Instrumentation adds observations or logging so a model’s computations become inspectable during execution. It turns loss, accuracy, probabilities, and internal states into evidence for evaluation and explanation. section 5.2.5 section 8.2.3
Practical instrumentation can:
- Measure intuitive performance alongside loss, such as binary accuracy after each training epoch. section 5.2.5
- Convert tensors into inspectable NumPy arrays, detaching them before recording values outside their device. section 8.2.3
- Expose prediction probabilities, hidden-state vectors, and intermediate outputs through model convenience methods. section 8.2.3
- Print selected values during execution when debugging or investigating how predictions arise. section 8.2.3
Tracing an RNN prediction
For “Khalid,” call predict_hidden and predict_category on each growing prefix. The resulting categories change from English to Chinese, Algerian, and Arabic, while the recorded 128-dimensional hidden states can be tabulated and compared with character position.
