lenet-ts

🧠 LeNet Convolutional Neural Network with Tensorflow and Typescript: 28x28 grayscale hand written digits recognition

/!\ This is a heavy repository as the MNIST dataset is pushed with it for convenience.

Neural Network Architecture

Layer Type	Input Shape	Info
2D Convolutional	28x28x1 (grayscale image size)	kernel of size 5, stride of 1, 8 filters, relu activation
2D Max Pooling	24x24x8	pool size of 2x2, stride of 2;2
2D Convolutional	12x12x8	kernel of size 5, stride of 1, 16 filters, relu activation
2D Max Pooling	8x8x128	pool size of 2x2, stride of 2;2
Fully Connected Layer	4x4x128	relu activation
Fully Connected Layer	64	softmax activation
Output	10	One-Hot style output

Usage

Install dependencies

npm install

Build Sources

npm run build

Train Model

npm run train ./dataset/train.csv <path-to-dump-model>

Predict

npm run start ./dataset/test.csv <path-to-dump-model>/model.json <amount-to-guess>

Output