Open source · ML infrastructure · Cross-framework adoption

Standalone TensorBoard

In 2016 the best way to watch a model train lived inside one framework. A team on MXNet or PyTorch could write the log file perfectly well and then had no way to open it — the renderer existed only inside a TensorFlow installation. I rebuilt it as a standalone tool that did both halves, logging and rendering, in one package. Amazon built its official MXNet tooling on that work, the PyTorch ecosystem referenced it, and TensorBoard itself now installs without TensorFlow.

A tool you could not open

In late 2016 I was training models on MXNet and doing what everyone else did: plotting metrics by hand with matplotlib, resizing images, mailing PNGs to colleagues. TensorBoard already solved all of that, and I could not use it.

I scoped the work in public, on the MXNet issue tracker, and the sticking point was there in the first pass:

That is the whole problem in one line. Asking a framework’s users to install a competing framework in order to see their own charts is not a dependency, it is a tax — and one paid by every non-TensorFlow community at once, which is what made it worth fixing properly rather than fixing for MXNet.

The widget below is that decision, made concrete. Pick a framework, then try to get a chart on screen.

What already existed, and what did not

The SOCKEYE paper describes it as “a standalone Tensorboard fork,” and in open-source usage that word describes provenance — a repository derived from another — rather than a judgement about originality. Still, it is worth being exact, because the halves are not equally hard. The format was Google’s and public, and logging adapters already existed. What the field lacked was a renderer that ran on its own — and, with it, a single tool that took you from a training loop to a chart without TensorFlow anywhere in the install.

What the field did with it

Two rival ecosystems ended up on it. AWS released mxboard in March 2018 as the official way to log MXNet data for TensorBoard, crediting this project as its origin. In the PyTorch world, tensorboardX — still one of the most widely used visualisation libraries outside TensorFlow — lists it as a reference. Amazon’s research toolkit SOCKEYE cited it in a peer-reviewed paper for rendering training statistics.

Two of these numbers are not mine to claim — they count other people’s packages. What they measure is whether the field still runs through this layer nearly a decade later.

The mxboard figure is the small number that says the most. Its last release was May 2018, and Apache MXNet — the framework it served — was retired to the Apache Attic in September 2023. Eight years without an update and two years after its framework was formally shut down, it is still installed around two thousand times a month by people who need to see inside a training run. Every figure above links to a live counter rather than a screenshot; counts and rankings move, and these were read in May and June 2026.

Then the incumbent moved. In mid-2017 Google split TensorBoard out of TensorFlow into a project of its own, and in a public thread on this repository the engineering director leading TensorFlow acknowledged the work and addressed the dependency directly:

They did remove it. TensorBoard’s own requirements file no longer lists TensorFlow, and the code keeps a stub that stands in when TensorFlow is absent — running without it is a maintained path, not an accident. Google’s README states it directly: TensorBoard “can be run with a reduced feature set if you do not have TensorFlow installed.” A few plugins and Cloud Storage log directories still want it. For the ordinary case of writing a log and reading it back, the dependency this project existed to work around is gone from the original too. The strongest evidence that a piece of infrastructure was right is usually not that it was adopted downstream but that the incumbent adopted its position, and that is what happened here.

Why a charting tool was a competitive problem

That reframes what this project was. Not developer convenience, but the removal of a switching cost that was steering enterprise customers toward one vendor. It is also why the work survived its author’s involvement: AWS needed MXNet to be a credible choice, the PyTorch community needed the same, and both got there through the same layer.

There is a durable lesson in it for the work I do now. Inspectability decides whether a model can be trusted in production, and inspectability that only exists inside one vendor’s stack is not inspectability for the field — it is a lock. The same instinct runs through the memory-network NLU work, where attention weights were exposed so operators could see why a request routed the way it did, and through the Indicator Loss patent, where the whole point is that a biasing decision is a logged value rather than a guess.

Timeline

• 27 Nov 2016Opened apache/mxnet#4003, scoping a TensorBoard-class tool for MXNet by taking the renderer out of TensorFlow, with a working proof of concept.
• 2017dmlc/tensorboard released — logging interface and renderer in one package, installable in one command with no TensorFlow present, and published on PyPI under the tensorboard name. Apache MXNet's documentation names me the repository's primary author.
• 3 Jul 2017The PyPI name is transferred to Google. By my own count in the thread that day, the repository was drawing several hundred visitors daily.
• 4 Jul 2017Google splits TensorBoard into a standalone project; its engineering director acknowledges this work publicly and commits, in writing, to removing the TensorFlow dependency.
• May 2018AWS releases mxboard as official MXNet tooling, crediting this project as its origin; Amazon’s SOCKEYE paper cites the standalone fork at AMTA for rendering training statistics.
• 2026Nine years on: tensorboardX records 5.9 million downloads a month and mxboard another 2,075 despite no release since 2018, while dmlc/tensorboard still ranks in the top 3% of machine-learning visualisation repositories on GitHub.

The underlying evidence

• Repositorydmlc/tensorboard — the project itself
• Originapache/mxnet#4003 — where the work was proposed and scoped in public, 27 November 2016
• AuthorshipApache MXNet documentation — “Zihao Zheng is the primary author of this repo”
• Downstreamawslabs/mxboard — AWS's official MXNet visualisation library, crediting this project
• DownstreamtensorboardX — the PyTorch ecosystem's equivalent, listing this project as a reference
• AcademicSOCKEYE (AMTA 2018) — Amazon researchers citing the standalone fork for rendering training statistics
• Recognitiondmlc/tensorboard#50 — Google's TensorFlow engineering director on the work, and on removing the dependency
• DemandAmazon's mxboard launch post, 18 May 2018 — users “started using TensorFlow because of the rich feature set offered in TensorBoard”, and a closing credit by name
• ScaletensorboardX and mxboard download statistics — live PyPI counts, re-checkable at any time
• StandingGitHub topic search, by stars: visualization + machine-learning and visualization — top 3% and top 2.4% respectively
• Todaytensorboard on PyPI — TensorFlow no longer appears in its requirements

← Back to home · All projects