Lapiz helps you send data to Tensorboard over HTTP (without requiring Tensorflow at the send point): it can then be used from PyTorch for instance. Of course, being a bridge, it adds flexibility at the cost of some latency (not yet measured).

The name Lapiz simply means "crayon" in Spanish, because this cool idea comes from crayon client/server library. I decided to rewrite crayon instead of making endless breaking PR because I was not convinced by the API or naming ("experiment" vs "run", etc) and its server code required quite some rewriting to adopt Tensorboard v1.3.0 HTTP API.

The client, a basic Python class that simply requires Python request library, sends json data to a server over HTTP. The server, a Flask app, then writes that data to Tensorboard using the Tensorflow tf.Summary API.

The server and embarked Tensorboard can be dockerized, which really makes it easy to use Tensorboard from outside of the Tensorflow ecosystem.

On the server instance, run the dockerized server

docker run -d -p 6006:6006 -p 6007:6007 --name lapiz bosr/lapiz

The first exposed port 6006 is for the browser to connect to Tensorboard web interface, the second 6007 is for lapiz Flask server.

In your project virtualenv, install the lapiz-client python library:

pip install lapiz-client  # available on Pypi

From a client:

import lapiz as lz

cl = lz.Client('192.168.78.36', port=6007)  # check connection to lapiz server (6007 by default)

with cl.run('run0,lr=1e-3,lambda=1e-1') as run:
    for _ in range(num_steps):

        # ...

        # the global_step starts at 0 by default, but override it with run.step(1)
        run.step()  # now step is 1

        run.add_scalar('accuracy', 1.3e-3)            # numpy types accepted
        run.add_histogram('hidden_weights', weights)  # flatten your weights
        run.add_image('sample', img)
        run.add_audio('audio', waveform)

        # run.save()  # not needed unless you cannot wait for the next run.step()

Each time the run.step() instruction is executed, the client saves all unsaved summaries, if any. You can save those summaries as early as you want with run.save(). The context manager makes it sure unsaved summaries will be saved at exit.

If the context manager style adopted by run does not fit you, use it directly but just don't forget to save and close:

run = cl.run('run0,lr=1e-3,lambda=1e-1')
run.step(36)
run.add_scalar('accuracy', 1e-3)
run.save()  # send to lapiz server
run.step(37)
run.add_scalar('G_loss', 1e-1)
run.wall_time(11.3)  # arbitrary wall_time
run.save()
run.close()  # or del run

The wall time associated to events is taken automatically when run.save() is executed, but it can be overwritten with run.wall_time(time.time()).

• Format of json messages: doc/JSON.md
• API specs for the Flask server (not yet updated): doc/specs.md
• More server-side options: doc/SERVER.md
• lapiz-client docs: https://github.com/bosr/lapiz-client