Closed ArgyChris closed 3 years ago
Hello,
perform a basic environment vectorization with a DummyVecEnv followed by VecNormalize.
Why do you use VecNormalize
with image? You should deactivate at least the obs normalization at it is done automatically for images (cf. doc) using norm_obs=False
.
Hello,
Thank you for the reply. I will keep in mind to deactivate the obs normalization.
Upon further investigation, it seems that there is a problem if I use the latest pytorch that supports CUDA=11.3. I found that if I use the same configuration as the older system I overcome the problem, and the loading of the robot model in the reset works fine. So, I created a new conda environment with the following that works.
3.8.5 Python
PyTorch version: 1.7.1+cu110
Gym version: 0.18.3
stable_baselines3 version: 1.0
With the previous configuration there is support for the latest GPU.
Thank you, Argy
it seems that there is a problem if I use the latest pytorch that supports CUDA=11.3.
So, I created a new conda environment with the following that works.
then it does not seem related to SB3...
then it does not seem related to SB3...
Closing because probably not related to SB3
Hello everyone,
I have just started learning the baselines. I am encountering the following problem, and I would appreciate if you could help me. I am trying to learn the reach of an object using the Kuka iiwa robot on pybullet using image based observations. First, I took the KukaGymEnv example from pybullet, I modified some steps, I choose PPO algorithm for Reinforcement Learning, and then I changed the policy network to a custom CNNPolicy. I perform a basic environment vectorization with a DummyVecEnv followed by VecNormalize. To monitor what is happening I get the observation (image), which is correct in the first reset. However, the problem that I have is that when VecNormalize is called there is a reset of the environment, and the robot turns into a pile of things (check output Figures 2, 3). Each subsequent reset creates the same error. I was not able to identify the source of error, and I would appreciate any input. I have to add that the code works fines, without any problem, if I use an older system without CUDA/GPU support (system 2).
Script for running the robot (stable baselines)
Observation and action space
custom_kukaGymEnv_reach_debug_CNN (reset, step, getExtendedObservation)
Kuka_CNN
custom_kuka_debug_2
Output
System Info
Describe the characteristic of your environment: System 1 (with error):
Additional context
System 2 where I do not encounter the problem: