🎯 ResNet50 ImageNet Classifier
Trained from Scratch on ImageNet-1K | 75%+ Top-1 Accuracy
1000 classes • 25.6M parameters • 98MB model
📊 Dataset
1.28M training images
1000 ImageNet classes
🎯 Performance
75-77% top-1 accuracy
92-94% top-5 accuracy
⚡ Architecture
ResNet50 (Bottleneck)
25.6M parameters
📸 Upload an Image for Classification
📌 Try these example images
1 10
💡 Tips for Best Results
- Upload clear, well-lit images
- Works best with centered objects
- Supports 1000 ImageNet categories
- Processing time: ~1-2 seconds
Model Architecture
ResNet50 trained from scratch (no pre-trained weights) on ImageNet-1K
Training Configuration:
- Optimizer: SGD with momentum (0.9), weight decay (1e-4)
- Learning Rate: Cosine annealing with warmup (0.1 → 0.0005)
- Augmentation: AutoAugment (ImageNet), RandomErasing, Mixup
- Precision: Mixed FP16 with gradient scaling
- Epochs: 75 with early stopping
Architecture Details:
Input (224×224×3)
↓
Conv1 (7×7, stride=2) + BN + ReLU → 112×112×64
MaxPool (3×3, stride=2) → 56×56×64
↓
Layer1: 3× Bottleneck → 56×56×256
Layer2: 4× Bottleneck → 28×28×512
Layer3: 6× Bottleneck → 14×14×1024
Layer4: 3× Bottleneck → 7×7×2048
↓
Global Average Pool → 1×1×2048
Fully Connected → 1000 classes
Project Links
Citation
@inproceedings{he2016deep,
title={Deep residual learning for image recognition},
author={He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
booktitle={CVPR},
year={2016}
}
💜 Built with Gradio • Trained on AWS EC2 • Deployed on 🤗 Hugging Face Spaces
Model trained from scratch achieving 76.12% top-1 accuracy on ImageNet-1K