Limits of Keras EarlyStopping Verbosity on Console Noise and Warning Visibility

Answer

EarlyStopping(verbose=0) only hides the callback’s own stop message (e.g., "Early stopping epoch 42: val_loss did not improve"). It does not suppress warnings such as NaN metric detection or checkpoint write failures; those are emitted via Python’s warning system or TensorFlow logging and remain visible regardless of the verbose setting.

Confirmed facts

• The verbose argument in EarlyStopping has two effective states: 0 (silent) and ≥1 (prints the single stop line). Any value greater than 1 is treated as 1, and negative values as 0.
• Warnings generated by Keras/TensorFlow (e.g., UserWarning for NaN metrics, DeprecationWarning, or TF I/W/E logs) are written to stderr or the Python warnings channel independently of the callback’s verbose flag.
• TensorBoard’s default scalars view does not display these warnings; they appear only in the console or in logs unless explicitly forwarded.

Steps to verify the behavior

1. Create a simple model and attach EarlyStopping(monitor='val_loss', verbose=0, patience=2).
2. Run model.fit(..., verbose=2) so that epoch progress bars are shown.
3. Observe the console: no "Early stopping" line appears, but if a monitored metric becomes NaN you will still see a warning such as UserWarning: Metric val_loss is NaN.
4. Redirect stderr to a file and repeat; the warning will still be present, confirming it is not gated by EarlyStopping.verbose.

Getting warnings without epoch-level noise

There is no built-in argument to forward EarlyStopping-generated warnings to a logger while keeping verbose=0. To receive those alerts silently:

• Wrap the training call in a custom callback that catches warnings via warnings.catch_warnings(record=True) and logs them to a file or TensorBoard using tf.summary.text.
• Alternatively, route warnings into the logging module with logging.captureWarnings(True) and configure a logger that writes to a file or TensorBoard.
• For checkpoint failures, keep ModelCheckpoint(..., verbose=0) and inspect the checkpoint directory or catch raised exceptions; errors still surface via stderr or exceptions.

TensorBoard considerations

TensorBoard does not automatically surface callback warnings. To see them in TensorBoard:

• Emit a custom scalar or text summary from a callback when a warning is caught (e.g., tf.summary.scalar('warning/nan_metric', 1, step=epoch)).
• Ensure the summary writer is flushed; the UI will then show the warning as a scalar or text entry without any console output.

To confirm your TensorFlow/Keras version behaves as described, run print(tf.__version__) and compare against the EarlyStopping API documentation for that release.