MIR.models.Deformable_Swin_Transformer_v2ΒΆ

Deformable Swin Transformer v2 backbone for 3D registration.

Functions

deform_window_partition(x, window_size)

Partition a per-head channel-last 3D map into non-overlapping windows.

window_partition(x, window_size)

Partition a channel-last 3D feature map into non-overlapping windows.

window_reverse(windows, window_size, H, W, L)

Merge windows produced by window_partition() back into a 3D feature map.

Classes

BasicLayer(*args, **kwargs)

A basic Swin Transformer layer for one stage.

DefSwinTransformerV2(*args, **kwargs)

Deformable 3D Swin Transformer (v2) encoder for a moving/fixed image pair.

LayerNormProxy(*args, **kwargs)

LayerNorm over the channel dimension of a channel-first 3D tensor.

Mlp(*args, **kwargs)

Feed-forward MLP block used inside transformer blocks.

Offset_block(*args, **kwargs)

Offset prediction block for deformable attention.

Offset_block0(*args, **kwargs)

Offset prediction block for deformable attention (lightweight).

PatchEmbed(*args, **kwargs)

Image to Patch Embedding using a strided 3D convolution.

PatchMerging(*args, **kwargs)

Patch Merging Layer.

SinPositionalEncoding3D(*args, **kwargs)

Fixed 3D sinusoidal positional encoding.

SinusoidalPositionEmbedding(*args, **kwargs)

Fixed sinusoidal (1D, absolute) position embedding for a token sequence.

SwinTransformerBlock(*args, **kwargs)

Swin Transformer Block.

WindowAttention(*args, **kwargs)

Window based multi-head cross attention between moving and fixed features with relative position bias.

XMlp(*args, **kwargs)

Cross-gated MLP conditioned on a second token stream.