vit-small-cifar100-lora

ViT-Small CIFAR-100 (LoRA Fine-tuned)

This model is a vit_small_patch16_224 from timm, fine-tuned on CIFAR-100 using LoRA (Low-Rank Adaptation) via the PEFT library.

Training Details

• Base model: vit_small_patch16_224 (ImageNet pretrained)
• Dataset: CIFAR-100 (100 classes)
• Method: LoRA injected into attention qkv layers
• WandB project: mlops-assignment5

Usage

import torch
import timm
from peft import LoraConfig, get_peft_model

model = timm.create_model("vit_small_patch16_224", pretrained=False, num_classes=100)
lora_config = LoraConfig(r=RANK, lora_alpha=ALPHA, target_modules=["qkv"],
                         lora_dropout=0.1, bias="none", modules_to_save=["head"])
model = get_peft_model(model, lora_config)

ckpt = torch.load("pytorch_model.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()