class

SwinTransformerForImageClassification

extendsPretrainedModelClassificationHeadMixin

SwinTransformerForImageClassification(config: SwinConfig)

source edit

Swin Transformer with a linear classification head (Liu et al., 2021).

Wraps the same hierarchical trunk as SwinTransformer (patch embedding → four stages of window / shifted-window attention → LayerNorm) and adds a global average pool plus a single nn.Linear classification head that maps the $(B, 8C)$ feature to config.num_classes logits:

\text{logits} = W_{\text{head}}\, \mathrm{GAP}(\mathrm{LN}(z^{L})) + b_{\text{head}}, \qquad W_{\text{head}} \in \mathbb{R}^{C_{\text{out}} \times 8C}.

Pass labels to forward to compute the cross-entropy loss in the same pass.

Parameters

configSwinConfig

Architecture specification. Must set num_classes to the desired number of output categories (default 1000 for ImageNet). See SwinConfig.

Attributes

patch_embed_PatchEmbed

Strided patch-extraction convolution + LayerNorm.

stagesnn.ModuleList

Four hierarchical Swin stages with patch-merging downsamplers.

normnn.LayerNorm

Final LayerNorm applied to the channel-last feature map.

avgpoolnn.AdaptiveAvgPool2d

1 \times 1

adaptive average pool over spatial dims.

classifiernn.Linear

Final linear projection of width (num_classes, 8 * embed_dim) built via ClassificationHeadMixin._build_classifier.

Notes

Reference: Ze Liu et al., "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows", ICCV 2021, arXiv:2103.14030. Swin-T / S / B / L reach 81.3 / 83.0 / 83.5 / 86.4 % top-1 on ImageNet-1k (Table 1 of the paper, 224x224 input).

Examples

End-to-end inference with the default Swin-T classifier:
>>> import lucid
>>> from lucid.models.vision.swin import (
...     SwinConfig, SwinTransformerForImageClassification,
... )
>>> model = SwinTransformerForImageClassification(SwinConfig())
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.logits.shape
(1, 1000)

Used by 2

Constructors

dunder

init

→None

__init__(config: SwinConfig)

source edit

Instance methods

forward

→ImageClassificationOutput

forward(x: Tensor, labels: Tensor | None = None)

source edit

End-to-end inference with the default Swin-T classifier: >>> import lucid >>> from lucid.models.vision.swin import ( ... SwinConfig, SwinTransformerForImageClassification, ... ) >>> model = SwinTransformerForImageClassification(SwinConfig()) >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.logits.shape (1, 1000)