Image Classification with CIFAR 10 Dataset

Introduction

This very beginner friendly deep learning project for those trying to get into the world of deep learning. We are using the CIFAR 10 dataset to train a convolutional neural network to classify images.

Dataset

Cifar 10 is a dataset of 60,000 32x32 color images in 10 classes. Each image is labeled with one of the 10 digits. It is a subset of the ImageNet dataset.

Downloading the dataset

To download the dataset, you can use the following command:

sh download/download_cifar10.sh

Unpickling the dataset

To unpickle the dataset, you can use the following command:

python data/unpickle_cifar10.py

Usage

Training

To train the network, you can use the following command and retrain model:

python main.py

Model will be saved as model.pth in the current directory.

Ps. There is no current support for training model on starting from current checkpoint, but it'll be easy to add.

Testing

To test the network and reproduce the results, you can use the following command:

python test.py

Results:

epoch	train_loss	val_loss	train_acc	val_acc
1	0.287	0.192	0.896	0.937
2	0.155	0.174	0.948	0.945
3	0.111	0.116	0.963	0.960
4	0.095	0.180	0.968	0.946
5	0.081	0.391	0.973	0.910
6	0.074	0.123	0.975	0.960

Predict Custom Image

To predict the any image, you can use the following command:

python predict.py --fp <path to image>

If you don't specify the path to image, it'll select random image from test set. and will save the result as prediction.png in the docs directory.

Examples

Airplane:

> python predict.py --fp docs/plane.jpg

Most possible class: airplane
Probabilities: {'airplane': '0.89', 'automobile': '0.00', 'bird': '0.00', 'cat': '0.00', 'deer': '0.02', 'dog': '0.00', 'frog': '0.00', 'horse': '0.00', 'ship': '0.09', 'truck': '0.00'}

Bird:

> python predict.py --fp docs/bird.png
Most possible class: bird
Probabilities: {'airplane': '0.06', 'automobile': '0.00', 'bird': '0.94', 'cat': '0.00', 'deer': '0.00', 'dog': '0.00', 'frog': '0.00', 'horse': '0.00', 'ship': '0.00', 'truck': '0.00'}

> python predict.py --fp docs/cat.jpeg
Most possible class: ship
Probabilities: {'airplane': '0.08', 'automobile': '0.08', 'bird': '0.19', 'cat': '0.13', 'deer': '0.00', 'dog': '0.00', 'frog': '0.11', 'horse': '0.00', 'ship': '0.37', 'truck': '0.03'}

😲

Conclusion

It seems this is not very good model, probably dataset not have complicated pictures. However, it's very easy and fun to get started with deep learning with this dataset. I hope you enjoy it and find it useful.

iboraham/image-classification-CIFAR10