Daniel Filan comments

Results 48 comments of


                                            Daniel Filan

CNN reward functions

Just realized that constant-length environments can be made much more simply.

Just want to flag that I just added a CNN potential function in `reward_nets.py`, as well as initializing and testing potentials in `test_reward_nets.py` (specifically in the `_make_env_and_save_reward_net` function, which is...

CNN reward functions

Have modified CNNs to only accept image observations and discrete actions (when actions are used), and when actions are used to output one reward value per possible action rather than...

CNN reward functions

Oops - tests turned out to not actually cover the changes to the new net output, and there are some errors there. Lemme fix those.

CNN reward functions

OK, errors fixed. This involved minimal changes to the reward inference code.

CNN reward functions

Anybody know why codecov isn't returning a report?

CNN reward functions

> Normally it waits until all tests pass to do so. But all other tests have passed.

CNN reward functions

Looks like that worked.

CNN reward functions

> The reduction in test coverage at https://app.codecov.io/gh/HumanCompatibleAI/imitation/compare/519/changes seems to be driven by: > > * forward of BasicPotentialCNN not being tested. We should probably add a test for that?...

CNN reward functions

Turns out no potentials were being covered by tests, since the code was calling the `predict` method of the `ShapedRewardNet`s, which in turn called the base network's `predict`, which called...