deepali.networks.unet

U-net model architectures.

Module Contents

Classes

UNetLayerConfig	Base class of configuration data classes.
UNetDownsampleConfig	Base class of configuration data classes.
UNetUpsampleConfig	Base class of configuration data classes.
UNetOutputConfig	Base class of configuration data classes.
UNetEncoderConfig	Base class of configuration data classes.
UNetDecoderConfig	Base class of configuration data classes.
UNetConfig	Base class of configuration data classes.
UNetEncoder	Downsampling path of U-net model.
UNetDecoder	Upsampling path of U-net model.
SequentialUNet	Sequential U-net architecture.
UNet	U-net with optionally multiple output layers.

Functions

unet_conv_block(→ torch.nn.Module)

Create U-net block of convolutional layers.

class UNetLayerConfig

Base class of configuration data classes.

class UNetDownsampleConfig

Base class of configuration data classes.

class UNetUpsampleConfig

Base class of configuration data classes.

class UNetOutputConfig

Base class of configuration data classes.

class UNetEncoderConfig

Base class of configuration data classes.

property num_levels: int

Number of spatial encoder levels.

class UNetDecoderConfig

Base class of configuration data classes.

property num_levels: int

Number of spatial decoder levels, including bottleneck input.

classmethod from_encoder(encoder: Union[UNetEncoder, UNetEncoderConfig], residual: Optional[bool] = None, **kwargs) → UNetDecoderConfig

Derive decoder configuration from U-net encoder configuration.

class UNetConfig

Base class of configuration data classes.

unet_conv_block(spatial_dims: int, in_channels: int, out_channels: int, kernel_size: deepali.core.typing.ScalarOrTuple[int] = 3, stride: deepali.core.typing.ScalarOrTuple[int] = 1, padding: Optional[deepali.core.typing.ScalarOrTuple[int]] = None, padding_mode: Union[deepali.core.enum.PaddingMode, str] = 'zeros', dilation: deepali.core.typing.ScalarOrTuple[int] = 1, groups: int = 1, init: str = 'default', bias: Optional[Union[bool, str]] = None, norm: deepali.networks.layers.NormArg = None, acti: deepali.networks.layers.ActivationArg = None, order: str = 'CNA', num_layers: Optional[int] = None) → torch.nn.Module

Create U-net block of convolutional layers.

class UNetEncoder(spatial_dims: int, in_channels: Optional[int] = None, config: Optional[UNetEncoderConfig] = None, conv_block: Optional[ModuleFactory] = None, input_layer: Optional[ModuleFactory] = None)

Downsampling path of U-net model.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

output_size(in_size: deepali.core.typing.ScalarOrTuple[int]) → deepali.core.typing.ScalarOrTuple[int]

Calculate output size of last feature map for a tensor of given spatial input size.

output_sizes(in_size: deepali.core.typing.ScalarOrTuple[int]) → List[deepali.core.typing.ScalarOrTuple[int]]

Calculate output sizes of feature maps for a tensor of given spatial input size.

class UNetDecoder(spatial_dims: int, in_channels: Optional[int] = None, config: Optional[UNetDecoderConfig] = None, conv_block: Optional[ModuleFactory] = None, input_layer: Optional[ModuleFactory] = None, output_all: bool = False)

Upsampling path of U-net model.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

classmethod from_encoder(encoder: Union[UNetEncoder, UNetEncoderConfig], residual: Optional[bool] = None, **kwargs) → UNetDecoder

Create U-net decoder given U-net encoder configuration.

output_size(in_size: deepali.core.typing.ScalarOrTuple[int]) → deepali.core.typing.ScalarOrTuple[int]

Calculate output size for an initial feature map of given spatial input size.

output_sizes(in_size: deepali.core.typing.ScalarOrTuple[int]) → List[deepali.core.typing.ScalarOrTuple[int]]

Calculate output sizes for an initial feature map of given spatial input size.

class SequentialUNet(spatial_dims: int, in_channels: Optional[int] = None, out_channels: Optional[int] = None, config: Optional[UNetConfig] = None, conv_block: Optional[ModuleFactory] = None, output_layer: Optional[ModuleFactory] = None, bridge_layer: Optional[ModuleFactory] = None)

Sequential U-net architecture.

The final module of this sequential module either outputs a tuple of feature maps at the different resolution levels (out_channels=None), the final decoded feature map at the highest resolution level (out_channels == config.decoder.out_channels and output_layers=None), or a tensor with specified number of out_channels as produced by a final output layer otherwise. Note that additional layers (e.g., a custom output layer or post-output layers) can be added to the initialized sequential U-net using add_module().

Initializes internal Module state, shared by both nn.Module and ScriptModule.

output_size(in_size: deepali.core.typing.ScalarOrTuple[int]) → deepali.core.typing.ScalarOrTuple[int]

Calculate spatial output size given an input tensor with specified spatial size.

class UNet(spatial_dims: int, in_channels: Optional[int] = None, out_channels: Optional[int] = None, output_modules: Optional[Mapping[str, torch.nn.Module]] = None, output_indices: Optional[Union[Mapping[str, int], int]] = None, config: Optional[UNetConfig] = None, conv_block: Optional[ModuleFactory] = None, bridge_layer: Optional[ModuleFactory] = None, output_layer: Optional[ModuleFactory] = None, output_name: str = 'output')

U-net with optionally multiple output layers.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

output_is_dict() → bool

Whether model output is dictionary of output tensors.

output_is_tensor() → bool

Whether model output is a single output tensor.

output_is_tuple() → bool

Whether model output is tuple of decoded feature maps.