vit-cifar100-lora

ViT-S Fine-tuned on CIFAR-100 with LoRA

This model is a Vision Transformer Small (ViT-S/16) pretrained on ImageNet and fine-tuned on CIFAR-100 using LoRA (Low-Rank Adaptation) via the PEFT library.

Model Details

• Base Model: vit_small_patch16_224 (timm)
• Dataset: CIFAR-100 (100 classes)
• Fine-tuning Method: LoRA (PEFT)
• LoRA Configuration:
  • Rank: 8
  • Alpha: 8
  • Dropout: 0.1
  • Target Modules: QKV attention weights

Training Details

• Epochs: 10
• Batch Size: 16
• Learning Rate: 0.0001
• Optimizer: AdamW
• Scheduler: Cosine Annealing with Warmup

Results

• Best Validation Accuracy: 88.26%

Usage

import timm
from peft import PeftModel, LoraConfig
import torch

# Load base model
base_model = timm.create_model('vit_small_patch16_224', pretrained=True, num_classes=100)

# Load fine-tuned weights
checkpoint = torch.load('best_model.pth', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

Assignment Info

This model was trained as part of DLops Assignment 5.