You can deploy this model and start earning money today!

maqsudxo1ja/
cifar100-cnn-resnet18

Views

No views yet

CIFAR-100 Image Classification — CNN & ResNet18

This repository contains two image classification models trained on the CIFAR-100 dataset: a Custom CNN built from scratch, and a ResNet18 model adapted via transfer learning.

📊 Model Results

Model	Top-1 Accuracy	Top-5 Accuracy	Parameters	File Size
Custom CNN	55.21%	84.12%	~2.1M	5.57 MB
ResNet18 (Transfer Learning)	71.13%	—	~11.2M	45 MB

Independently verified: on the full CIFAR-100 test set (10,000 images), the ResNet18 model achieved 70.53% accuracy.

🗂️ About the Dataset

CIFAR-100 consists of 60,000 color images (32x32 pixels) across 100 classes: 50,000 training images and 10,000 test images, with 600 images per class.

🏗️ Model Architecture

Custom CNN

Input (3x32x32)
  → [Conv2D(32) + BatchNorm + ReLU] x2 → MaxPool → Dropout(0.25)
  → [Conv2D(64) + BatchNorm + ReLU] x2 → MaxPool → Dropout(0.3)
  → [Conv2D(128) + BatchNorm + ReLU] x2 → MaxPool → Dropout(0.4)
  → Flatten → Dense(512) → Dropout(0.5) → Dense(100)

ResNet18 (Adapted)

Based on a ResNet18 pretrained on ImageNet, adapted for small (32x32) images as follows:

First convolutional layer changed from 7x7, stride=2 to 3x3, stride=1
Initial MaxPool layer removed (replaced with nn.Identity())
Final fully connected layer resized for 100 output classes

🚀 Usage (Inference)

python

import torch
import torch.nn as nn
from torchvision.models import resnet18
from huggingface_hub import hf_hub_download

# Download the model from Hugging Face
resnet_path = hf_hub_download(
    repo_id="maqsudxo1ja/cifar100-cnn-resnet18",
    filename="resnet18_cifar100.pth"
)

# Rebuild the architecture
model = resnet18(weights=None)
model.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
model.maxpool = nn.Identity()
model.fc = nn.Linear(model.fc.in_features, 100)

# Load the weights
model.load_state_dict(torch.load(resnet_path, map_location="cpu"))
model.eval()

Preprocessing

python

import torchvision.transforms as transforms

transform = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
    transforms.Normalize((0.5071, 0.4865, 0.4409), (0.2673, 0.2564, 0.2762))
])

Full Inference Example

python

from PIL import Image
import torch

image = Image.open("image.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    output = model(input_tensor)
    probs = torch.softmax(output, dim=1)[0]
    top5_prob, top5_idx = torch.topk(probs, 5)

# class names can be obtained via torchvision.datasets.CIFAR100(...).classes
for i in range(5):
    print(f"{classes[top5_idx[i]]}: {top5_prob[i]*100:.2f}%")

🎯 Training Details

Parameter	Value
Optimizer	Adam (weight_decay=1e-4)
Learning rate	0.001 (with ReduceLROnPlateau)
Loss function	CrossEntropyLoss
Batch size	128
Data augmentation	RandomCrop, RandomHorizontalFlip
Early stopping	Patience = 5 epochs
Environment	Kaggle, GPU T4 x2

📁 Files in this Repository

cnn_cifar100.pth — weights of the Custom CNN trained from scratch
resnet18_cifar100.pth — weights of the fine-tuned ResNet18 (transfer learning)

⚠️ Limitations

The model only recognizes CIFAR-100's 100 classes (not general real-world objects)
Optimized for small (32x32) images — higher-resolution images must be resized to 32x32 first
Higher confusion is likely between visually similar classes (e.g., different animal species)

📄 License

MIT License

✍️ Author

@maqsudxo1ja