codecov[bot]
commented
4 years ago Codecov Report
Merging #34 into master will increase coverage by
0.98%. The diff coverage is84.47%.
@@ Coverage Diff @@
## master #34 +/- ##
==========================================
+ Coverage 60.42% 61.41% +0.98%
==========================================
Files 63 55 -8
Lines 5226 4901 -325
==========================================
- Hits 3158 3010 -148
+ Misses 2068 1891 -177| Flag | Coverage Δ | |
|---|---|---|
| #aprl | ? |
|
| #modelfree | ? |
| Impacted Files | Coverage Δ | |
|---|---|---|
| src/aprl/multi/common_worker.py | 100% <ø> (ø) |
|
| src/aprl/training/victim_envs.py | 94.44% <ø> (ø) |
|
| src/aprl/training/scheduling.py | 85.18% <ø> (ø) |
|
| tests/test_common.py | 100% <ø> (ø) |
|
| tests/test_agents.py | 98.97% <ø> (ø) |
|
| src/aprl/policies/base.py | 85.91% <ø> (ø) |
|
| src/aprl/training/gail_dataset.py | 100% <ø> (ø) |
|
| src/aprl/envs/sumo_auto_contact.py | 100% <ø> (ø) |
|
| src/aprl/activations/density/visualize.py | 0% <0%> (ø) |
|
| src/aprl/visualize/scores.py | 0% <0%> (ø) |
|
| ... and 42 more |
Continue to review full report at Codecov.
Legend - Click here to learn more
Δ = absolute <relative> (impact),ø = not affected,? = missing dataPowered by Codecov. Last update e9e8d16...aa03e9c. Read the comment docs.
This codebase depends on Bansal et al's environments in https://github.com/humancompatibleai/multiagent-competition which use MuJoCo 1.31. We cannot upgrade this since recent MuJoCo versions change the dynamics and we do not have code to retrain the policies. However, Gym and other packages have long since upgraded to more recent versions. So far, we have been working around this by having two virtual environments: one with the old versions, and one with the new versions, and backporting some fixes to try and keep the versions in sync as much as possible.
In branch https://github.com/AdamGleave/mujoco-py/tree/mj131 we introduce a new package
mujoco-py-131with the old bindings, that can be simultaneously installed with a modern version ofmujoco-py. In PR https://github.com/HumanCompatibleAI/multiagent-competition/pull/5 themultiagent-competitionenvironments are changed to use a modern version of Gym, with this old version ofmujoco-py-131. This PR upgradesadversarial-policiesto the latest version of Gym andmultiagent-competition. We now have a single requirements.txt and virtual environment in the CI.There is still room for further improvement: for example, the codebase could be reorganized now the distinction between
aprlandmodelfreeis less sharp.Fixes #1 (shrinks Docker image by 1.4 GB, although it's still quite large)