inference.py

import torch
from torchvision import transforms
from model import MNISTModel


class InferenceWrapper:
    def __init__(self, model_path: str):
        """
        Initialize the inference wrapper with a model path.
        
        Args:
            model_path (str): Path to the model weights file
        """
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.model_path = model_path
        self.model = self._load_model()
        self.transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.1307,), (0.3081,))
        ])
        
    def _load_model(self):
        """Load and return the model."""
        model = MNISTModel().to(self.device)
        model.load_state_dict(
            torch.load(self.model_path, map_location=self.device, weights_only=True)
        )
        model.eval()
        return model

    def predict_tensor(self, input_tensor: torch.Tensor):
        """
        Run inference on a single input tensor.
        
        Args:
            input_tensor (torch.Tensor): Input tensor of shape [1, 28, 28] or [N, 1, 28, 28]
            
        Returns:
            tuple: (prediction, confidence)
        """
        with torch.no_grad():
            if input_tensor.dim() == 3:
                input_tensor = input_tensor.unsqueeze(0)
            
            input_tensor = input_tensor.to(self.device)
            output = self.model(input_tensor)
            probs = torch.softmax(output, dim=1)
            prediction = output.argmax(1).item()
            confidence = probs[0][prediction].item()
            return prediction, confidence

    def predict_batch(self, input_tensors: torch.Tensor):
        """
        Run inference on a batch of input tensors.
        
        Args:
            input_tensors (torch.Tensor): Batch of input tensors of shape [N, 1, 28, 28]
            
        Returns:
            tuple: (predictions, confidences)
        """
        with torch.no_grad():
            input_tensors = input_tensors.to(self.device)
            output = self.model(input_tensors)
            probs = torch.softmax(output, dim=1)
            predictions = output.argmax(1)
            confidences = torch.gather(probs, 1, predictions.unsqueeze(1)).squeeze(1)
            return predictions.cpu().numpy(), confidences.cpu().numpy()


def main():
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument('--model-path', required=True, help='Path to the model weights')
    args = parser.parse_args()
    
    # Example usage
    wrapper = InferenceWrapper(args.model_path)
    
    # Example single inference
    test_input = torch.randn(1, 28, 28)
    prediction, confidence = wrapper.predict_tensor(test_input)
    print(f"Single prediction: {prediction}, confidence: {confidence:.4f}")
    
    # Example batch inference
    batch_input = torch.randn(4, 1, 28, 28)
    predictions, confidences = wrapper.predict_batch(batch_input)
    print(f"Batch predictions: {predictions}")
    print(f"Batch confidences: {confidences}")

if __name__ == "__main__":
    main()