Closed Boubside closed 4 years ago
I saw this and thought I'd throw out this comment. Would rl-coach-slim work for you? this was a packaged designed for robomaker/sagemaker. I believe it removes the tensorflow requirements.
when I run help on this package.
NAME rl_coach
PACKAGE CONTENTS agents (package) architectures (package) base_parameters checkpoint coach core_types dashboard dashboard_components (package) data_stores (package) debug_utils environments (package) exploration_policies (package) filters (package) graph_managers (package) level_manager logger memories (package) off_policy_evaluators (package) orchestrators (package) plot_atari presets (package) renderer rollout_worker run_multiple_seeds saver schedules spaces tests (package) training_worker utilities (package) utils
Hi @Lacan82, Thanks for your answer. I might try this in the future. Something that worked for me was to pull the source code of coach from the Git, bypassing the need for installation. I then use coach with a dummy environment (that only purpose is to define action and state spaces) to perform the inference. It work fine and i didn't add dependency issue, even if I don"t have Intel_tensorflow installed.
I successfully manage to train an obstacle avoidance with distributed reinforcement learning using coach on AWS Robomaker/Sagemaker. Now I want to retrieve my model to use it for inference on the real robot. The brain of the robot Is a raspberry Pi 4.
So I tried to deploy on the Pi but the issue is I didn't manage to install rl_coach on the raspberry pi as there is no install wheel for intel_tensorflow on Raspbian. So my question is : is there a known way to install coach on a raspberry for inference ?
Alternatively, can I save my file to a Tensorflow file format to use it for inference ? I already found the way to save the graph as a .pb file, but then i can't load it back as I don't know the names of input and output tensors (bascally, my issue for this is the same than #374 and #71, any improvement on these ones ?). Do you have any guidance on the best way to do it with a clipped PPO algorithm and a custom environment?
Thanks a lot.