Crate microcnn 

A minimal CNN framework in Rust with INT8 and INT4 quantization support.

This crate provides building blocks for constructing and running convolutional neural networks in FP32, INT8, and INT4 precision. It includes a reference LeNet-5 implementation for MNIST digit classification.

Example

use microcnn::lenet::lenet;

let mut net = lenet(false);
net.load("data/lenet.raw");

Re-exports

pub use tensor::TensorI8;

pub use tensor::TensorI4;

Modules

arc

Quantization utilities and model architecture (LeNet).

benchmark

conv

Convolution algorithm implementations (Naive, Im2col, Winograd, FFT).

lenet

loader

MNIST dataset loaders.

metrics

Benchmarking utilities for comparing FP32/INT8/INT4 performance.

mnist

network

Neural network layers and network types (FP32, INT8, INT4).

quantization

Re-export at legacy paths.

tensor

FP32, INT8, and INT4 tensors.

tensor_i4

tensor_i8

Alias modules so crate::tensor_i8::TensorI8 etc. still resolve.