[Tune] status tables during training aren't showing values for tunable, nested config parameters #9307
Description
What is the problem?
Ray 0.8.6. Python 3.7.7.
Running this script (actually run with ipython):
from ray import tune
ray.init()
tune.run(
"PPO",
stop={"episode_reward_mean": 400},
config={
"env": "CartPole-v1",
"num_gpus": 0,
"num_workers": 1,
# "lr": tune.grid_search([0.01, 0.001, 0.0001]),
"model": {
'fcnet_hiddens': [tune.grid_search([20, 40, 60, 80]), tune.grid_search([20, 40, 60, 80])]
},
"eager": False,
},
)It appears that Tune easily drives the 16 different combinations for the two elements of fcnet_hidden (number of weights for two NN layers). All combinations easily reach the reward of 400, but some do so with lower iterations and/or elapsed times, which is what I want to determine.
The bug is in the status table printed. It has two columns for the tuning parameters, model/fcnet_hiddens/0 and model/fcnet_hiddens/1, but they are empty. I can determine the actual values by looking at the experiment_tag, e.g., experiment_tag: 8_fcnet_hiddens_0=20,fcnet_hiddens_1=60