Module mnist 

LeNet-5 — Yann LeCun’s convolutional network for MNIST (1998).

Original paper: “Gradient-Based Learning Applied to Document Recognition” LeCun, Bottou, Bengio, Haffner (1998).

Architecture (adapted to use ReLU in place of the original tanh/sigmoid):

Input         [N,  1, 28, 28]
Conv2d(1→6,   5×5, pad=2)  →  [N,  6, 28, 28]   (same-padding keeps spatial dims)
ReLU
AvgPool2d(2×2, stride=2)   →  [N,  6, 14, 14]
Conv2d(6→16,  5×5, pad=0)  →  [N, 16, 10, 10]
ReLU
AvgPool2d(2×2, stride=2)   →  [N, 16,  5,  5]
Flatten                    →  [N, 400]
Linear(400→120)
ReLU
Linear(120→84)
ReLU
Linear(84→10)
Softmax(dim=1)             →  [N, 10]  class probabilities

This example traces the model symbolically using SymTensor, extracts the computation graph, and prints every node in topological order.

Re-exports

pub use dataset::MnistBatch;

pub use dataset::MnistDataset;

Modules

dataset

MNIST parquet dataset reading. Reader for HuggingFace MNIST parquet files.

Functions

mnist

Another LeNet-5 pipeline (see mnist_lenet5), with biased convolutions and same-padding on the first layer.

mnist_lenet5

LeNet-5 (valid-convolution variant) — no padding, no bias, returns raw logits. Use this for training with a cross-entropy loss kernel that computes log-softmax internally.

mnist_mlp

Fully-connected MLP for MNIST — no conv layers, suitable for end-to-end training.

mnist_valid

LeNet-5 (valid-convolution variant) — no padding, no bias, with final softmax. Use this for inference when you need class probabilities.