This repository contains two different CNN image classifier trained using two different architectures. The first model is trained on All-CNN architecture which achieves 90% accuracy using Keras framework. The second model is trained on LeNet-5 architecture which achieves 74% accuracy using PyTorch.
This model is motivated from the Striving for Simplicity - All Convolution Net paper. The paper achieves 95.6% accuracy using the All-CNN architecture. My model (YGNet) has few changes in the architecture than the All-CNN architecture used in the paper. I've used max-pooling instead of making it a fully convolutional network. The total number of the trainable parameters remains the same i.e. ~1.3M. The changes used in my architecture were made on purpose to understand the effect and importance of certain layers in the convolutional neural network.
The source code allows to train the model from scratch. Also, the pre-trained model is present in this repo which generates 90% accuracy. The steps to build the source code is as follows,
$ git clone git@github.com:yogeshgajjar/CNN-CIFAR-10-image-classification.git
$ cd YGNet_Keras
$ python3 CIFAR-10_CNN.py train
This will train the model from scratch. The hyper-parameters can be tweaked as per required. To load the pre-trained model,
$ cd YGNet_Keras
$ python3 CIFAR-10_CNN.py load
The accuracy-epoch performance curve after running for 200 epochs
The loss-epoch performance curve after running for 200 epochs
LeNet-5 is a basic architecture which performance moderately well on CIFAR-10 dataset. LeNet-5 has around ~395k learnable parameters. My model achieved 74% accuracy using PyTorch. This model is trained on Google Colab.