Upload README.md with huggingface_hub

README.md

@@ -3,8 +3,11 @@ license: apache-2.0
 tags:
 - object-detection
 - medical-imaging
+- rf-detr
 - heart-anatomy
 - computer-vision
+datasets:
+- roboflow/heart-anatomy-types
 metrics:
 - mean-average-precision
 model-index:
@@ -21,7 +24,7 @@ model-index:
       name: mAP@50
 ---
 
-# Heartformer: Heart Anatomy Type Detection
+# Heartformer: Heart Anatomy Type Detection with RF-DETR
 
 **Heartformer** is a specialized object detection model for identifying and localizing different types of heart anatomy visualizations in medical images. Built on the RF-DETR (Roboflow Detection Transformer) architecture, this model can detect and classify seven distinct categories of cardiac imaging and illustration modalities.
 
@@ -93,9 +96,9 @@ Detection Heads
 
 ## 📊 Dataset
 
-### Heart Anatomy Types v2 (Self-Sourced)
+### Heart Anatomy Types v2 (Roboflow)
 
-The model was trained on a curated dataset of 621 annotated images from diffrent sources, specifically designed to capture the diversity of cardiac anatomy representations.
+The model was trained on a curated dataset of 621 annotated images from [Roboflow Universe](https://universe.roboflow.com/), specifically designed to capture the diversity of cardiac anatomy representations.
 
 #### Dataset Statistics
 
@@ -121,6 +124,7 @@ The model was trained on a curated dataset of 621 annotated images from diffrent
 #### Data Sources
 
 - Medical textbooks (openly licensed)
+- Roboflow Universe community contributions
 - Educational anatomy databases
 - All images verified for appropriate licensing
 
@@ -176,7 +180,7 @@ Annotations follow the COCO format:
 
 ### Training Details
 
-- **Hardware**: Apple M3 MacBook (MPS backend)
+- **Hardware**: Apple M3 MacBook Pro (MPS backend)
 - **Training Time**: ~1 hour 50 minutes
 - **Best Epoch**: Epoch 4 (with EMA weights)
 - **Early Stopping**: Triggered at epoch 11 (no improvement for 8 epochs)
@@ -250,19 +254,33 @@ The model shows excellent class separation with minimal confusion:
 
 ```bash
 pip install torch torchvision
+pip install git+https://github.com/roboflow/rf-detr.git
+pip install safetensors  # For loading .safetensors format
+```
+
+### Download Model
+
+**Recommended: SafeTensors format (safer, smaller, faster)**
+```bash
+wget https://huggingface.co/giannisan/heartformer/resolve/main/heartformer-v0.1.safetensors
+```
+
+**Alternative: PyTorch format**
+```bash
+wget https://huggingface.co/giannisan/heartformer/resolve/main/checkpoint_best_ema.pth
 ```
 
 ### Inference
 
+**Using SafeTensors (Recommended)**
 ```python
 from rfdetr import RFDETRNano
-from PIL import Image
+from safetensors.torch import load_file
 
 # Load model
-model = RFDETRNano(
-    pretrain_weights="path/to/checkpoint_best_ema.pth",
-    num_classes=8  # 7 classes + 1 background
-)
+model = RFDETRNano(num_classes=8)
+state_dict = load_file("heartformer-v0.1.safetensors")
+model.load_state_dict(state_dict)
 
 # Run inference
 detections = model.predict("heart_image.jpg", threshold=0.3)
@@ -278,6 +296,7 @@ for bbox, confidence, class_id in zip(
     print(f"BBox: {bbox}")
 ```
 
+**Using PyTorch Checkpoint
 ### Class Names
 
 ```python
@@ -354,7 +373,7 @@ If you use Heartformer in your research or application, please cite:
 
 ### Acknowledgments
 
-- **RF-DETR**: Based on RF-DETR architecture
+- **RF-DETR**: Based on Roboflow's RF-DETR architecture
   ```bibtex
   @misc{rfdetr2024,
     title={RF-DETR: Real-time Detection Transformer},
@@ -364,12 +383,12 @@ If you use Heartformer in your research or application, please cite:
     howpublished={\url{https://github.com/roboflow/rf-detr}}
   }
   ```
-- **Dataset**: Heart Anatomy Types v2
+- **Dataset**: Heart Anatomy Types v2 from Roboflow Universe
 - **DINOv2 Backbone**: Meta AI's self-supervised vision transformer
 
 ## 📄 License
 
-This model is released under the **Apache License 2.0**,
+This model is released under the **Apache License 2.0**, the same license as RF-DETR.
 
 ```
 Copyright 2024 Giannisan
@@ -408,4 +427,4 @@ For questions, issues, or collaboration opportunities:
 
 ---
 
-**Note**: This model is continuously being improved. Check back for updates and new versions!
+**Note**: This model is continuously being improved. Check back for updates and new versions!