what are the argument required by this script? Any example?

as the code written [here](def logit_bernoulli_entropy(logits_B):
ent_B = (1.-tensor.nnet.sigmoid(logits_B))*logits_B - logsigmoid(logits_B)
return ent_B), bernoulli was computed by this way

but it's different to the equation of binary entropy:
$-p\log p - (1-p)\log(1-p)$

is there any relationship between these two expressions? or why does openai compute bernoulli entropy that way? is there any theoretical equation support?

For anyone struggling to run the code:

1. Set floatX=float64 in your Theano Flags
2. The code as it is right now only works on a cluster. If you want to run individual experiments instead of doing phase1_train, do something like:

for x, y in zip(cmd_templates, argdicts):
    print(x.format(**y))

and then pick the commands you want and run them individually. In step2_eval, write something like:

if not os.path.exists(checkptfile):
    nonexistent_checkptfiles.append(checkptfile)
else:
    evals_to_do.append((task, alg, num_trajs, run, checkptfile))

instead of the assertion that would prevent you from running only some of the experiments.
3) I had to disable multithreading. If you run into these issues you can set the number of threads to 1 where you find it.
4) Some things are set up only for the humanoid environment in mujoco. For instance, if you are using a different environment, in environments/rlgymenv.py, track_body_name should be set to None and there is no sim.env.viewer -- use other methods like sim.env.render() if you want to view things.

Hello @hojonathanho,

does the code here swapped the order of step 4. and step 5. of the algorithm presented in the paper?

Hi,
Could you please provide a detailed documentation of the organization of the code and what all algorithms are available? Also, a "getting started" tutorial on running im_pipeline.py with standard/default command line arguments would be very helpful.
Thank you,
Anirban

Dear authors,

Thanks for sharing this high-quality code. l create expert policies by running scripts/scripts/run_rl_mj.py on CartPole-v0 successfully, but when l run the MountainCar-v0 with the same script, the score remains -200. Do you make some extra configurations over MountainCar-v0 during your expriments?

Thanks!

Hi, I am confused about this file when I using it. Is there any information about the detail description of API?

I'm getting this error when executing "log.write_snapshot(policy,i)" in line 393 of policyopt.nn.py (while running the main method in imitate_mj.py). Anyone know how to deal with this issue?

Hi,
Great job! it would be helpful to see an example json.
Thanks,
Nir

We try to run your project, but we are not sure about what 'dataset' in pipelines is exactly.
I set the parameters under the instructions in pipelines like this:
python imitate_mj.py
--mode ga
--env Humanoid-v1
--data log_humanoid_1.h5
--limit_trajs 160
--data_subsamp_freq 20
--favor_zero_expert_reward 0
--min_total_sa 50000
--max_iter 1501
--reward_include_time 0
--reward_lr .01
-log {out}

where is my fault?
Thank you very much.