.bias,
data_format=self.data_format)
if self.activation is not None:
output = self.activation(output)
return output
if isinstance(inputs, list):
if len(inputs) == 2:
return _interval_call(inputs, _conv)
elif len(inputs) == 3:
return _both_call(inputs, _conv)
else:
return _single_call(inputs, _conv)
def call(self, inputs):
assert type(inputs) is list, 'must input a list containing two tensors'
assert len(inputs) == 2, 'must input 2 inputs tensors, one for real'\
+ ', and the other for image.'
if self.rank == 1:
outputs_real = K.conv1d(
inputs[0],
self.kernel[0],
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])\
- K.conv1d(
inputs[1],
self.kernel[1],
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
outputs_imag = K.conv1d(
inputs[0],
self.kernel[1],
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])\
+ K.conv1d(
inputs[1],
self.kernel[0],
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
outputs_real = K.conv2d(
inputs[0],
self.kernel[0],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)\
- K.conv2d(
inputs[1],
self.kernel[1],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs_imag = K.conv2d(
inputs[0],
self.kernel[1],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)\
+ K.conv2d(
inputs[1],
self.kernel[0],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
outputs_real = K.conv3d(
inputs[0],
self.kernel[0],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)\
- K.conv3d(
inputs[1],
self.kernel[1],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
outputs_imag = K.conv3d(
inputs[0],
self.kernel[1],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)\
+ K.conv3d(
inputs[1],
self.kernel[0],
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.output_merge:
outputs = K.concatenate((K.expand_dims(
outputs_real, axis=1), K.expand_dims(outputs_imag, axis=1)),
axis=1)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
else:
if self.use_bias:
outputs_real = K.bias_add(outputs_real,
self.bias[0],
data_format=self.data_format)
outputs_imag = K.bias_add(outputs_imag,
self.bias[1],
data_format=self.data_format)
if self.activation is not None:
return [
self.activation(outputs_real),
self.activation(outputs_imag)
]
return [outputs_real, outputs_imag]
def call(self, inputs, training=None):
if self.spectral_norm:
w_shape = K.shape(self.kernel)
w_reshaped = K.reshape(self.kernel, (-1, w_shape[-1]))
u_hat = self.u
v_hat = None
for i in range(1):
"""
power iteration
Usually iteration = 1 will be enough
"""
v_ = K.dot(u_hat, K.transpose(w_reshaped))
v_hat = l2_norm(v_)
u_ = K.dot(v_hat, w_reshaped)
u_hat = l2_norm(u_)
sigma = K.dot(v_hat, w_reshaped)
sigma = K.dot(sigma, K.transpose(u_hat))
# normalize it
w_bar = w_reshaped / sigma
# reshape weight tensor
if training in {0, False}:
w_bar = K.reshape(w_bar, w_shape)
else:
with tf.control_dependencies([self.u.assign(u_)]):
w_bar = K.reshape(w_bar, w_shape)
self.kernel = w_bar
if self.rank == 1:
outputs = K.conv1d(inputs,
self.kernel,
strides=self.strides[0],
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate[0])
if self.rank == 2:
outputs = K.conv2d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.rank == 3:
outputs = K.conv3d(inputs,
self.kernel,
strides=self.strides,
padding=self.padding,
data_format=self.data_format,
dilation_rate=self.dilation_rate)
if self.use_bias:
outputs = K.bias_add(outputs,
self.bias,
data_format=self.data_format)
if self.activation is not None:
return self.activation(outputs)
return outputs
def call(self, x):
# print('input shape:',x.shape) # None, 16,12,512
exp_x = K.expand_dims(x, axis=-1)
# print('expanded input shape:',exp_x.shape) # N, 16, 12, 512, 1
# filters = Attention Heads
c3d = K.conv3d(exp_x,
kernel=self.kernel_conv3d,
strides=(1, 1, self.i_shape[-1]),
padding='same',
data_format='channels_last')
conv3d = K.bias_add(c3d, self.bias_conv3d)
# print('c3d shape:', c3d.shape)
# conv3d = kl.Conv3D(padding='same',filters=self.multiheads, kernel_size=(3,3,self.i_shape[-1]), strides=(1,1,self.i_shape[-1]),kernel_initializer='he_normal',activation='relu')(exp_x)
# print('conv3d shape:', conv3d.shape)
conv3d = K.permute_dimensions(conv3d, pattern=(0, 4, 1, 2, 3))
# print('conv3d shape:', conv3d.shape)
# conv3d = K.flatten(conv3d)
conv3d = K.squeeze(conv3d, axis=-1)
conv3d = K.reshape(conv3d,
shape=(-1, self.multiheads,
self.i_shape[1] * self.i_shape[2]))
# print('conv3d shape:', conv3d.shape)
# conv3d = K.expand_dims(conv3d,axis=1) # N, 1, 16, 12
# print('conv3d shape:', conv3d.shape)
softmax_alpha = K.softmax(conv3d, axis=-1) # attention map # N, 16x12
# print('softmax_alpha shape:', softmax_alpha.shape)
softmax_alpha = K.reshape(softmax_alpha,
shape=(-1, self.multiheads, self.i_shape[1],
self.i_shape[2]))
# print('softmax_alpha shape:', softmax_alpha.shape)
if self.aggregate_channels == False:
exp_softmax_alpha = K.expand_dims(
softmax_alpha, axis=-1) # for elementwise multiplication
# print('exp_softmax_alpha shape:', exp_softmax_alpha.shape)
exp_softmax_alpha = K.permute_dimensions(exp_softmax_alpha,
pattern=(0, 2, 3, 1, 4))
# print('exp_softmax_alpha shape:', exp_softmax_alpha.shape)
x_exp = K.expand_dims(x, axis=-2)
# print('x_exp shape:', x_exp.shape)
u = multiply([exp_softmax_alpha, x_exp])
# print('u shape:', u.shape)
u = K.reshape(u,
shape=(-1, self.i_shape[1], self.i_shape[2],
u.shape[-1] * u.shape[-2]))
else:
exp_softmax_alpha = K.permute_dimensions(softmax_alpha,
pattern=(0, 2, 3, 1))
exp_softmax_alpha = K.sum(exp_softmax_alpha, axis=-1)
# print('exp_softmax_alpha shape:', exp_softmax_alpha.shape)
exp_softmax_alpha = K.expand_dims(exp_softmax_alpha, axis=-1)
# print('exp_softmax_alpha shape:', exp_softmax_alpha.shape)
u = multiply([exp_softmax_alpha, x])
# print('u shape:', u.shape)
if self.concat_input_with_scaled:
o = K.concatenate([u, x], axis=-1)
else:
o = u
# print('o shape:', o.shape)
# u = kl.Conv2D(activation='relu',filters=self.i_shape[-1],kernel_size=(1,1),padding='valid')(u)
# u = self.gamma2 * u + x
# print('u shape:', u.shape)
# u = self.gamma2 * u
# u = K.tanh(u)
# print('u shape:', u.shape)
# return [u, softmax_alpha]
# self.out_features_shape = tuple(u.shape.as_list())
# self.out_attention_maps_shape = tuple(softmax_alpha.shape.as_list())
# print(self.out_features_shape, self.out_attention_maps_shape)
return [o, softmax_alpha]