steffen/marl-factory-grid
1
0
Fork 0
mirror of https://github.com/illiumst/marl-factory-grid.git synced 2025-10-31 04:37:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

new observation properties for testing of technical limitations '' Debuggings

Browse Source
This commit is contained in:
Steffen Illium
2021-11-07 18:15:31 +01:00
parent d69cf75c15
commit ab9cde7853
4 changed files with 28 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
environments/factory/env_default_param.yaml
View File

@@ -1,23 +0,0 @@
combin_agent_slices_in_obs: true
dirt_properties: !!python/object/new:environments.factory.simple_factory.DirtProperties
- 1
- 0.05
- 0.1
- 3
- 1
- 20
- 0.0
done_at_collision: false
frames_to_stack: 0
level_name: rooms
max_steps: 400
movement_properties: !!python/object/new:environments.utility_classes.MovementProperties
- true
- true
- false
n_agents: 1
omit_agent_slice_in_obs: true
parse_doors: false
pomdp_radius: 3
record_episodes: false
verbose: false

29
environments/factory/env_item_default_param.json
View File

@@ -1,29 +0,0 @@
{
"item_properties": {
"n_items": 5,
"spawn_frequency": 10,
"n_drop_off_locations": 5,
"max_dropoff_storage_size": 0,
"max_agent_inventory_capacity": 5,
"agent_can_interact": true
},
"env_seed": 2,
"movement_properties": {
"allow_square_movement": true,
"allow_diagonal_movement": true,
"allow_no_op": false
},
"level_name": "rooms",
"verbose": false,
"n_agents": 1,
"max_steps": 400,
"pomdp_r": 2,
"combin_agent_obs": true,
"omit_agent_in_obs": true,
"cast_shadows": true,
"frames_to_stack": 3,
"done_at_collision": false,
"record_episodes": false,
"parse_doors": false,
"doors_have_area": false
}

2
environments/utility_classes.py
View File

@@ -3,7 +3,7 @@ from typing import NamedTuple, Union
class AgentRenderOptions(object):
SEPERATE = 'each'
SEPERATE = 'seperate'
COMBINED = 'combined'
LEVEL = 'lvl'
NOT = 'not'

45
studies/e_1.py
View File

@@ -96,7 +96,7 @@ def load_model_run_baseline(seed_path, env_to_run):
# retrieve model class
model_cls = next(val for key, val in h.MODEL_MAP.items() if key in seed_path.parent.name)
# Load both agents
model = model_cls.load(seed_path / 'model.zip')
model = model_cls.load(seed_path / 'model.zip', device='cpu')
# Load old env kwargs
with next(seed_path.glob('*.json')).open('r') as f:
env_kwargs = simplejson.load(f)
@@ -128,7 +128,7 @@ def load_model_run_study(seed_path, env_to_run, additional_kwargs_dict):
# retrieve model class
model_cls = next(val for key, val in h.MODEL_MAP.items() if key in seed_path.parent.name)
# Load both agents
models = [model_cls.load(seed_path / 'model.zip') for _ in range(n_agents)]
models = [model_cls.load(seed_path / 'model.zip', device='cpu') for _ in range(n_agents)]
# Load old env kwargs
with next(seed_path.glob('*.json')).open('r') as f:
env_kwargs = simplejson.load(f)
@@ -179,6 +179,7 @@ if __name__ == '__main__':
# Define properties object parameters
obs_props = ObservationProperties(render_agents=AgentRenderOptions.NOT,
omit_agent_self=True,
additional_agent_placeholder=None,
frames_to_stack=3,
pomdp_r=2
)
@@ -202,12 +203,12 @@ if __name__ == '__main__':
# Bundle both environments with global kwargs and parameters
env_map = {'dirt': (DirtFactory, dict(dirt_prop=dirt_props,
**factory_kwargs)),
**factory_kwargs.copy())),
'item': (ItemFactory, dict(item_prop=item_props,
**factory_kwargs)),
**factory_kwargs.copy())),
'itemdirt': (DirtItemFactory, dict(dirt_prop=dirt_props,
item_prop=item_props,
**factory_kwargs))}
**factory_kwargs.copy()))}
env_names = list(env_map.keys())
# Define parameter versions according with #1,2[1,0,N],3
@@ -240,6 +241,7 @@ if __name__ == '__main__':
dict(obs_prop=ObservationProperties(
render_agents=AgentRenderOptions.LEVEL,
omit_agent_self=True,
additional_agent_placeholder=None,
frames_to_stack=3,
pomdp_r=2)
)
@@ -249,6 +251,7 @@ if __name__ == '__main__':
post_training_kwargs=
dict(obs_prop=ObservationProperties(
render_agents=AgentRenderOptions.NOT,
additional_agent_placeholder=None,
omit_agent_self=True,
frames_to_stack=3,
pomdp_r=2)
@@ -259,18 +262,18 @@ if __name__ == '__main__':
# Train starts here ############################################################
# Build Major Loop parameters, parameter versions, Env Classes and models
if True:
for observation_mode in observation_modes.keys():
for obs_mode in observation_modes.keys():
for env_name in env_names:
for model_cls in [h.MODEL_MAP['A2C'], h.MODEL_MAP['DQN']]:
# Create an identifier, which is unique for every combination and easy to read in filesystem
identifier = f'{model_cls.__name__}_{start_time}'
# Train each combination per seed
combination_path = study_root_path / observation_mode / env_name / identifier
combination_path = study_root_path / obs_mode / env_name / identifier
env_class, env_kwargs = env_map[env_name]
env_kwargs = env_kwargs.copy()
# Retrieve and set the observation mode specific env parameters
if observation_mode_kwargs := observation_modes.get(observation_mode, None):
if additional_env_kwargs := observation_mode_kwargs.get("additional_env_kwargs", None):
env_kwargs.update(additional_env_kwargs)
additional_kwargs = observation_modes.get(obs_mode, {}).get("additional_env_kwargs", {})
env_kwargs.update(additional_kwargs)
for seed in range(5):
env_kwargs.update(env_seed=seed)
# Output folder
@@ -320,7 +323,10 @@ if __name__ == '__main__':
gc.collect()
# Compare performance runs, for each seed within a model
compare_seed_runs(combination_path, use_tex=False)
try:
compare_seed_runs(combination_path, use_tex=False)
except ValueError:
pass
# Better be save then sorry: Clean up!
try:
del env_kwargs
@@ -332,8 +338,11 @@ if __name__ == '__main__':
# Compare performance runs, for each model
# FIXME: Check THIS!!!!
compare_model_runs(study_root_path / observation_mode / env_name, f'{start_time}', 'step_reward',
use_tex=False)
try:
compare_model_runs(study_root_path / obs_mode / env_name, f'{start_time}', 'step_reward',
use_tex=False)
except ValueError:
pass
pass
pass
pass
@@ -343,8 +352,8 @@ if __name__ == '__main__':
# Evaluation starts here #####################################################
# First Iterate over every model and monitor "as trained"
if True:
for observation_mode in observation_modes:
obs_mode_path = next(x for x in study_root_path.iterdir() if x.is_dir() and x.name == observation_mode)
for obs_mode in observation_modes:
obs_mode_path = next(x for x in study_root_path.iterdir() if x.is_dir() and x.name == obs_mode)
# For trained policy in study_root_path / identifier
for env_path in [x for x in obs_mode_path.iterdir() if x.is_dir()]:
for policy_path in [x for x in env_path.iterdir() if x. is_dir()]:
@@ -364,8 +373,8 @@ if __name__ == '__main__':
# Then iterate over every model and monitor "ood behavior" - "is it ood?"
if True:
for observation_mode in observation_modes:
obs_mode_path = next(x for x in study_root_path.iterdir() if x.is_dir() and x.name == observation_mode)
for obs_mode in observation_modes:
obs_mode_path = next(x for x in study_root_path.iterdir() if x.is_dir() and x.name == obs_mode)
# For trained policy in study_root_path / identifier
for env_path in [x for x in obs_mode_path.iterdir() if x.is_dir()]:
for policy_path in [x for x in env_path.iterdir() if x. is_dir()]:
@@ -381,7 +390,7 @@ if __name__ == '__main__':
result = pool.starmap(load_model_run_study,
it.product(paths,
(env_map[env_path.name][0],),
(observation_modes[observation_mode],))
(observation_modes[obs_mode],))
)
# for seed_path in (y for y in policy_path.iterdir() if y.is_dir()):
# load_model_run_study(seed_path)