class

EfficientNet

extendsPretrainedModelBackboneMixin

EfficientNet(config: EfficientNetConfig)

source edit

EfficientNet feature-extracting backbone (no classification head).

Implements the compound-scaled topology from Tan & Le, "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks", ICML 2019 (arXiv:1905.11946). The baseline (B0) is a NAS-designed network of MBConv blocks — inverted-residual bottlenecks (MobileNet-v2 style) augmented with a squeeze-and-excitation module and the swish (SiLU) activation. Successive B-variants (B0 through B7) apply the paper's compound-scaling rule

d = \alpha^\phi, \quad w = \beta^\phi, \quad r = \gamma^\phi, \qquad \alpha \beta^2 \gamma^2 \approx 2,

to scale depth, width, and resolution jointly along a single coefficient $\phi$ .

The body is a 3×3 stem (stride 2) followed by seven MBConv stages and a final 1×1 head expansion to round(1280 · width_mult) channels. Stochastic depth (drop_connect_rate) is applied linearly across blocks as a strong regulariser at large depth.

Parameters

configEfficientNetConfig

Frozen architecture spec. Use the factory functions (efficientnet_b0 through efficientnet_b7) for paper-cited variants.

Attributes

configEfficientNetConfig

Stored copy of the config that built this model.

featuresnn.Sequential

Stem + seven MBConv stages + 1×1 head. Channel and depth counts scale with config.width_mult and config.depth_mult.

avgpoolnn.AdaptiveAvgPool2d

Global average pool collapsing the final feature map to (B, C, 1, 1).

feature_infolist[FeatureInfo]

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

Notes

Each MBConv block applies

x \to \mathrm{Expand}_{1\times1} \to \mathrm{DW}_{k\times k} \to \mathrm{SE} \to \mathrm{Proj}_{1\times1},

with a residual when stride is 1 and channel counts match. During training the residual branch is randomly dropped with probability drop_connect_rate * block_idx / (total - 1) — deeper blocks get a higher drop probability, matching the "deep network regularisation" recipe of stochastic depth.

Examples

Build an EfficientNet-B0 backbone and run a forward pass:
>>> import lucid
>>> from lucid.models.vision.efficientnet import efficientnet_b0
>>> backbone = efficientnet_b0()
>>> x = lucid.randn(2, 3, 224, 224)
>>> out = backbone(x)
>>> out.last_hidden_state.shape
(2, 1280, 1, 1)

Used by 2

Constructors

dunder

init

→None

__init__(config: EfficientNetConfig)

source edit

Initialise the ReLU module. See the class docstring for parameter semantics.

Properties

prop

feature_info

→list[FeatureInfo]

feature_info: list[FeatureInfo]

source edit

Instance methods

forward

→BaseModelOutput

forward(x: Tensor)

source edit

Parameters

features

[P3, P4, P5, P6, P7] (finest → coarsest).

Returns

BaseModelOutput

Fused [P3_out, P4_out, P5_out, P6_out, P7_out].

forward_features

→Tensor

forward_features(x: Tensor)

source edit

Build an EfficientNet-B0 backbone and run a forward pass: >>> import lucid >>> from lucid.models.vision.efficientnet import efficientnet_b0 >>> backbone = efficientnet_b0() >>> x = lucid.randn(2, 3, 224, 224) >>> out = backbone(x) >>> out.last_hidden_state.shape (2, 1280, 1, 1)