WandbMacros.jl

Convenient Julia macros for logging to weights and biases (Wandb) dashboard.

The macros provide a Julian way of interfacing with Wandb's Python API via PyCall. The functionality of each macro is transparent and obvious.

using Pkg
Pkg.add("WandbMacros")

using WandbMacros  # Automatically installs wandb (if not already installed) if PyCall.conda is true, else raises a prompt to install it.

@wandbinit project="Project1" name="Run1"  # Specify any keyword arguments that you would pass to wandb.init() in Python.

@wandbconfig seed=42 learning_rate=0.001 network_size=[64,32]  # Specify the config dictionary. Equivalent to wandb.config.update({"seed": 42, "learning_rate": 0.001, "network_size": [64,32]}, allow_val_change=True) in Python.

@wandbsave "file1.txt"  # Equivalent to wandb.save("file1.txt") in Python

@wandblog loss=0.1 accuracy=0.91 Validation/accuracy=0.75 step=100  # Equivalent to wandb.log({"loss":1, "accuracy":0.91, "Validation/accuracy":0.75}, step=100) in Python. `step` is an optional and a reserved keyword.

@wandbfinish  # equivalent to wandb.finish() in Python

1. using WandbMacros also exports the PyCall object wandb=pyimport("wandb"), which can be used to call Wandb functions that are not covered by the macros provided in this package.

2. @wandbinit, @wandbconfig and @wanbdlog work like Julia's @info macro. This allows for some powerful functionality:

   • Suppose you have the config parameters stored in a Julia dictionary named config_dict and it has keys of type Symbol, then you can splat the values by doing @wandbconfig config_dict.... You can also specify additional key-value pairs while splatting e.g., @wandbconfig param1=100 config_dict.... Similar functionality is available for @wandblog and @wandbinit.
   • If you have a julia variable named loss, then instead of logging it using @wandblog loss=loss, you can simply do @wandblog loss. This can be combined with other ways of specifying the arguments e.g., @wandblog accuracy=0.1 loss foo=x some_dict....

3. To run multiple instances of wandb in a process, do run1 = @wandbinit project="project1" name="run1" reinit=true and close the instance by calling @wandfinish run1.

4. Wandb logging can be disabled entirely (without commenting out the code) by setting environment variable JULIA_NO_WANDB=true, and enabled again by either unsetting the environment variable or setting it to JULIA_NO_WANDB=false. The environment variable can be set within the code using ENV["JULIA_NO_WANDB"]=true.

wandb.init() with PyCall is known to crash on Windows, unless you specify a keyword argument settings=wandb.Settings(start_method="thread"). But no worries! @wandbinit macro handles it automatically for you.