class

ResNet

extendsPretrainedModelBackboneMixin

ResNet(config: ResNetConfig)

source edit

ResNet feature-extracting backbone (no classification head).

Implements the ResNet topology from He et al. in its v1.5 form: the downsampling stride sits on the bottleneck's 3x3 convolution. The paper says only that "downsampling is performed by conv3_1, conv4_1, and conv5_1 with a stride of 2" without naming which conv carries it; the original MSRA release put it on the first 1x1 (v1). v1.5 is what every modern reference builds and what the shipped checkpoints were trained as., "Deep Residual Learning for Image Recognition", CVPR 2016 (arXiv:1512.03385), restricted to the convolutional trunk: a 7×7 stem at stride 2, a 3×3 max-pool at stride 2, then four stages of residual blocks producing feature maps at strides 4, 8, 16, and 32 relative to the input. No global pooling, no flatten, and no linear classifier — those live on ResNetForImageClassification.

The block topology is selected by config.block_type: ResNet-18 and ResNet-34 use _BasicBlock, while the deeper ResNet-50 / 101 / 152 / 200 / 269 and Wide ResNet variants use _Bottleneck. This class is the canonical feature extractor consumed by detection / segmentation heads (FPN, Faster R-CNN, Mask R-CNN, …) and by self-supervised pretraining pipelines.

Parameters

configResNetConfig

Frozen architecture spec. Use the factory functions (resnet_18, resnet_50, …) for paper-cited variants — they instantiate the correct config for you.

Attributes

configResNetConfig

Stored copy of the config that built this model.

stemnn.Sequential

7×7 conv (stride 2) → BatchNorm → ReLU — reduces input from

H \times W

H/2 \times W/2

maxpoolnn.MaxPool2d

3×3 max-pool with stride 2 and padding 1 — further reduces spatial size to

H/4 \times W/4

layer1, layer2, layer3, layer4nn.Sequential

The four residual stages. layer1 keeps the input spatial size; layer2/layer3/layer4 each halve it. Output channel counts are hidden_sizes[i] * block.expansion.

feature_infolist[FeatureInfo]

Per-stage descriptor (channels + reduction factor) exposed via BackboneMixin for downstream FPN / decoder modules.

Notes

Each residual stage applies the standard skip-connection update

y_{l+1} = \sigma\!\left(F(y_l, W_l) + W_s\, y_l\right),

where $F$ is the in-block transformation, $\sigma$ is ReLU, and $W_s$ is either the identity (matched dimensions) or a 1×1 strided projection (dimension change at stage boundaries). The residual formulation lets gradients flow directly through the identity branch, making it practical to train networks with over a hundred layers — the central contribution of the original paper.

The output of forward is a BaseModelOutput whose last_hidden_state has shape (B, C_4, H/32, W/32) with $C_4 = \text{hidden\_sizes}[3] \times \text{expansion}$ — e.g. (B, 512, H/32, W/32) for ResNet-18 and (B, 2048, H/32, W/32) for ResNet-50.

Examples

Build a ResNet-50 backbone and run a single forward pass:
>>> import lucid
>>> from lucid.models.vision.resnet import resnet_50
>>> backbone = resnet_50()
>>> x = lucid.randn(2, 3, 224, 224)   # (batch, channels, H, W)
>>> out = backbone(x)
>>> out.last_hidden_state.shape
(2, 2048, 7, 7)
Inspect the per-stage feature map descriptors for FPN integration:
>>> info = backbone.feature_info
>>> [(fi.stage, fi.num_channels, fi.reduction) for fi in info]
[(1, 256, 4), (2, 512, 8), (3, 1024, 16), (4, 2048, 32)]

Used by 2

Constructors

dunder

init

→None

__init__(config: ResNetConfig)

source edit

Initialise the module and validate the supplied config.

Parameters

configModelConfig

Must be an instance of the subclass's declared config_class.

Raises

TypeError

If config_class has not been set on the concrete subclass, or if config is not an instance of config_class.

Properties

prop

feature_info

→list[FeatureInfo]

feature_info: list[FeatureInfo]

source edit

Instance methods

forward

→BaseModelOutput

forward(x: Tensor)

source edit

Run the model with every covered weight compressed.

forward_features

→Tensor

forward_features(x: Tensor)

source edit

Pooled (B, C) feature vector, post head-norm.

The per-stage maps this backbone's feature_info describes are left on stage_features for pyramid consumers.

Build a ResNet-50 backbone and run a single forward pass: >>> import lucid >>> from lucid.models.vision.resnet import resnet_50 >>> backbone = resnet_50() >>> x = lucid.randn(2, 3, 224, 224) # (batch, channels, H, W) >>> out = backbone(x) >>> out.last_hidden_state.shape (2, 2048, 7, 7) Inspect the per-stage feature map descriptors for FPN integration: >>> info = backbone.feature_info >>> [(fi.stage, fi.num_channels, fi.reduction) for fi in info] [(1, 256, 4), (2, 512, 8), (3, 1024, 16), (4, 2048, 32)]