tialdrine/resnet18-intel-classifier

🏞️ ResNet18 Intel Image Classifier

📌 A ResNet18-based image classification model trained on the Intel Image Classification dataset, capable of recognizing six types of natural scenes. The model was fine-tuned using PyTorch, optimized for reproducibility and deployment in educational and practical scenarios.

🏷️ Classes

• Buildings
• Forest
• Glacier
• Mountain
• Sea
• Street

🧰 Training Procedure

1. Loaded a pretrained ResNet18 model from torchvision.models.
2. Replaced the final classification layer with a 6-unit fully connected layer.
3. Resized all input images to 224x224 and applied ImageNet normalization.
4. Used ImageFolder and random_split() to divide the dataset:
   • 70% Training
   • 15% Validation
   • 15% Testing
5. Training setup:
   • Optimizer: Adam
   • Loss Function: CrossEntropyLoss
   • Batch size: 32
   • Learning rate: 0.001
   • Epochs: 5
6. Saved the final model as pytorch_model.bin.

📊 Performance

Metric	Value
Final Train Accuracy	90.08%
Final Val Accuracy	88.74%

⚙️ Framework & Environment

• Python: 3.10.12
• PyTorch: 2.0.1+cu118
• Torchvision: 0.15.2+cu118
• Platform: Google Colab (GPU enabled, CUDA support)

🧪 Hyperparameters

Parameter	Value
Epochs	5
Batch Size	32
Optimizer	Adam
Learning Rate	0.001
Loss Function	CrossEntropy
Image Size	224x224
Data Split	70% Train / 15% Val / 15% Test

---