TensorBoardLogger.jl

Log data to TensorBoard from Julia

TensorBoard is an interactive dashboard that can be used to visualize data generated from your programs. TensorBoardLogger is a Julia package that allows you to log data to TensorBoard through the standard Julia Logging system.

Installation

To install this Julia package run the following command in the julia REPL:

] add TensorBoardLogger
Note

If you need to log audio files, you must also install WAV. If you need to log Images, you must also install ImageMagick (on MacOS, you will also need QuartzImageIO). You only need to install those dependencies, and you will never need to load them.

] add ImageMagick
Note

TensorBoardLogger only contains code to serialize data to TensorBoard. To run the TensorBoard interface, you should install it and run it. The easiest way to install it is to install TensorFlow. Instructions for running TensorBoard can be found on the GitHub page.

Basic Usage

The fundamental type defined in this package is a TBLogger, which behaves like other standard loggers in Julia such as ConsoleLogger or TextLogger. You can create one by passing it the path to the folder where you want to store the data. You can also pass an optional second argument to specify the behaviour in case there given path exists.

Once you have created a TBLogger, you can use it as you would use any other logger in Julia:

  • You can set it to be your global logger with the function global_logger
  • You can set it to be the current logger in a scope with the function with_logger
  • You can combine it with other Loggers using LoggingExtras.jl, so that messages are logged to TensorBoard and to other backends at the same time.

Every TBLogger has an internal counter to store the current step, which is initially set to 1. All the data logged with the same @log call will be logged with the same step, and then it will increment the internal counter by 1.

If you want to increase the counter by a different amount, or prevent it from increasing, you can log the additional message log_step_increment=N. The default behaviour corresponds to N=1. If you set N=0 the internal counter will not be modified. The defaul behaviour for logger lg can be changed executing set_step_increment!(lg, N).

See the example below:

using TensorBoardLogger, Logging, Random

lg=TBLogger("tensorboard_logs/run", min_level=Logging.Info)

struct sample_struct first_field; other_field; end

with_logger(lg) do
    for i=1:100
        x0          = 0.5+i/30; s0 = 0.5/(i/20);
        edges       = collect(-5:0.1:5)
        centers     = collect(edges[1:end-1] .+0.05)
        histvals    = [exp(-((c-x0)/s0)^2) for c=centers]
        data_tuple  = (edges, histvals)
        data_struct = sample_struct(i^2, i^1.5-0.3*i)


        @info "test" i=i j=i^2 dd=rand(10).+0.1*i hh=data_tuple
        @info "test_2" i=i j=2^i hh=data_tuple log_step_increment=0
        @info "" my_weird_struct=data_struct   log_step_increment=0
        @debug "debug_msg" this_wont_show_up=i
    end
end

Backends

TensorBoardLogger supports five backends: - Scalars - Distributions - Images - Audio - Text/Markdown

When you log data to tensorboard, it is processed to convert it to the format used by TensorBoard. In particular, fields in structures are logged individually and complex numbers/arrays are split into their real and imaginary parts. Depending on its type, an object is serialized to the relevant backend, with the text backend being the default fallback.

For more detailed informations on how data is dispatched to a certain backend, and how to specify a desired backend refer to Specifying a backend.

If you want to define a new default behaviour for a custom type refer to section Extending TensorBoardLogger.

Deserialization

If you want to read back the data stored in TensorBoard/.proto files, have a look at Reading back TensorBoard data

Third-party packages

We also support logging custom types from a the following third-party libraries:

  • Plots.jl: the Plots.Plot type will be rendered to PNG at the resolution specified by the object and logged as an image
  • PyPlot.jl: the PyPlot.Figure type will be rendered to PNG at the resolution specified by the object and logged as an image
  • Gadfly.jl: the Gadfly.Plot type will be rendered to PNG at the resolution specified by the object and logged as an image. Cairo and Fontconfig packages must be imported for this functionality to work as it is required by Gadfly.
  • Tracker.jl: the TrackedReal and TrackedArray types will be logged as vector data
  • ValueHistories.jl: the MVHistory type is used to store the deserialized content of .proto files.

Explicit logging

As an alternative, you can also log data to TensorBoard through its functional interface, by calling the relevant method with a tag string and the data. For information on this interface refer to Explicit interface.

Hyperparameter tuning

Many experiments rely on hyperparameters, which can be difficult to tune. Tensorboard allows you to visualise the effect of your hyperparameters on your metrics, giving you an intuition for the correct hyperparameters for your task. For information on this API, see the Hyperparameter logging manual page.

Other exported methods

For documentation on other exported methods, see the Exported manual page.