MarcelBruckner
commented
4 years ago - Expert trajectories and pretrained agents now in https://github.com/GAIL-4-BARK/large_data_store
- They get pulled via bazel with the dependency: data = [ "@com_github_gail_4_bark_large_data_store", ... ]
- They are added to the gail.py, tfa_generate.py and docs/report BUILDs and the paths to load them are adjusted
- Added examples/tfa_generate.py which we now use to generate sac expert trajectories.
- Reverted examples/tfa.py to be the original from Patrick
- I removed the command line flags in examples/gail.py, examples/tfa.py and examples/tfa_generate.py. We now set all settings except the mode via the params.json files
- Removed many launch configs as we now dont use command line arguments anymore
- Cleaned up and commented the usage of our examples/gail.py and examples/tfa_generate.py scripts
- Load save utils now use relative paths
- When training the log dirs are now in the bazel-bin/.../...runfiles/... folders
- Fixed the single_agent_runtime problem with a generic observation space
- If no expert trajectories are loaded (visualize mode) we can now pass None for them and the constructor creates an empty list for compatibility
- TF2RLRunner now creates the necessary model and log dirs
- Added SACRunnerGenerator subclass of SACRunner as suggested for the GenerateExpertTrajectories function
- Rendered tests are now a py_binary so that they do not fail on bazel test //... as they are anyways no real tests, but to visually confirm the replaying of the dataset
- Removed the pretrained agent from the docs/report as we now use the agent in https://github.com/GAIL-4-BARK/large_data_store
- Some small fixes to the report to be runnable with the new path loading for trajectories and experts