def build_model(self):
"""Build the model to be saved as self._model."""
assert self.labeled
# build inputs
self._inputs = self.build_inputs()
moving_image = self._inputs["moving_image"]
fixed_image = self._inputs["fixed_image"]
moving_label = self._inputs["moving_label"]
# build ddf
backbone_inputs = self.concat_images(moving_image, fixed_image,
moving_label)
backbone = REGISTRY.build_backbone(
config=self.config["backbone"],
default_args=dict(
image_size=self.fixed_image_size,
out_channels=1,
out_kernel_initializer="glorot_uniform",
out_activation="sigmoid",
),
)
# (batch, f_dim1, f_dim2, f_dim3)
pred_fixed_label = backbone(inputs=backbone_inputs)
pred_fixed_label = tf.squeeze(pred_fixed_label, axis=4)
self._outputs = dict(pred_fixed_label=pred_fixed_label)
return tf.keras.Model(inputs=self._inputs, outputs=self._outputs)
def build_model(self):
"""Build the model to be saved as self._model."""
# build inputs
self._inputs = self.build_inputs()
moving_image = self._inputs[
"moving_image"] # (batch, m_dim1, m_dim2, m_dim3)
fixed_image = self._inputs[
"fixed_image"] # (batch, f_dim1, f_dim2, f_dim3)
# build ddf
control_points = self.config["backbone"].pop("control_points", False)
backbone_inputs = self.concat_images(moving_image, fixed_image)
backbone = REGISTRY.build_backbone(
config=self.config["backbone"],
default_args=dict(
image_size=self.fixed_image_size,
out_channels=3,
out_kernel_initializer="zeros",
out_activation=None,
),
)
if isinstance(backbone, GlobalNet):
# (f_dim1, f_dim2, f_dim3, 3), (4, 3)
ddf, theta = backbone(inputs=backbone_inputs)
self._outputs = dict(ddf=ddf, theta=theta)
else:
# (f_dim1, f_dim2, f_dim3, 3)
ddf = backbone(inputs=backbone_inputs)
ddf = (self._resize_interpolate(ddf, control_points)
if control_points else ddf)
self._outputs = dict(ddf=ddf)
# build outputs
warping = layer.Warping(fixed_image_size=self.fixed_image_size)
# (f_dim1, f_dim2, f_dim3)
pred_fixed_image = warping(inputs=[ddf, moving_image])
self._outputs["pred_fixed_image"] = pred_fixed_image
if not self.labeled:
return tf.keras.Model(inputs=self._inputs, outputs=self._outputs)
# (f_dim1, f_dim2, f_dim3)
moving_label = self._inputs["moving_label"]
pred_fixed_label = warping(inputs=[ddf, moving_label])
self._outputs["pred_fixed_label"] = pred_fixed_label
return tf.keras.Model(inputs=self._inputs, outputs=self._outputs)
def build_model(self):
"""Build the model to be saved as self._model."""
# build inputs
self._inputs = self.build_inputs()
moving_image = self._inputs["moving_image"]
fixed_image = self._inputs["fixed_image"]
control_points = self.config["backbone"].pop("control_points", False)
# build ddf
backbone_inputs = self.concat_images(moving_image, fixed_image)
backbone = REGISTRY.build_backbone(
config=self.config["backbone"],
default_args=dict(
image_size=self.fixed_image_size,
out_channels=3,
out_kernel_initializer="zeros",
out_activation=None,
),
)
dvf = backbone(inputs=backbone_inputs)
dvf = self._resize_interpolate(
dvf, control_points) if control_points else dvf
ddf = layer.IntDVF(fixed_image_size=self.fixed_image_size)(dvf)
# build outputs
self._warping = layer.Warping(fixed_image_size=self.fixed_image_size)
# (f_dim1, f_dim2, f_dim3, 3)
pred_fixed_image = self._warping(inputs=[ddf, moving_image])
self._outputs = dict(dvf=dvf,
ddf=ddf,
pred_fixed_image=pred_fixed_image)
if not self.labeled:
return tf.keras.Model(inputs=self._inputs, outputs=self._outputs)
# (f_dim1, f_dim2, f_dim3, 3)
moving_label = self._inputs["moving_label"]
pred_fixed_label = self._warping(inputs=[ddf, moving_label])
self._outputs["pred_fixed_label"] = pred_fixed_label
return tf.keras.Model(inputs=self._inputs, outputs=self._outputs)