lenet_5

→LeNet

lenet_5(pretrained: bool = False, overrides: object = {})

source edit

LeNet-5 feature-extracting backbone (no classification head).

Builds a LeNet with the paper-cited LeCun 1998 topology: three convolutions (C1: 1→6, C3: 6→16, C5: 16→120) interleaved with two $2\times2$ sub-sampling layers, using $\tanh$ activations and parameter-free average pooling. Approximately 44 k convolutional parameters (≈60 k including the F6 + output layers of the classifier variant). Designed for $1\times32 \times32$ grayscale digit inputs.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored — the returned model is randomly initialised.

**overridesobject= {}

Keyword overrides forwarded into LeNetConfig to customise individual fields. Use activation="relu" and pooling="max" for the modern ReLU + max-pool reimplementation, or in_channels=3 for RGB inputs.

Returns

LeNet

Backbone with the LeNet-5 configuration applied (or with overrides merged on top of it).

Notes

See LeCun et al., "Gradient-Based Learning Applied to Document Recognition", Proc. IEEE 86(11):2278–2324, 1998, Figure 2. The canonical metric is MNIST test-set error: the original paper reports 0.95% with the full distortion-augmented training recipe. No paper-cited "tiny / large" LeNet variants exist (H11) — pass activation/pooling overrides to get the common modernised flavours.

Examples

>>> import lucid
>>> from lucid.models.vision.lenet import lenet_5
>>> model = lenet_5()
>>> x = lucid.randn(1, 1, 32, 32)
>>> out = model(x)
>>> out.last_hidden_state.shape   # (B, 120, 1, 1)
(1, 120, 1, 1)

Used by 1

lucid.models.vision.lenet

>>> import lucid >>> from lucid.models.vision.lenet import lenet_5 >>> model = lenet_5() >>> x = lucid.randn(1, 1, 32, 32) >>> out = model(x) >>> out.last_hidden_state.shape # (B, 120, 1, 1) (1, 120, 1, 1)