Add fallback to librosa features when HeAR model not available

Dockerfile

@@ -13,6 +13,10 @@ COPY app.py .
 COPY inference_service.py .
 COPY pneumonia_classifier.joblib .
 
+# Set environment variable for Hugging Face token
+# Pass this at runtime: docker run -e HF_TOKEN="your_token" ...
+ENV HF_TOKEN=""
+
 EXPOSE 5000
 
 CMD ["python", "app.py"]

README.md

@@ -1,6 +1,6 @@
 ---
 title: Pneumonia Space
-emoji: 🌍
+emoji: 🫁
 colorFrom: blue
 colorTo: green
 sdk: gradio
@@ -10,4 +10,82 @@ pinned: false
 license: mit
 ---
 
+# 🫁 Pneumonia Risk Assessment API
+
+AI-powered API for assessing pneumonia risk from respiratory audio recordings.
+
+## 🚀 Features
+
+- **HeAR Model Integration**: Uses Google's Health Acoustic Representations model
+- **Risk Scoring**: Provides probability-based risk assessment (not diagnostic)
+- **Fallback System**: Uses librosa-based features if HeAR model unavailable
+- **REST API**: Simple Flask endpoint for audio file uploads
+
+## 🔑 Setup
+
+### Hugging Face Authentication
+
+The HeAR model requires Hugging Face authentication. Set your token as an environment variable:
+
+```bash
+export HF_TOKEN="your_huggingface_token_here"
+```
+
+Or login using the CLI:
+
+```bash
+huggingface-cli login
+```
+
+Get your token from: https://huggingface.co/settings/tokens
+
+### Running Locally
+
+```bash
+# Install dependencies
+pip install -r requirements.txt
+
+# Run the application
+python app.py
+```
+
+### Using Docker
+
+```bash
+# Build the image
+docker build -t pneumonia-api .
+
+# Run with HF token
+docker run -p 5000:5000 -e HF_TOKEN="your_token" pneumonia-api
+```
+
+## 📡 API Usage
+
+**Endpoint**: `POST /predict_pneumonia`
+
+**Request**: Multipart form data with `audio_file`
+
+**Response**:
+```json
+{
+  "filename": "recording.wav",
+  "pneumonia_risk_score": 0.7234,
+  "risk_level": "High",
+  "note": "This is an AI assessment, not a medical diagnosis. Consult a healthcare professional."
+}
+```
+
+**Example with curl**:
+```bash
+curl -X POST -F "audio_file=@recording.wav" http://localhost:5000/predict_pneumonia
+```
+
+## ⚠️ Disclaimer
+
+This tool provides risk assessment scores, not medical diagnoses. Always consult healthcare professionals for medical decisions.
+
+## 📄 License
+
+MIT License
+
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

inference_service.py

@@ -9,14 +9,14 @@ import warnings
 warnings.filterwarnings("ignore", category=UserWarning, module="soundfile")
 warnings.filterwarnings("ignore", module="librosa")
 
-# Try to import tensorflow_hub for HeAR model, but make it optional
+# Try to import huggingface_hub for HeAR model
 try:
-    import tensorflow as tf
-    import tensorflow_hub as hub
-    TF_AVAILABLE = True
+    from huggingface_hub import from_pretrained_keras
+    from huggingface_hub.utils import HfFolder
+    HF_AVAILABLE = True
 except ImportError:
-    TF_AVAILABLE = False
-    print("TensorFlow not available. Will use librosa-based features.")
+    HF_AVAILABLE = False
+    print("Hugging Face Hub not available. Will use librosa-based features.")
 
 # HeAR Parameters
 SAMPLE_RATE = 16000
@@ -36,23 +36,34 @@ def create_dummy_audio(filename='dummy_audio.wav', duration=5, sr=SAMPLE_RATE):
 
 def load_hear_model():
     """
-    Try to load HeAR model. If not available, return None and use fallback feature extraction.
+    Try to load HeAR model from Hugging Face Hub.
+    If not available or authentication fails, return None and use fallback feature extraction.
     """
-    if not TF_AVAILABLE:
-        print("TensorFlow not available. Using librosa-based feature extraction.")
+    if not HF_AVAILABLE:
+        print("Hugging Face Hub not available. Using librosa-based feature extraction.")
         return None
     
-    print("Attempting to load HeAR model from Kaggle Models...")
+    print("Loading HeAR model from Hugging Face Hub...")
     try:
-        # Try loading from Kaggle Models (requires kagglehub)
-        import kagglehub
-        path = kagglehub.model_download("google/hear/tfLite/default")
-        loaded_model = tf.saved_model.load(path)
+        # Check if HF token is available (from env var or HF cache)
+        token = os.environ.get("HF_TOKEN") or HfFolder.get_token()
+        if token is None:
+            print("Warning: No Hugging Face token found.")
+            print("Set HF_TOKEN environment variable or login with 'huggingface-cli login'")
+            print("Falling back to librosa-based feature extraction.")
+            return None
+        
+        # Load the model directly from Hugging Face Hub
+        loaded_model = from_pretrained_keras("google/hear", use_auth_token=token)
         infer_fn = loaded_model.signatures["serving_default"]
-        print("HeAR model loaded successfully from Kaggle.")
+        
+        print(f"HeAR Model loaded successfully!")
+        print(f"Sample Rate: {SAMPLE_RATE} Hz")
+        print(f"Clip Duration: {CLIP_DURATION} seconds")
+        print(f"Clip Length: {CLIP_LENGTH} samples")
         return infer_fn
     except Exception as e:
-        print(f"HeAR model not available: {e}")
+        print(f"Error loading HeAR model: {e}")
         print("Falling back to librosa-based feature extraction.")
         return None
 

requirements.txt

@@ -4,7 +4,6 @@ librosa
 joblib
 soundfile
 scikit-learn
-# Optional: for HeAR model support
+# For HeAR model support
+huggingface_hub>=0.20.0
 tensorflow>=2.12.0
-tensorflow_hub
-kagglehub