prenorm-resnet-cifar10-v4

PreNormResNet v4 – CIFAR-10 (From Scratch)

📌 Overview

PreNormResNet_v4 is a custom ResNet-style architecture trained from scratch on CIFAR-10 using modern stabilization and regularization techniques.

Unlike fine-tuned ImageNet models, this network learns all representations directly from CIFAR-10, making it suitable for architectural research and controlled experimentation.

---

🧠 Architecture Highlights

• Pre-Normalization Residual Bottlenecks
• Group Normalization (batch-size independent)
• Early spatial downsampling
• Custom stage design (1–2–2 blocks)
• Label smoothing
• Cosine LR scheduling with warmup
• Early stopping

---

🏗 Architecture Summary

📊 Performance

Note: This model is trained from scratch, not fine-tuned.

---

🔬 Training Details

• Dataset: CIFAR-10
• Input resolution: 32×32
• Normalization: CIFAR statistics
• Loss: CrossEntropy with label smoothing (0.1)
• Scheduler: Linear warmup → Cosine Annealing
• Batch size: 128
• Hardware: NVIDIA GTX 1650

---

🚀 Usage

Load model

import torch
from model import ResNet50Style_v4, PreNormBottleneck_v4

ckpt = torch.load("prenorm_resnet_cifar_v4.pt", map_location="cpu")

model = ResNet50Style_v4(
    block=PreNormBottleneck_v4,
    n_blocks=[1, 2, 2],
    num_classes=10
)
model.load_state_dict(ckpt["state_dict"])
model.eval()

🧪 Intended Use

• Research on residual architectures
• Studying normalization strategies
• Training dynamics from scratch
• CIFAR-scale experimentation

⚠ Limitations

• Not pretrained
• Lower accuracy than ImageNet fine-tuned models
• CIFAR-specific resolution

👤 Author

AnjanSB

Built with PyTorch, MLflow, DVC, and Dagshub. Experiment Repo : https://dagshub.com/AnjanSB/cifar-image-classification-modeling-dlops