CIFAR-10 Image Classifier
This project demonstrates a CIFAR-10 image classification model. You can view images from the dataset, see the model's predictions, and compare them with the actual labels.
Image
Prediction
Label
Training Method
The model was trained using a ResNet architecture on the CIFAR-10 dataset. The ResNet architecture used in this project includes:
- Convolutional layers with batch normalization and ReLU activation.
- Residual blocks to mitigate the vanishing gradient problem in deep networks.
- Global average pooling layer followed by a fully connected layer for classification.
The training process involved data augmentation techniques such as:
- Random horizontal flipping and rotations to make the model invariant to these transformations.
- Random cropping and scaling to handle different object sizes and aspect ratios.
- Color jittering to simulate different lighting conditions.
Detailed Training Steps:
- Implemented basic network layers including ReLU activation, fully-connected layers, and softmax activation for output.
- Derived gradients and performed backpropagation for each layer to update weights.
- Applied cross-entropy loss for classification tasks and optimized using AdamW optimizer with learning rate decay.
- Experimented with different hyperparameters such as learning rate and hidden layer size to achieve the best performance.
Model Performance:
Achieved a test accuracy of 93.6% on the CIFAR-10 test set after 100 epochs, contributing to a score that placed top 3 on the Kaggle leaderboard during a university competition.