efficientformer_l1

→EfficientFormer

efficientformer_l1(pretrained: bool = False, overrides: object = {})

source edit

EfficientFormer-L1 backbone (Li et al., 2022).

Builds the canonical EfficientFormer-L1 configuration: depths=(3, 2, 6, 4), embed_dims=(48, 96, 224, 448), mlp_ratios=(4.0, 4.0, 4.0, 4.0), drop_path_rate=0.0. Approximately 12.3M parameters — the smallest, lowest-latency variant in the paper.

Model Size

Parameters

pretrainedbool= False

If True, loads ImageNet-1k pretrained weights when available in the model zoo. Defaults to False.

**overridesobject= {}

Keyword overrides on top of the canonical L1 config. Each override must match a field of EfficientFormerConfig.

Returns

EfficientFormer

An EfficientFormer backbone returning a flat $(B, 448)$ feature.

Notes

EfficientFormer-L1 reaches 79.2% top-1 on ImageNet-1k at MobileNetV2-class on-device latency (Li et al., 2022, Table 4). See arXiv:2206.01191.

Examples

>>> import lucid
>>> from lucid.models.vision.efficientformer import efficientformer_l1
>>> model = efficientformer_l1()
>>> x = lucid.randn(1, 3, 224, 224)
>>> feat = model.forward_features(x)
>>> feat.shape
(1, 448)

Used by 1

lucid.models.vision.efficientformer

>>> import lucid >>> from lucid.models.vision.efficientformer import efficientformer_l1 >>> model = efficientformer_l1() >>> x = lucid.randn(1, 3, 224, 224) >>> feat = model.forward_features(x) >>> feat.shape (1, 448)