AnjanSB/finetune-resnet-cifar10

ResNet50 Fine-Tuned – CIFAR-10

📌 Overview

ResNet50 Fine-Tuned for CIFAR-10 is a high-performance image classification model built by adapting an ImageNet-pretrained ResNet50 to the CIFAR-10 dataset.

Instead of training from scratch, this model leverages pretrained visual representations and selectively fine-tunes higher layers, achieving strong accuracy with faster convergence and improved generalization.

This model represents a production-oriented baseline for small-image classification tasks.

---

🏗 Architecture Summary

Input (224×224)
└─ Pretrained ResNet50 Backbone
├─ Conv → BN → ReLU
├─ Residual Blocks (Frozen)
├─ Residual Blocks (Fine-tuned – layer4)
└─ Global Average Pool
└─ Custom FC Head
├─ Linear (2048 → 512)
├─ ReLU
├─ Dropout (0.5)
└─ Linear (512 → 10)

---

📊 Performance

Metric	Value
Validation Accuracy	92.46%
Validation Loss	0.257
Training Epochs	40
Optimizer	AdamW
Learning Rate	1e-4
Weight Decay	1e-4

This model outperforms all scratch-trained models by leveraging pretrained ImageNet features.

---

🔬 Training Details

• Dataset: CIFAR-10
• Input resolution: 224×224
• Normalization: ImageNet statistics
• Loss: CrossEntropyLoss
• Scheduler: Cosine Annealing
• Batch size: 64
• Hardware: NVIDIA GTX 1650

---

🚀 Usage

Load model

import torch
from model import ResNet50FineTune

model = ResNet50FineTune(num_classes=10)

ckpt = torch.load("resnet50_finetune_cifar.pt", map_location="cpu")
model.load_state_dict(ckpt["state_dict"])

model.eval()

Input preprocessing

from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=(0.485, 0.456, 0.406),
        std=(0.229, 0.224, 0.225)
    )
])

---

🧪 Intended Use

• Production-ready CIFAR-10 classification
• Transfer learning benchmarks
• Fine-tuning experiments
• Educational demonstrations of pretrained CNN adaptation

---

⚠ Limitations

• Input images must be resized to 224×224
• Depends on ImageNet-pretrained weights
• Less suitable for architecture research compared to scratch models

---

🔍 Comparison with Scratch Model

Model	Training	Val Acc
PreNormResNet v4	From scratch	89.29%
ResNet50 Fine-Tuned	Transfer learning	92.46%

Pretraining shifts the learning regime and provides a strong inductive bias, especially effective for small datasets like CIFAR-10.

---

👤 Author

AnjanSB

Built with PyTorch, MLflow, DVC, and Dagshub

📊 Experiment Repo: https://dagshub.com/AnjanSB/cifar-image-classification-modeling-dlops

🚀 Live Demo (Custom Scratch Model): https://huggingface.co/spaces/AnjanSB/cifar10-prenorm-resnet-demo

---

🏁 Final Note

This model represents a real-world deployment strategy, while the custom PreNormResNet v4 highlights architecture engineering and research depth. Together, they form a complete and well-rounded ML portfolio.