class

LeNetForImageClassification

extendsPretrainedModelClassificationHeadMixin

LeNetForImageClassification(config: LeNetConfig)

source edit

LeNet-5 with F6 fully-connected layer and final classifier.

Combines a LeNet convolutional trunk with the original paper's two-layer fully-connected head: a hidden layer F6 mapping 120 → 84 followed by the output layer mapping 84 → config.num_classes. When labels are supplied to forward, a cross-entropy loss is returned alongside the logits; otherwise loss is None. Used as a teaching baseline on MNIST / Fashion-MNIST and as a parity reference for LeNet-style architectures.

Parameters

configLeNetConfig

Architecture spec. Use lenet_5_cls for the paper-cited configuration (10-way classifier, tanh, average pooling); pass a custom config to retarget num_classes or to switch the activation/pooling family.

Attributes

configLeNetConfig

Stored copy of the config that built this model.

featuresnn.Sequential

The C1-S2-C3-S4-C5 convolutional stack inherited from the backbone.

f6nn.Linear

Hidden fully-connected layer projecting 120 → 84.

act_f6nn.Module

Activation applied after F6 — lucid.nn.Tanh in the paper, lucid.nn.ReLU if config.activation="relu".

classifiernn.Module

Final linear projection 84 → num_classes, built by ClassificationHeadMixin._build_classifier.

Notes

From LeCun et al., "Gradient-Based Learning Applied to Document Recognition", Proc. IEEE 86(11):2278–2324, 1998, Figure 2. The original output layer was a Gaussian-RBF unit comparing F6 activations against fixed digit templates; this implementation uses the modern linear + cross-entropy head, which gave essentially the same accuracy in subsequent reimplementations. Loss is the standard categorical cross-entropy

\mathcal{L} = -\frac{1}{N} \sum_{n=1}^{N} \log \operatorname{softmax}(\text{logits}_n)_{\,y_n},

computed only when labels is not None.

Examples

Run inference on a small MNIST-style batch:
>>> import lucid
>>> from lucid.models.vision.lenet import lenet_5_cls
>>> model = lenet_5_cls()
>>> x = lucid.randn(4, 1, 32, 32)
>>> out = model(x)
>>> out.logits.shape
(4, 10)
>>> out.loss is None
True
Compute a training loss given labels:
>>> labels = lucid.tensor([0, 1, 2, 3], dtype=lucid.int64)
>>> out = model(x, labels=labels)
>>> out.loss.shape
()

Used by 2

Constructors

dunder

init

→None

__init__(config: LeNetConfig)

source edit

Instance methods

forward

→ImageClassificationOutput

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

source edit

Run inference on a small MNIST-style batch: >>> import lucid >>> from lucid.models.vision.lenet import lenet_5_cls >>> model = lenet_5_cls() >>> x = lucid.randn(4, 1, 32, 32) >>> out = model(x) >>> out.logits.shape (4, 10) >>> out.loss is None True Compute a training loss given labels: >>> labels = lucid.tensor([0, 1, 2, 3], dtype=lucid.int64) >>> out = model(x, labels=labels) >>> out.loss.shape ()