Egg Inspector

This repository implements egg object detection using Faster R-CNN with a ResNet-50 backbone.

Configuration

• Number of GPUs: 4
• Architecture:
  • Type: Faster R-CNN
  • Backbone: ResNet-50
• Number of Classes: 6

Data Loader

• Type: EggDataLoader
• Arguments:
  • Data Directory: "data/"
  • Batch Size: 1024
  • Shuffle: True
  • Validation Split: 0.1
  • Number of Workers: 4

Optimizer

• Type: SGD
• Arguments:
  • Learning Rate: 0.005
  • Momentum: 0.9
  • Weight Decay: 0.005

Loss Function

• Cross-Entropy Loss

Evaluation Metrics

• Precision, Recall, F1 Score

Learning Rate Scheduler

• Type: StepLR
• Arguments:
  • Step Size: 50
  • Gamma: 0.1

Data Transformations

Training

• Random Rotation: [-45, 45] degrees
• Color Adjustments: Brightness=0.5, Contrast=0.5, Saturation=0.5, Hue=0.5

Testing

• Resize: Size=256
• Center Crop: Size=224

Trainer

• Epochs: 10
• Save Directory: "saved/"
• Save Frequency: Save model every epoch
• Verbosity: Level 2
• Monitoring: Minimize Validation Loss
• Early Stopping: Wait 10 epochs before stopping
• TensorBoard: Enabled

Miscellaneous

• Data Path: AI-hub Egg Dataset
• Batch Size: 4
• Collate Function: collate_fn_custom
• Number of Classes: 6
• Class Weights: [1.0, 1.5, 1.0, 1.0, 1.0, 1.0]

---

This project aims to effectively detect various egg objects using Faster R-CNN with a ResNet-50 backbone. The provided configuration includes essential elements for training and evaluating the model on a specific dataset.