Skip to content

Commit

Permalink
Merge pull request #574 from BDonnot/bd_dev
Browse files Browse the repository at this point in the history
Fix MaskedEnvironment
  • Loading branch information
BDonnot committed Jan 11, 2024
2 parents 5c42ffb + f1310c5 commit 628547b
Show file tree
Hide file tree
Showing 11 changed files with 264 additions and 161 deletions.
1 change: 1 addition & 0 deletions CHANGELOG.rst
Expand Up @@ -39,6 +39,7 @@ Change Log
- [FIXED] `PandapowerBackend`, when no slack was present
- [FIXED] the "BaseBackendTest" class did not correctly detect divergence in most cases (which lead
to weird bugs in failing tests)
- [FIXED] an issue with imageio having deprecated the `fps` kwargs (see https://github.com/rte-france/Grid2Op/issues/569)
- [ADDED] A type of environment that does not perform the "emulation of the protections"
for some part of the grid (`MaskedEnvironment`) see https://github.com/rte-france/Grid2Op/issues/571
- [IMPROVED] the CI speed: by not testing every possible numpy version but only most ancient and most recent
Expand Down
2 changes: 1 addition & 1 deletion docs/action.rst
Expand Up @@ -85,7 +85,7 @@ you want to perform on the grid. For more information you can consult the help o

To avoid extremely verbose things, as of grid2op 1.5.0, we introduced some convenience functions to allow
easier action construction. You can now do `act.load_set_bus = ...` instead of the previously way
more verbose `act.update({"set_bus": {"loads_id": ...}}`
more verbose `act.update({"set_bus": {"loads_id": ...}})`

.. _action-module-examples:

Expand Down
2 changes: 1 addition & 1 deletion docs/environment.rst
Expand Up @@ -101,7 +101,7 @@ be equivalent to starting into the "middle" of a video game. If that is the case
Finally, you might have noticed that each call to "env.reset" might take a while. This can dramatically
increase the training time, especially at the beginning. This is due to the fact that each time
`env.reset` is called, the whole chronics is read from the hard drive. If you want to lower this
impact then you might consult the `Optimize the data pipeline`_ section.
impact then you might consult the :ref:`environment-module-data-pipeline` page of the doc.

.. _environment-module-chronics-info:

Expand Down
59 changes: 59 additions & 0 deletions grid2op/Environment/baseEnv.py
Expand Up @@ -84,6 +84,65 @@ class BaseEnv(GridObjects, RandomObject, ABC):
The documentation is showed here to document the common attributes of an "BaseEnvironment".
.. _danger-env-ownership:
Notes
------------------------
Note en environment data ownership
.. danger::
A non pythonic decision has been implemented in grid2op for various reasons: an environment
owns everything created from it.
This means that if you (or the python interpreter) deletes the environment, you might not
use some data generate with this environment.
More precisely, you cannot do something like:
.. code-block:: python
import grid2op
env = grid2op.make("l2rpn_case14_sandbox")
saved_obs = []
obs = env.reset()
saved_obs.append(obs)
obs2, reward, done, info = env.step(env.action_space())
saved_obs.append(obs2)
saved_obs[0].simulate(env.action_space()) # works
del env
saved_obs[0].simulate(env.action_space()) # DOES NOT WORK
It will raise an error like `Grid2OpException EnvError "This environment is closed. You cannot use it anymore."`
This will also happen if you do things inside functions, for example like this:
.. code-block:: python
import grid2op
def foo(manager):
env = grid2op.make("l2rpn_case14_sandbox")
obs = env.reset()
manager.append(obs)
obs2, reward, done, info = env.step(env.action_space())
manager.append(obs2)
manager[0].simulate(env.action_space()) # works
return manager
manager = []
manager = foo(manager)
manager[0].simulate(env.action_space()) # DOES NOT WORK
The same error is raised because the environment `env` is automatically deleted by python when the function `foo` ends
(well it might work on some cases, if the function is called before the variable `env` is actually deleted but you
should not rely on this behaviour.)
Attributes
----------
Expand Down
8 changes: 8 additions & 0 deletions grid2op/Environment/environment.py
Expand Up @@ -37,6 +37,14 @@ class Environment(BaseEnv):
"""
This class is the grid2op implementation of the "Environment" entity in the RL framework.
.. danger::
Long story short, once a environment is deleted, you cannot use anything it "holds" including,
but not limited to the capacity to perform `obs.simulate(...)` even if the `obs` is still
referenced.
See :ref:`danger-env-ownership` (first danger block).
Attributes
----------
Expand Down
21 changes: 15 additions & 6 deletions grid2op/Environment/maskedEnvironment.py
Expand Up @@ -27,7 +27,14 @@ class MaskedEnvironment(Environment): # TODO heritage ou alors on met un truc d
.. warning::
At time of writing, the behaviour of "obs.simulate" is not modified
"""
CAN_SKIP_TS = False # some steps can be more than one time steps
# some kind of infinity value
# NB we multiply np.finfo(dt_float).max by a small number (1e-7) to avoid overflow
# indeed, _hard_overflow_threshold is multiply by the flow on the lines
INF_VAL_THM_LIM = 1e-7 * np.finfo(dt_float).max

# some kind of infinity value
INF_VAL_TS_OVERFLOW_ALLOW = np.iinfo(dt_int).max - 1

def __init__(self,
grid2op_env: Union[Environment, dict],
lines_of_interest):
Expand All @@ -38,7 +45,7 @@ def __init__(self,
elif isinstance(grid2op_env, dict):
super().__init__(**grid2op_env)
else:
raise EnvError(f"For TimedOutEnvironment you need to provide "
raise EnvError(f"For MaskedEnvironment you need to provide "
f"either an Environment or a dict "
f"for grid2op_env. You provided: {type(grid2op_env)}")

Expand All @@ -62,10 +69,8 @@ def _make_lines_of_interest(self, lines_of_interest):

def _reset_vectors_and_timings(self):
super()._reset_vectors_and_timings()
self._hard_overflow_threshold[~self._lines_of_interest] = 1e-7 * np.finfo(dt_float).max # some kind of infinity value
# NB we multiply np.finfo(dt_float).max by a small number to avoid overflow
# indeed, _hard_overflow_threshold is multiply by the flow on the lines
self._nb_timestep_overflow_allowed[~self._lines_of_interest] = np.iinfo(dt_int).max - 1 # some kind of infinity value
self._hard_overflow_threshold[~self._lines_of_interest] = type(self).INF_VAL_THM_LIM
self._nb_timestep_overflow_allowed[~self._lines_of_interest] = type(self).INF_VAL_TS_OVERFLOW_ALLOW

def get_kwargs(self, with_backend=True, with_chronics_handler=True):
res = {}
Expand All @@ -79,6 +84,10 @@ def get_params_for_runner(self):
res["other_env_kwargs"] = {"lines_of_interest": copy.deepcopy(self._lines_of_interest)}
return res

def _custom_deepcopy_for_copy(self, new_obj):
super()._custom_deepcopy_for_copy(new_obj)
new_obj._lines_of_interest = copy.deepcopy(self._lines_of_interest)

@classmethod
def init_obj_from_kwargs(cls,
other_env_kwargs,
Expand Down
13 changes: 9 additions & 4 deletions grid2op/Episode/EpisodeReplay.py
Expand Up @@ -102,15 +102,15 @@ def replay_episode(
load_info: ``str``
Defaults to "p". What kind of values to show on loads.
Can be oneof `["p", "v", None]`
Can be one of `["p", "v", None]`
gen_info: ``str``
Defaults to "p". What kind of values to show on generators.
Can be oneof `["p", "v", None]`
Can be one of `["p", "v", None]`
line_info: ``str``
Defaults to "rho". What kind of values to show on lines.
Can be oneof `["rho", "a", "p", "v", None]`
Can be one of `["rho", "a", "p", "v", None]`
resolution: ``tuple``
Defaults to (1280, 720). The resolution to use for the gif.
Expand Down Expand Up @@ -187,7 +187,12 @@ def replay_episode(
# Export all frames as gif if enabled
if gif_name is not None and len(frames) > 0:
try:
imageio.mimwrite(gif_path, frames, fps=fps)
try:
# with imageio > 2.5 you need to compute the duration
imageio.mimwrite(gif_path, frames, duration=1000./fps)
except TypeError:
# imageio <= 2.5 can be given fps directly
imageio.mimwrite(gif_path, frames, fps=fps)
# Try to compress
try:
from pygifsicle import optimize
Expand Down
33 changes: 31 additions & 2 deletions grid2op/Observation/baseObservation.py
Expand Up @@ -4207,7 +4207,18 @@ def get_forecast_env(self) -> "grid2op.Environment.Environment":
f_obs_3, *_ = forecast_env.step(act_3)
sim_obs_3, *_ = sim_obs_2.simulate(act_3)
# f_obs_3 should be sim_obs_3
.. danger::
Long story short, once a environment (and a forecast_env is one)
is deleted, you cannot use anything it "holds" including,
but not limited to the capacity to perform `obs.simulate(...)` even if the `obs` is still
referenced.
See :ref:`danger-env-ownership` (first danger block).
This caused issue https://github.com/rte-france/Grid2Op/issues/568 for example.
Returns
-------
grid2op.Environment.Environment
Expand Down Expand Up @@ -4339,8 +4350,26 @@ def get_env_from_external_forecasts(self,
you have 100 rows then you have 100 steps.
.. warning::
We remind that, if you provide some forecasts, it is expected that
We remind that, if you provide some forecasts, it is expected that they allow some powerflow to converge.
The balance between total generation on one side and total demand and losses on the other should also
make "as close as possible" to reduce some modeling artifact (by the backend, grid2op does not check
anything here).
Finally, make sure that your input data meet the constraints on the generators (pmin, pmax and ramps)
otherwise you might end up with incorrect behaviour. Grid2op supposes that data fed to it
is consistent with its model. If not it's "undefined behaviour".
.. danger::
Long story short, once a environment (and a forecast_env is one)
is deleted, you cannot use anything it "holds" including,
but not limited to the capacity to perform `obs.simulate(...)` even if the `obs` is still
referenced.
See :ref:`danger-env-ownership` (first danger block).
This caused issue https://github.com/rte-france/Grid2Op/issues/568 for example.
Examples
--------
A typical use might look like
Expand Down
1 change: 1 addition & 0 deletions grid2op/Runner/runner.py
Expand Up @@ -1137,6 +1137,7 @@ def run(
returned list are not necessarily sorted by this value)
- "cum_reward" the cumulative reward obtained by the :attr:`Runner.Agent` on this episode i
- "nb_time_step": the number of time steps played in this episode.
- "total_step": the total number of time steps possible in this episode.
- "episode_data" : [Optional] The :class:`EpisodeData` corresponding to this episode run only
if `add_detailed_output=True`
- "add_nb_highres_sim": [Optional] The estimated number of calls to high resolution simulator made
Expand Down
2 changes: 1 addition & 1 deletion grid2op/tests/BaseBackendTest.py
Expand Up @@ -2741,7 +2741,7 @@ def test_issue_134(self):
}
)
obs, reward, done, info = env.step(action)
assert not done
assert not done, f"Episode should not have ended here, error : {info['exception']}"
assert obs.line_status[LINE_ID] == False
assert obs.topo_vect[obs.line_or_pos_topo_vect[LINE_ID]] == -1
assert obs.topo_vect[obs.line_ex_pos_topo_vect[LINE_ID]] == -1
Expand Down

0 comments on commit 628547b

Please sign in to comment.