Deep Learning Kaggle-CIFAR-10-Object-Recognition-in-Images

Kaggle-CIFAR-10-Object-Recognition-in-Images

Implementations of different CNN models to recognize CIFAR-10 Image by MXNet

Introduction

This repository is about different CNN architecture by using MXNet framework for CIFAR-10 dataset in kaggle. For now, the my resnet164_v2 model got 0.93370 which is over rank 7, and my resnet18 model got 0.90480. Some of the ideas come from gluon community, and welcome to join the family of MXNet gluon.

Requrements

• MXNet

Relative Papers

• Resnet18
  • Deep Residual Learning for Image Recognition
• Resnet164_v2
  • Identity Mappings in Deep Residual Networks

Accuracy

Before the training, we will do some image augmentations such as resize, random brightness, random hue, random resized crop and random flip left right. For both models, the total epochs are 200, batch size is 128, initial learning rate is 0.1, and the learning rate decay is 90.

Resnet18

The code of this model is resnet_18_cifar10.py or the jupyter notebook resnet_18_cifar10.ipynb After 200 epochs, loss is 0.004550, training accuracy is 0.998800, and kaggle score is 0.90480.

Resnet164_v2

The code of this model is resnet_164v2_cifar10.py or the jupyter notebook resnet_164v2_cifar10.ipynb After 200 epochs, loss is 0.004943, training accuracy is 0.998640, and kaggle score is 0.93370.

Future

• Add more image augmentatation and modify the hyperparameter to optimize the models
• Implement more models such as AlexNet, DenseNet, and GooLeNet to recognize CIFAR-10, and even CIFAR-100 and ImageNet.