ViT-Classification-CIFAR10

Model Description

This model is a Vision Transformer (ViT) architecture trained on the CIFAR-10 dataset for image classification. It is trained from scratch without pre-training on a larger dataset.

Metrics:

Test accuracy: 82.04%
Test loss: 0.5560

Training Configuration

Hardware: NVIDIA RTX 3090

Training parameters:

Epochs: 200
Batch size: 2048
Input size: 3x32x32
Patch size: 4
Sequence length: 8*8
Embed size: 128
Num of layers: 12
Num of heads: 4
Forward multiplier: 2
Dropout: 0.1
Optimizer: AdamW

Intended Uses & Limitations

This model is intended for practice purposes and exploration of ViT architectures on the CIFAR-10 dataset. It can be used for image classification tasks on similar datasets.

Limitations:

This model is trained on a relatively small dataset (CIFAR-10) and might not generalize well to unseen data.
Training is done without fine-tuning, potentially limiting its performance compared to a fine-tuned model.
Training is performed on a single RTX 3090.

Training Data

The model is trained on the CIFAR-10 dataset, containing 60,000 32x32 color images in 10 classes.

Training set: 50,000 images
Test set: 10,000 images

Data Source: https://paperswithcode.com/dataset/cifar-10

Documentation

GitHub Repository: ViT-Classification-CIFAR10