[tune] Cannot access logdir in functional API

[renatobellotti]

What is the problem?

I am trying to save something at the end of the training. To achieve this, I'm trying to get the save directory for the trial. The code from the documentation ("Trainable Logging" --> "Functional API") is as follows:

def trainable(config):
    library.init(
        name=trial_id,
        id=trial_id,
        resume=trial_id,
        reinit=True,
        allow_val_change=True)
    library.set_log_path(tune.track.logdir)

    for step in range(100):
        library.log_model(...)
        library.log(results, step=step)
        tune.track.log(results)

To me, it is not clear where the name library comes from. So I was assuming this had nothing to do with my problem, and thought the relevant part was the tune.track.logdir variable. However, that variable does not exist.

When I run the example code, I get the following error message:

AttributeError: module 'ray.tune' has no attribute 'track'

I'm new to ray and tune, so perhaps there is another way to get the log directory for the trial. I found the documentation a bit sparse about this issue if the functional API is used.

Reproduction (REQUIRED)

Ray version: 0.8.6 (from PyPI)

import ray
from ray import tune

def training_function(config):
    # see:
    # https://docs.ray.io/en/latest/tune/api_docs/logging.html
    # (section "Trainable Logging")
    tune.report(logdir=tune.track.logdir)

ray.init()

analysis = tune.run(training_function,
                    config={
                        'dummy': [1],
                    },
                    local_dir='test_local_dir',
                    num_samples=1,
                    resources_per_trial={'cpu': 1},
                    queue_trials=False)

[amogkam]

Hi, thanks for raising this issue. Yes I agree the documentation is pretty confusing here; we'll make sure to clear it up for the next release. tune.track has actually been deprecated in place of tune.report. In your example, there's actually no need to pass in a logdir to tune.report- it will automatically log to the trial directory. You simply need pass in a dictionary of the values you want to log https://docs.ray.io/en/latest/tune/api_docs/trainable.html#function-based-api. If you do want to get the logdir you can do so with tune.get_trial_dir() inside your training function.

The name library in the documentation is supposed to refer to any 3rd party library you would like to use for logging, it's not a variable name on its own.

[renatobellotti]

Thank you very much, now it is working well! 👍