Customize learning_rate in autogluon classification .fit()

I’m a student working on an autogluon project and can’t get the learning rate modified. I have an oscillating HPO chart and figured that was a good place to start.

Here’s my code:

import autogluon.core as ag
from autogluon import ImageClassification as task
from mxnet import optimizer as optim

dataset = task.Dataset('data/train')

@ag.obj(
        learning_rate=ag.space.Real(0.1111, 0.22222),
        momentum=ag.space.Real(0.86, 0.99),
        wd=ag.space.Real(1e-5, 1e-3, log=True)
        )        
class NAG(optim.NAG):
    pass
optimizer = NAG()

time_limit = 10*60
num_epochs = 30

classifier = task.fit(dataset, optimizer=optimizer, time_limits=time_limit, epochs=num_epochs, verbose=4)

[Innixma]

You probably aren’t using the pre-release then. I’d recommend creating a fresh venv without autogluon installed, and then following the pre-release install instructions.

[mhup115]

I just tried updating to the pre-release, it says i’m running the latest version and to restart the kernel.

When following these docs on customizing the search space, the logs produced don’t reflect any of these changes. Pretty stumped unfortunately