Network

class deepreg.model.network.ConditionalModel(*args: Any, **kwargs: Any)

A registration model predicts fixed image label without DDF or DVF.

Init.

Parameters

• moving_image_size – (m_dim1, m_dim2, m_dim3)

• fixed_image_size – (f_dim1, f_dim2, f_dim3)

• index_size – number of indices for identify each sample

• labeled – if the data is labeled

• batch_size – total number of samples consumed per step, over all devices. When using multiple devices, TensorFlow automatically split the tensors. Therefore, input shapes should be defined over batch_size.

• config – config for method, backbone, and loss.

• name – name of the model

build_model()

Build the model to be saved as self._model.

postprocess(inputs: Dict[str, tensorflow.Tensor], outputs: Dict[str, tensorflow.Tensor]) → Tuple[tensorflow.Tensor, Dict]

Return a dict used for saving inputs and outputs.

Parameters

• inputs – dict of model inputs

• outputs – dict of model outputs

Returns

tuple, indices and a dict. In the dict, each value is (tensor, normalize, on_label), where - normalize = True if the tensor need to be normalized to [0, 1] - on_label = True if the tensor depends on label

class deepreg.model.network.DDFModel(*args: Any, **kwargs: Any)

A registration model predicts DDF.

When using global net as backbone, the model predicts an affine transformation parameters, and a DDF is calculated based on that.

Init.

Parameters

• moving_image_size – (m_dim1, m_dim2, m_dim3)

• fixed_image_size – (f_dim1, f_dim2, f_dim3)

• index_size – number of indices for identify each sample

• labeled – if the data is labeled

• batch_size – total number of samples consumed per step, over all devices. When using multiple devices, TensorFlow automatically split the tensors. Therefore, input shapes should be defined over batch_size.

• config – config for method, backbone, and loss.

• name – name of the model

build_loss()

Build losses according to configs.

build_model()

Build the model to be saved as self._model.

postprocess(inputs: Dict[str, tensorflow.Tensor], outputs: Dict[str, tensorflow.Tensor]) → Tuple[tensorflow.Tensor, Dict]

Return a dict used for saving inputs and outputs.

Parameters

• inputs – dict of model inputs

• outputs – dict of model outputs

Returns

tuple, indices and a dict. In the dict, each value is (tensor, normalize, on_label), where - normalize = True if the tensor need to be normalized to [0, 1] - on_label = True if the tensor depends on label

class deepreg.model.network.DVFModel(*args: Any, **kwargs: Any)

A registration model predicts DVF.

DDF is calculated based on DVF.

Init.

Parameters

• moving_image_size – (m_dim1, m_dim2, m_dim3)

• fixed_image_size – (f_dim1, f_dim2, f_dim3)

• index_size – number of indices for identify each sample

• labeled – if the data is labeled

• batch_size – total number of samples consumed per step, over all devices. When using multiple devices, TensorFlow automatically split the tensors. Therefore, input shapes should be defined over batch_size.

• config – config for method, backbone, and loss.

• name – name of the model

build_loss()

Build losses according to configs.

build_model()

Build the model to be saved as self._model.

postprocess(inputs: Dict[str, tensorflow.Tensor], outputs: Dict[str, tensorflow.Tensor]) → Tuple[tensorflow.Tensor, Dict]

Return a dict used for saving inputs and outputs.

Parameters

• inputs – dict of model inputs

• outputs – dict of model outputs

Returns

tuple, indices and a dict. In the dict, each value is (tensor, normalize, on_label), where - normalize = True if the tensor need to be normalized to [0, 1] - on_label = True if the tensor depends on label

class deepreg.model.network.RegistrationModel(*args: Any, **kwargs: Any)

Interface for registration model.

Init.

Parameters

• moving_image_size – (m_dim1, m_dim2, m_dim3)

• fixed_image_size – (f_dim1, f_dim2, f_dim3)

• index_size – number of indices for identify each sample

• labeled – if the data is labeled

• batch_size – total number of samples consumed per step, over all devices. When using multiple devices, TensorFlow automatically split the tensors. Therefore, input shapes should be defined over batch_size.

• config – config for method, backbone, and loss.

• name – name of the model

build_inputs() → Dict[str, tensorflow.keras.layers.Input]

Build input tensors.

Returns

dict of inputs.

abstract build_loss()

Build losses according to configs.

abstract build_model()

Build the model to be saved as self._model.

call(inputs: Dict[str, tensorflow.Tensor], training=None, mask=None) → Dict[str, tensorflow.Tensor]

Call the self._model.

Parameters

• inputs – a dict of tensors.

• training – training or not.

• mask – maks for inputs.

Returns

concat_images(moving_image: tensorflow.Tensor, fixed_image: tensorflow.Tensor, moving_label: Optional[tensorflow.Tensor] = None) → tensorflow.Tensor

Adjust image shape and concatenate them together.

Parameters

• moving_image – registration source

• fixed_image – registration target

• moving_label – optional, only used for conditional model.

Returns

get_config() → dict

Return the config dictionary for recreating this class.

log_tensor_stats(tensor: tensorflow.Tensor, name: str)

Log statistics of a given tensor.

Parameters

• tensor – tensor to monitor.

• name – name of the tensor.

plot_model(output_dir: str)

Save model structure in png.

Parameters

output_dir – path to the output dir.

abstract postprocess(inputs: Dict[str, tensorflow.Tensor], outputs: Dict[str, tensorflow.Tensor]) → Tuple[tensorflow.Tensor, Dict]

Return a dict used for saving inputs and outputs.

Parameters

• inputs – dict of model inputs

• outputs – dict of model outputs

Returns

tuple, indices and a dict. In the dict, each value is (tensor, normalize, on_label), where - normalize = True if the tensor need to be normalized to [0, 1] - on_label = True if the tensor depends on label

deepreg.model.network.dict_without(d: dict, key) → dict

Return a copy of the given dict without a certain key.

Parameters

• d – dict to be copied.

• key – key to be removed.

Returns

the copy without a key