Skip to content

I built a Neural Network model with PyTorch library to recognize the CIFAR10 dataset, which consists of 10 different object images. Since PyTorch supports CIFAR10 dataset, accessing and loading of train and test images was easy.

Notifications You must be signed in to change notification settings

azadkaratas/CIFAR10-CNN-Model

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CIFAR10-CNN-Model

I built a Neural Network model with PyTorch library to recognize the CIFAR10 dataset, which consists of 10 different object images. Since PyTorch supports CIFAR10 dataset, accessing and loading of train and test images was easy.

In the neural network model, I achieved 83% performance by using 5 convolutional layers followed by 2 fully connected layers. Network is implemented with GPU accelerated PyTorch library. For fast training of the network, I used the fast computers supported by Google Colaboratory. It significantly reduces amount of training time.

CNN architecture is shown below:

alt text

Another simple representation of the proposed model is as shown below:

Input Image -> Conv (1x1x56) -> BN -> ReLU -> Conv (2x2x84) -> BN -> ReLU -> Conv (2x2x128) -> BN -> ReLU -> Max Pool (2x2) -> Conv (2x2x256) -> BN -> ReLU -> Max Pool (2x2) -> Conv (2x2x512) -> BN -> ReLU -> Max Pool (2x2) -> Dropout -> FC1 -> FC2 -> CrossEntropyLoss

Result

This designed neural network model gives 83% accuracy for all 10000 images in CIFAR10 dataset. Thanks to Google Colaboratory, I managed to tune my hyperparameters in a quick way.

About

I built a Neural Network model with PyTorch library to recognize the CIFAR10 dataset, which consists of 10 different object images. Since PyTorch supports CIFAR10 dataset, accessing and loading of train and test images was easy.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published