Make SemiCompactObsevations be the only option (Metta-AI#256)

Some months ago we added SemiCompactObservations, where layers would be determined by property_name, versus (type_id, property_name). We've been using this as the default for a while, so now we're going to go ahead and remove the competition, and steal its name. No unit tests for this, since its cython. We should add some once that's addressed. Tested by building and training locally.

Author: sasmith

configs/env/mettagrid/mettagrid.yaml

@@ -3,7 +3,6 @@ _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 report_stats_interval: 100
 normalize_rewards: false
-semi_compact_obs: true
 track_last_action: false
 
 sampling: 0

configs/env/mettagrid/puffer.yaml

@@ -3,7 +3,6 @@ _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 report_stats_interval: 100
 normalize_rewards: false
-semi_compact_obs: true
 track_last_action: false
 
 sampling: 0

configs/user/sasmith.yaml

@@ -33,8 +33,6 @@ wandb:
  track: true
 
 env:
- # clean this up
- semi_compact_obs: true
  game:
  max_steps: 2000
 

deps/mettagrid/configs/benchmark.yaml

@@ -1,7 +1,6 @@
 # This is a modified `sample.yaml` config with sampling removed, for benchmark_env_perf.py script.
 report_stats_interval: 100
 normalize_rewards: false
-semi_compact_obs: true
 track_last_action: false
 
 sampling: 0

deps/mettagrid/configs/show_map.yaml

@@ -1,7 +1,6 @@
 # This config is used by `tools.map.view` script.
 report_stats_interval: 100
 normalize_rewards: false
-semi_compact_obs: true
 track_last_action: false
 
 game:

deps/mettagrid/configs/test_basic.yaml

@@ -1,7 +1,6 @@
 _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 normalize_rewards: false
-semi_compact_obs: true
 track_last_action: false
 
 sampling: 0

deps/mettagrid/mettagrid/grid_env.pyx

@@ -318,7 +318,6 @@ cdef class GridEnv:
 
  @property
  def observation_space(self):
- space = self._obs_encoder.observation_space()
  return gym.spaces.Box(
  0,
  255,

deps/mettagrid/mettagrid/mettagrid.pyx

@@ -16,8 +16,7 @@ from mettagrid.grid_env cimport GridEnv
 from mettagrid.grid_object cimport GridObject
 from mettagrid.observation_encoder cimport (
  ObsType,
- ObservationEncoder,
- SemiCompactObservationEncoder
+ ObservationEncoder
 )
 
 # Object imports
@@ -51,8 +50,6 @@ cdef class MettaGrid(GridEnv):
  self._cfg = cfg
 
  obs_encoder = ObservationEncoder()
- if env_cfg.semi_compact_obs:
- obs_encoder = SemiCompactObservationEncoder()
  cdef vector[ActionHandler*] actions
  if cfg.actions.put_items.enabled:
  actions.push_back(new PutRecipeItems(cfg.actions.put_items))
@@ -167,7 +164,7 @@ cdef class MettaGrid(GridEnv):
 
  cpdef grid_objects(self):
  cdef GridObject *obj
- cdef ObsType[:] obj_data = np.zeros(len(self.grid_features()), dtype=self._obs_encoder.obs_np_type())
+ cdef ObsType[:] obj_data = np.zeros(len(self.grid_features()), dtype=np.uint8)
  cdef unsigned int obj_id, i
  cdef ObservationEncoder obs_encoder = self._obs_encoder
  cdef vector[unsigned int] offsets

deps/mettagrid/mettagrid/observation_encoder.pxd

@@ -14,9 +14,4 @@ cdef class ObservationEncoder:
  cdef encode(self, GridObject *obj, ObsType[:] obs)
  cdef _encode(self, GridObject *obj, ObsType[:] obs, vector[unsigned int] offsets)
  cdef vector[string] feature_names(self)
- cpdef observation_space(self)
- cpdef obs_np_type(self)
-
-cdef class SemiCompactObservationEncoder(ObservationEncoder):
- pass
 

deps/mettagrid/mettagrid/observation_encoder.pyx

@@ -1,6 +1,3 @@
-import numpy as np
-import gymnasium as gym
-
 from libcpp.string cimport string
 from libcpp.vector cimport vector
 from libcpp.map cimport map
@@ -14,17 +11,13 @@ from mettagrid.objects.converter cimport Converter
 from mettagrid.objects.wall cimport Wall
 
 cdef class ObservationEncoder:
- cpdef obs_np_type(self):
- return np.uint8
- 
  cdef init(self, unsigned int obs_width, unsigned int obs_height):
  self._obs_width = obs_width
  self._obs_height = obs_height
 
  def __init__(self) -> None:
  self._offsets.resize(ObjectType.Count)
  self._type_feature_names.resize(ObjectType.Count)
- features = []
 
  self._type_feature_names[ObjectType.AgentT] = Agent.feature_names()
  self._type_feature_names[ObjectType.WallT] = Wall.feature_names()
@@ -34,38 +27,7 @@ cdef class ObservationEncoder:
  # remove these types from code.
  for type_id in [ObjectType.AltarT, ObjectType.ArmoryT, ObjectType.FactoryT, ObjectType.GeneratorT, ObjectType.LabT, ObjectType.LaseryT, ObjectType.MineT, ObjectType.TempleT]:
  self._type_feature_names[type_id] = Converter.feature_names(type_id)
-
- for type_id in range(ObjectType.Count):
- for i in range(len(self._type_feature_names[type_id])):
- self._offsets[type_id].push_back(len(features))
- features.append(self._type_feature_names[type_id][i])
- self._feature_names = features
-
- cdef encode(self, GridObject *obj, ObsType[:] obs):
- self._encode(obj, obs, self._offsets[obj._type_id])
-
- cdef _encode(self, GridObject *obj, ObsType[:] obs, vector[unsigned int] offsets):
- obj.obs(&obs[0], offsets)
-
- cdef vector[string] feature_names(self):
- return self._feature_names
- 
- cpdef observation_space(self):
- type_info = np.iinfo(self.obs_np_type())
-
- return gym.spaces.Box(
- low=type_info.min, high=type_info.max,
- shape=(
- len(self.feature_names()),
- self._obs_height, self._obs_width),
- dtype=self.obs_np_type()
- )
-
-cdef class SemiCompactObservationEncoder(ObservationEncoder):
- def __init__(self) -> None:
- super().__init__()
- self._offsets.resize(ObjectType.Count)
- self._type_feature_names.resize(ObjectType.Count)
+ 
  # Generate an offset for each unique feature name.
  cdef map[string, int] features
  cdef vector[string] feature_names
@@ -77,6 +39,15 @@ cdef class SemiCompactObservationEncoder(ObservationEncoder):
  # Set the offset for each feature, using the global offsets.
  for type_id in range(ObjectType.Count):
  for i in range(len(self._type_feature_names[type_id])):
- self._offsets[type_id][i] = features[self._type_feature_names[type_id][i]]
+ self._offsets[type_id].push_back(features[self._type_feature_names[type_id][i]])
 
  self._feature_names = feature_names
+
+ cdef encode(self, GridObject *obj, ObsType[:] obs):
+ self._encode(obj, obs, self._offsets[obj._type_id])
+
+ cdef _encode(self, GridObject *obj, ObsType[:] obs, vector[unsigned int] offsets):
+ obj.obs(&obs[0], offsets)
+
+ cdef vector[string] feature_names(self):
+ return self._feature_names