The very first example that Tune course runs is defined like this:
analysis = tune.run(
"PPO", # Use proximal policy optimization to train
stop={"episode_reward_mean": 400}, # Stopping criteria, when average reward over the episodes
# of training equals 400 out of a maximum possible 500 score.
config={
"env": "CartPole-v1", # Tune can associate this string with the environment.
"num_gpus": 0, # If you have GPUs, go for it!
"num_workers": 3, # Number of Ray workers to use; Use one LESS than
# the number of cores you wan to use (or omit this argument)!
"model": { # The NN model we'll optimize.
'fcnet_hiddens': [ # "Fully-connected network with N hidden layers".
tune.grid_search([20, 40]), # Try these four values for layer one.
tune.grid_search([20, 40]) # Try these four values for layer one.
]
},
"eager": False, # Flag for TensorFlow; don't use eager evaluation.
},
verbose=1
)
This failed for me, saying that Tensorflow doesn't recognize eager parameter. So I've removed it, and it failed again, this time there was some TF/Numpy conversion issue. Then I found this issue , and it confirms that eager should not be used anymore. Following the conversation there, I got the following working code:
analysis = tune.run(
"PPO", # Use proximal policy optimization to train
stop={"episode_reward_mean": 400}, # Stopping criteria, when average reward over the episodes
# of training equals 400 out of a maximum possible 500 score.
config={
"env": "CartPole-v1", # Tune can associate this string with the environment.
"num_gpus": 0, # If you have GPUs, go for it!
"num_workers": 3, # Number of Ray workers to use; Use one LESS than
# the number of cores you wan to use (or omit this argument)!
"model": { # The NN model we'll optimize.
'fcnet_hiddens': [ # "Fully-connected network with N hidden layers".
tune.grid_search([20, 40]), # Try these four values for layer one.
tune.grid_search([20, 40]) # Try these four values for layer one.
]
},
"framework": "tfe",
#"eager": False, # Flag for TensorFlow; don't use eager evaluation.
},
verbose=1
)
Not sure if this is the proper fix, so just an issue here and not a PR. But hopefully that helps the next person going through the Tune course.
The very first example that Tune course runs is defined like this:
This failed for me, saying that Tensorflow doesn't recognize
eager
parameter. So I've removed it, and it failed again, this time there was some TF/Numpy conversion issue. Then I found this issue , and it confirms thateager
should not be used anymore. Following the conversation there, I got the following working code:Not sure if this is the proper fix, so just an issue here and not a PR. But hopefully that helps the next person going through the Tune course.