rujutashashikanjoshi/rfdetr-drone-detection-0205-mixed

RF-DETR MEDIUM Model

This is an RF-DETR (Real-time Transformer-based Object Detector) model fine-tuned on a custom dataset.

Model Details

• Model Architecture: RFDETRMedium
• Framework: RF-DETR (PyTorch)
• Image Resolution: 640x640
• Training Epochs: 50
• Batch Size: 4 (with gradient accumulation steps: 4)

Classes

The model is trained to detect the following classes: Classes not available

Usage

Using RF-DETR Library (Recommended)

from rfdetr import RFDETRMedium
from PIL import Image

# Load the model
model = RFDETRMedium(
    pretrain_weights="hf://rujutashashikanjoshi/rfdetr-drone-detection-0205/checkpoint_best_total.pth"
)
model.optimize_for_inference()

# Run inference
image = Image.open("your_image.jpg")
detections = model.predict(image, threshold=0.5)

print(f"Found {len(detections)} detections")

Using PyTorch Directly

import torch
from huggingface_hub import hf_hub_download

# Download the model weights
model_path = hf_hub_download(
    repo_id="rujutashashikanjoshi/rfdetr-drone-detection-0205",
    filename="pytorch_model.bin"
)

# Load the state dict
state_dict = torch.load(model_path, map_location='cpu')

Training Details

• Optimizer: AdamW
• Learning Rate Schedule: Cosine with warmup
• Data Format: COCO JSON
• Framework Version: rfdetr>=1.2.1

Files in this Repository

• checkpoint_best_total.pth: Original RF-DETR checkpoint (best model)
• pytorch_model.bin: Standard PyTorch weights for compatibility
• config.json: Model configuration
• class_names.txt: List of detection classes
• results.json: Training results and metrics
• metrics_plot.png: Visualization of training metrics
• log.txt: Full training log

License

This model is released under the Apache 2.0 License.