jiangsy/mbpo_pytorch

mbpo_pytorch's Issues

unable to find priv.yaml

Hi,
Firstly, thank you very much for a PyTorch implementation of MBPO. While I was going over your code, I noticed that you are using priv.yaml in run_mbpo.py, but I was unable to find this file in the config folder, apologies in advance if I have missed it.

Thanks,
Desik

AttributeError: 'SimpleUniversalBuffer' object has no attribute 'get_batch_generator'

I got this error report when I run "run_mbpo.py",and I actually didn't see this methed in class SimpleUniversalBuffer,there are only get_batch_generator_inf and get_batch_generator_epoch in this class,please tell me what should I do.

Could you show some results of using this repo on mujoco?

May I ask how long it takes to run it once? I only ran it 137 times in a whole day

May I ask how long it takes to run python ./mbpo_pytorch/scripts/run_mbpo.py once? I only ran it 137/1000 epochs in a whole day.

Updates of max_logvar and min_logvar parameters

Hi, thanks for this Pytorch version of MBPO.

In mbpo_pytorch/models/dynamics.py, the maximum and minimum log(var) are initalized as follows :
self.max_diff_state_logvar = nn.Parameter(torch.ones([1, state_dim]) / 2.)
self.min_diff_state_logvar = nn.Parameter(-torch.ones([1, state_dim]) * 10.)
self.max_reward_logvar = nn.Parameter(torch.ones([1, reward_dim]) / 2.)
self.min_reward_logvar = nn.Parameter(-torch.ones([1, reward_dim]) * 10.)

Are those parameters updated at any time ?

I couldn't find it, and then it wouldn't make sense to include them in the loss in mbpo_pytorch/algos/mbrl/mbpo.py
train_model_loss += \ 0.01 * (torch.sum(self.dynamics.max_diff_state_logvar) + torch.sum(self.dynamics.max_reward_logvar) - torch.sum(self.dynamics.min_diff_state_logvar) - torch.sum(self.dynamics.min_reward_logvar))

Best

Rewards in Hopper Benchmarking Environments

Hi,
I was going over the step function in hopper.py and noticed that the reward function used here is slightly different from the original one present in OpenAI gym, may I know the reason behind using a modified reward function?

Thanks,
Desik