Access Logits of a classification model

Once the classification model is trained, is there a way to access the logits for a given input image? This would be require to create an anomaly detector.