tf.summary.trace_on

Starts a trace to record computation graphs and profiling information.

tf.summary.trace_on(
    graph=True, profiler=False, profiler_outdir=None
)

Used in the notebooks

Used in the guide Used in the tutorials

Must be invoked in eager mode.

When enabled, TensorFlow runtime will collect information that can later be exported and consumed by TensorBoard. The trace is activated across the entire TensorFlow runtime and affects all threads of execution.

To stop the trace and export the collected information, use tf.summary.trace_export. To stop the trace without exporting, use tf.summary.trace_off.

graph If True, enables collection of executed graphs. It includes ones from tf.function invocation and ones from the legacy graph mode. The default is True.
profiler If True, enables the advanced profiler. Enabling profiler implicitly enables the graph collection. The profiler may incur a high memory overhead. The default is False.
profiler_outdir Output directory for profiler. It is required when profiler is enabled when trace was started. Otherwise, it is ignored.