def __init__(self, init_space, nb_bins):
if not isinstance(init_space, Box):
raise RuntimeError(
"Impossible to convert a gym space of type {} to a discrete space"
" (it should be of "
"type space.Box)"
"".format(type(init_space)))
if nb_bins < 2:
raise RuntimeError(
"This do not work with less that 1 bin (if you want to ignored some part "
"of the action_space or observation_space please use the "
"\"gym_space.ignore_attr\" or \"gym_space.keep_only_attr\"")
min_ = init_space.low
max_ = init_space.high
self._ignored = min_ == max_ # which component are ignored
self._res = min_
self._values = np.linspace(min_, max_, num=nb_bins + 2)
self._values = self._values[
1:-1, :] # the values that will be used when using #gym_to_glop
# TODO there might a cleaner approach here
self._bins_size = np.linspace(min_, max_, num=2 * nb_bins + 1)
self._bins_size = self._bins_size[
2:-1:2, :] # the values defining the "cuts"
self._gen_idx = np.arange(self._bins_size.shape[-1])
n_bins = np.ones(min_.shape[0]) * nb_bins
n_bins[
self.
_ignored] = 1 # if min and max are equal, i don't want to have multiple variable
BaseGymAttrConverter.__init__(
self,
space=MultiDiscrete(n_bins),
)
def __init__(self, init_space=None):
self.size = None
BaseGymAttrConverter.__init__(self, space=None)
if init_space is not None:
self.initialize_space(init_space)
self.previous_fun = self._previous_fun
self.after_fun = self._after_fun
def __init__(self, substract, divide, dtype=None, init_space=None):
BaseGymAttrConverter.__init__(self,
g2op_to_gym=None,
gym_to_g2op=None,
space=None)
self._substract = np.array(substract)
self._divide = np.array(divide)
self.dtype = dtype if dtype is not None else dt_float
if init_space is not None:
self.initialize_space(init_space)
def __init__(self, init_space):
if not isinstance(init_space, (MultiBinary, MultiDiscrete)):
raise RuntimeError("Impossible to convert a gym space of type {} to a Tuple (it should be of "
"type space.MultiBinary or space.MultiDiscrete)"
"".format(type(init_space)))
self.size = init_space.n
li = []
for i in range(self.size):
tmp_sz = 2
if isinstance(init_space, MultiDiscrete):
tmp_sz = init_space.nvec[i]
li.append(Discrete(tmp_sz))
BaseGymAttrConverter.__init__(self, space=Tuple(li))
def __init__(self, nb_bins, init_space=None):
BaseGymAttrConverter.__init__(self,
g2op_to_gym=None,
gym_to_g2op=None,
space=None)
if nb_bins < 2:
raise RuntimeError("This do not work with less that 1 bin (if you want to ignored some part "
"of the action_space or observation_space please use the "
"\"gym_space.ignore_attr\" or \"gym_space.keep_only_attr\"")
self._nb_bins = nb_bins
self._ignored = None
self._res = None
self._values = None
self._bins_size = None
self._gen_idx = None
if init_space is not None:
self.initialize_space(init_space)
def __init__(self, substract, divide, init_space, dtype=None):
self._substract = np.array(substract)
self._divide = np.array(divide)
self.dtype = dtype if dtype is not None else dt_float
if not isinstance(init_space, Box):
raise RuntimeError(
"Impossible to scale a converter if this one is not from type space.Box"
)
tmp_space = copy.deepcopy(init_space)
tmp_space.low = self.scale(tmp_space.low)
tmp_space.high = self.scale(tmp_space.high)
if dtype is not None:
tmp_space.dtype = dtype
tmp_space.low = tmp_space.low.astype(self.dtype)
tmp_space.high = tmp_space.high.astype(self.dtype)
BaseGymAttrConverter.__init__(self,
g2op_to_gym=self.scale,
gym_to_g2op=self.unscale,
space=tmp_space)
self.dtype = self.my_space.dtype
self._substract = self._substract.astype(self.dtype)
self._divide = self._divide.astype(self.dtype)