TensorBoard

Any training loop already calling writer.add_scalar(tag, value, step) can get read-only OmniLoop dashboard readouts for free. There is no need to edit your loop or define a separate OmniLoop schema.

Usage

Wrap your existing SummaryWriter once:

from torch.utils.tensorboard import SummaryWriter
from omniloop.integrations.tensorboard import OmniLoopSummaryWriter

# Was: writer = SummaryWriter(log_dir)
writer = OmniLoopSummaryWriter(SummaryWriter(log_dir))

# ... inside your loop ...
writer.add_scalar("Loss/Policy", loss, it)  # -> Sent to TensorBoard AND OmniLoop

Every scalar is forwarded unchanged to the wrapped TensorBoard writer. It is also published as an OmniLoop telemetry frame. The tag is automatically declared as a read-only parameter (e.g., "Loss/Policy" becomes group “Loss”, label “Policy”) so it appears in the dashboard’s readout section automatically.

Installation

Install the TensorBoard extra:

pip install omniloop[tensorboard]

Alternative: In-place patching

If you don’t control where the writer is instantiated (e.g., inside a framework’s runner), you can patch an existing SummaryWriter in place:

from omniloop.integrations.tensorboard import mirror_summary_writer

mirror_summary_writer(runner.writer)

Coalesce Mode

By default, OmniLoopSummaryWriter operates with Coalesce mode enabled.

Scalars are buffered and flushed as a single telemetry frame per global_step. This ensures that a dashboard tick shows a consistent row of metrics that all correspond to the exact same training step.