attention_unet

→AttentionUNetForSemanticSegmentation

attention_unet(pretrained: bool = False, overrides: object = {})

source edit

Attention U-Net (Oktay et al., MIDL 2018).

Builds an AttentionUNetForSemanticSegmentation with the standard configuration: 4-level encoder / decoder, base_channels = 64 (channel schedule 64 -> 128 -> 256 -> 512 -> 1024), in_channels = 1 (medical imaging default), and num_classes = 2. Soft attention gates on every skip connection suppress irrelevant encoder activations.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored.

**overridesobject= {}

Keyword overrides forwarded into AttentionUNetConfig (num_classes, in_channels for RGB inputs, base_channels, depth, bilinear).

Returns

AttentionUNetForSemanticSegmentation

Segmentation model with the standard Attention U-Net configuration applied (or with overrides merged on top of it).

Notes

See Oktay et al., "Attention U-Net: Learning Where to Look for the Pancreas", MIDL 2018 (arXiv:1804.03999). The defining attention-gate update is

\hat{x}^\ell = \sigma\!\bigl(\psi^\top \tanh(W_x x^\ell + W_g g^\ell)\bigr) \odot x^\ell,

where $x^\ell$ is the encoder feature at level $\ell$ and $g^\ell$ is the up-sampled decoder feature serving as the gating signal.

Examples

>>> import lucid
>>> from lucid.models.vision.attention_unet import attention_unet
>>> model = attention_unet(num_classes=4, in_channels=3)
>>> x = lucid.randn(1, 3, 256, 256)
>>> out = model(x)
>>> out.logits.shape
(1, 4, 256, 256)

Used by 1

lucid.models.vision.attention_unet

>>> import lucid >>> from lucid.models.vision.attention_unet import attention_unet >>> model = attention_unet(num_classes=4, in_channels=3) >>> x = lucid.randn(1, 3, 256, 256) >>> out = model(x) >>> out.logits.shape (1, 4, 256, 256)