Description

This repository contains implementations of a few convolutional neural networks (CNNs) for the MNIST dataset and CIFAR-10 dataset data sets. These CNNs use PyTorch and the Keras API with TensorFlow backend. Note that TensorFlow versions >= 2.0 will throw deprecated software warnings.

The PyTorch MNIST CNN includes 2 convolutional layers, a linear layer with ReLU activation, and a linear layer with log_softmax. The model is based on this architecture and achieves an accuracy of 99.0%.

The Keras MNIST CNN includes 2 convolution layers and a flatten layer and is based on this architecture from the Keras dev team. This model achieves an accuracy of 99.2%.

The Keras CIFAR-10 CNN includes 6 convolution layers and a flatten layer and is based on this architecture by Abhijeet Kumar. This model achieves an accuracy of 89%.

CNN Architectures

MNIST with PyTorch	MNIST with Keras	CIFAR-10 with Keras

Demo Tutorial

• To run the demo, call the function RunDemo.main(model_name=model_name, API=API, use_cached=use_cached).

• model_name can take values CIFAR, MNIST, or [] (default).

• API can take values Keras or PyTorch.

• use_cached is boolean (default True).

Dependencies

• torch, torchvision, torchsummary
• graphviz, torchviz
• keras
• TensorFlow, CNTK, or Theano