Upload inference.py with huggingface_hub

inference.py

@@ -37,6 +37,7 @@ def load_model_and_artifacts(
     feature_names_path="feature_names.json",
     config_path="model_config.json"
 ):
+    """Load model, scaler, feature names, and configuration."""
     if not all(os.path.exists(path) for path in [model_path, scaler_path, feature_names_path, config_path]):
         missing = [path for path in [model_path, scaler_path, feature_names_path, config_path] if not os.path.exists(path)]
         raise FileNotFoundError(f"Missing files: {missing}")
@@ -63,6 +64,7 @@ def load_model_and_artifacts(
     return model, scaler, feature_names, config
 
 def predict(model, scaler, sequences):
+    """Generate 13-week sales forecasts from input sequences."""
     device = torch.device("cpu")
     model.to(device)
     
@@ -86,9 +88,9 @@ def predict(model, scaler, sequences):
     rescaled = np.maximum(rescaled, 0)
     rescaled = np.clip(rescaled, 3000, 19372)  # Training sales range: $3069–19372
     
-    # Estimate uncertainty: std of predictions per sequence, repeated for each timestep
-    uncertainties = np.std(rescaled, axis=1, keepdims=True) + 100  # Shape: (batch_size, 1)
+    # Estimate uncertainty: scaled std of predictions per sequence, repeated for each timestep
+    uncertainties = 0.2 * np.std(rescaled, axis=1, keepdims=True) + 100  # Shape: (batch_size, 1)
     uncertainties = np.repeat(uncertainties, rescaled.shape[1], axis=1)  # Shape: (batch_size, 13)
-    uncertainties = np.clip(uncertainties, 100, 500)  # Ensure reasonable bounds
+    uncertainties = np.clip(uncertainties, 100, 1000)  # Wider bounds to avoid constant clipping
     
-    return rescaled, uncertainties
+    return rescaled, uncertainties