Add benchmark metrics to simulator #823
Conversation
| float: partial success if supported, -1.0 otherwise | ||
| """ | ||
| assert self._done is not None, "At least one step() must occur before partial_success can be calculated!" | ||
| return len(satisfied) / (len(satisfied) + len(unsatisfied)) |
There was a problem hiding this comment.
Try to implement the satisfied etc. things - ask Eric how to do this
| super().__init__() | ||
| self._reward = 0 | ||
| self.initialized = False | ||
| self.state_cache = {} |
| if self.measure_work: | ||
| self._reward = 0 |
| if not self.measure_work: | ||
| self.state_cache = new_state_cache | ||
|
|
||
| self._reward += work_metric |
There was a problem hiding this comment.
Here do self._reward = work_metric if self.measure_work else (self._reward + work_metric)
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
omnigibson/envs/env_base.py
Outdated
| # record the start time of the simulation step in the beginning of the step | ||
| self._cur_sim_start_ts = time.perf_counter() |
There was a problem hiding this comment.
should we move these logic into the metrics themselves? e.g. the env class just call a bunch of pre_step for each of the metrics, something like this:
for metric in self.task.metrics:
metric.pre_step()
omnigibson/metrics/energy_metric.py
Outdated
| # TODO: this computation is very slow, consider using a more efficient method | ||
| # TODO: this method needs to be updated to account for object addition and removal |
There was a problem hiding this comment.
can we actually finish these two TODOs?
omnigibson/metrics/energy_metric.py
Outdated
| # TODO: this computation is very slow, consider using a more efficient method | ||
| # TODO: this method needs to be updated to account for object addition and removal | ||
| posrot2 = self.state_cache[linkname] | ||
| work_metric += np.linalg.norm(posrot[0] - posrot2[0]) * self.link_masses[linkname] |
There was a problem hiding this comment.
Maybe we should consider rotation changes as well, e.g. a weighted sum of translation changes and rotation changes, where the weights are hyperparams that can be configured.
There was a problem hiding this comment.
The work done by rotation should be W = torque * rotation. Without knowing torque, I added a rough calculation of the work/energy metric. I am happy to discuss this calculation more specifically with you tomorrow.
position, orientation = T.pose_transform(posrot[0], posrot[1], init_posrot[0], init_posrot[1])
work_metric += np.linalg.norm(orientation) * self.link_masses[linkname] * self.metric_config["rotation"]
| @@ -576,7 +576,7 @@ def get_control_dict(self): | |||
| # TODO: Move gravity force computation dummy to this class instead of BaseRobot | |||
| fcns["gravity_force"] = lambda: ( | |||
| ControllableObjectViewAPI.get_generalized_gravity_forces(self.articulation_root_path) | |||
| if self.fixed_base | |||
| if self.fixed_base or self._dummy is None | |||
There was a problem hiding this comment.
I believe that self._dummy is None should be there to account for the fact that dummy is not loaded at this stage of the simulator initialization. However, dummy will be loaded later, after the gravity force computation.
I also resolved this TODO, moving the dummy initialization from the BaseRobot to ControllableObject.
tests/test_metric_functions.py
Outdated
| assert metrics["work"] == 0, "Work metric was not reset" | ||
|
|
||
|
|
||
| def test_behavior_task_work_metric(): |
There was a problem hiding this comment.
can we add tests for all the metrics, e.g. wall time, energy, steps, etc.?
There was a problem hiding this comment.
env.step()
number of time step * action_time_step
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
initial commit to benchmark metrics