class

FCNForSemanticSegmentation

extendsSemanticSegmentationModel

FCNForSemanticSegmentation(config: FCNConfig)

source edit

Fully Convolutional Network for semantic segmentation (Long et al., CVPR 2015).

The original deep-learning semantic segmentation architecture: a classification backbone is repurposed as a fully convolutional network by replacing fully-connected layers with 1x1 convolutions and adding a final upsampling stage to recover input resolution. In Lucid's implementation the backbone is a dilated ResNet (layer3 dilation = 2, layer4 dilation = 4) so the deepest feature map stays at stride 8 instead of stride 32, giving denser spatial sampling without sacrificing the receptive field.

A primary FCN head produces num_classes logits from the C5 feature, and an auxiliary head on C4 stabilises optimisation (used at training time only). Bilinear interpolation lifts both heads to the input resolution.

Parameters

configFCNConfig

Frozen architecture spec. Use fcn_resnet50 / fcn_resnet101 for the standard variants.

Attributes

configFCNConfig

Stored copy of the config that built this model.

backbone_DilatedResNet

Dilated ResNet trunk producing C4 (stride 8, used by aux head) and C5 (stride 8, used by main head) — both at the same resolution thanks to atrous convolution in layer4.

classifier_FCNHead

3x3 conv -> BatchNorm -> ReLU -> Dropout -> 1x1 conv producing num_classes channels from the C5 feature.

aux_classifier_FCNHead

Auxiliary head with the same structure applied to C4; its loss is weighted by 0.4 and added to the main loss only during training.

Notes

See Long et al., "Fully Convolutional Networks for Semantic Segmentation", CVPR 2015 (arXiv:1411.4038). The bilinear upsampling at the end is equivalent to a learnable transpose convolution initialised to bilinear weights — but kept fixed in this modernised dilated-ResNet variant. Training loss is

\mathcal{L} = \mathcal{L}_\mathrm{CE}(\mathrm{main}, y) + 0.4\, \mathcal{L}_\mathrm{CE}(\mathrm{aux}, y),

with categorical cross-entropy averaged over the non-ignored pixels. The paper (§4.1) defines the objective as an unnormalised spatial sum, and its quoted learning rates are calibrated to that scale; the reference recipe that produced the shipped checkpoints uses the mean instead, and this implementation follows the reference so those weights reproduce. Scale the learning rate accordingly if switching to the paper's formulation.

Examples

>>> import lucid
>>> from lucid.models.vision.fcn import fcn_resnet50
>>> model = fcn_resnet50()
>>> x = lucid.randn(1, 3, 512, 512)
>>> out = model(x)
>>> out.logits.shape    # (B, num_classes, H, W)
(1, 21, 512, 512)

Used by 2

Constructors

dunder

init

→None

__init__(config: FCNConfig)

source edit

Instance methods

forward

→SemanticSegmentationOutput

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

source edit

Run FCN forward pass.

Parameters

xTensor

(B, C, H, W) image batch.

targetsTensor | None= None

Optional (B, H, W) integer ground-truth masks.

Returns

SemanticSegmentationOutput

SemanticSegmentationOutput with logits (B, num_classes, H, W)

>>> import lucid >>> from lucid.models.vision.fcn import fcn_resnet50 >>> model = fcn_resnet50() >>> x = lucid.randn(1, 3, 512, 512) >>> out = model(x) >>> out.logits.shape # (B, num_classes, H, W) (1, 21, 512, 512)