MNIST Cls Cpp

A demo of using libtorch to train and test models on MNIST dataset for classification.

Build

cd MNIST-cls-cpp

cmake -S . -B build

cmake --build build

Train and test

# train
./build/train -p path/to/mnist/dataset

# test
./build/test -p path/to/mnist/dataset -m path/to/saved/model

Included Models

Simple Net

A simple classifier with three Fully Connected Layer.

LeNet 5

Implement of the classical model LeNet5, according to LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.

AlexNet

Implement of the classical model AlexNet, according to Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[C] International Conference on Neural Information Processing Systems. Curran Associates Inc. 2012:1097-1105. with a little modified.

Benchmark

Epoch: 5, Batch size: 32, Learing rate: 0.01.

The Train Time is measured on the train set with CPU i5-9300H, and the Correct% is measured on the test set.

	Correct%	Train Time
Simple Net	88.66	13723ms
LeNet-5	99.05	56465ms
AlexNet	98.47	too long