Limits of Keras EarlyStopping Verbosity on Console Noise and Warning Visibility

Goal: silence the per‑epoch messages produced by Keras callbacks during model.fit while preserving the ability to be notified of serious training anomalies such as NaN metric values or checkpoint write failures.

Constraint: setting verbose=0 on callbacks like EarlyStopping, ModelCheckpoint, or CSVLogger removes all stdout output from those callbacks, but it is unclear whether internally generated warnings (e.g., when a monitored metric becomes NaN) are also suppressed. TensorBoard provides non‑console logging, yet its default configuration does not surface those warnings in the UI.

Uncertainty: does the EarlyStopping implementation still emit warning prints when verbose=0, or are they gated by the same flag? If they are hidden, what documented mechanism allows users to receive those alerts without reintroducing epoch‑level noise?

Specific questions:

• Does EarlyStopping(verbose=0) suppress warning messages about NaN metrics or failed checkpoints?
• Is there a supported callback argument or configuration that forwards such warnings to a logger or TensorBoard while keeping verbose=0 for epoch chatter?
• Can TensorBoard be tuned to display callback‑generated warnings in its scalars or logs without causing console output?