Refactor NPO #1189
Refactor NPO #1189
Conversation
ghost
removed their request for review
src/garage/misc/tensor_utils.py
Outdated
|
|
||
| Returns: | ||
| numpy.ndarray: Stacked and padded tensor. Shape: :math:`(N, D, S^*)` | ||
| where K is the len of input paths. |
|
Can you describe in your commit message/subject in a little more detail what and why you are refactoring NPO? |
Done |
Codecov Report
@@ Coverage Diff @@
## master #1189 +/- ##
==========================================
- Coverage 87.99% 87.97% -0.02%
==========================================
Files 184 184
Lines 8771 8776 +5
Branches 1108 1111 +3
==========================================
+ Hits 7718 7721 +3
- Misses 854 855 +1
- Partials 199 200 +1
Continue to review full report at Codecov.
|
ahtsan
force-pushed
the
refactor_npo
branch
2 times, most recently
from
be4f360
to
8417901
Compare
src/garage/misc/tensor_utils.py
Outdated
| @@ -23,20 +23,26 @@ def discount_cumsum(x, discount): | |||
| axis=0)[::-1] | |||
|
|
|||
|
|
|||
| def explained_variance_1d(ypred, y): | |||
| def explained_variance_1d(ypred, y, valids): | |||
There was a problem hiding this comment.
hmm perhaps valids should be optional?
ahtsan
force-pushed
the
refactor_npo
branch
3 times, most recently
from
9a275f4
to
4c4ec69
Compare
| # pylint: disable=abstract-class-instantiated, no-member | ||
| # This test cause low memory with some reason | ||
| @pytest.mark.flaky |
There was a problem hiding this comment.
that would be really surprising. you seen this in the wild on master?
It seems likely that some other part of the test suite is using all if your memory, and this test just became a victim.
There was a problem hiding this comment.
that makes sense, I will remove the decorator
* Rename fit_baseline to fit_baseline_with_data to avoid naming conflicts with argument names in RL2NPO * Unify returns and rewards in sample_data to dense form
This PR refactors NPO, as discussed in #1172 , in preparation of RL2NPO.