Thomas Knoll
11 months ago
8 changed files with 420 additions and 182 deletions
-
46examples/shields/rl/11_minigridrl.py
-
33examples/shields/rl/12_minigridrl_tune.py
-
23examples/shields/rl/13_minigridsb.py
-
34examples/shields/rl/14_train_eval.py
-
65examples/shields/rl/15_train_eval_tune.py
-
209examples/shields/rl/rllibutils.py
-
68examples/shields/rl/sb3utils.py
-
124examples/shields/rl/utils.py
@ -0,0 +1,209 @@ |
|||
import gymnasium as gym |
|||
import numpy as np |
|||
import random |
|||
|
|||
from minigrid.core.actions import Actions |
|||
from minigrid.core.constants import COLORS, OBJECT_TO_IDX, STATE_TO_IDX |
|||
|
|||
from gymnasium.spaces import Dict, Box |
|||
from collections import deque |
|||
from ray.rllib.utils.numpy import one_hot |
|||
|
|||
from helpers import get_action_index_mapping |
|||
from shieldhandlers import ShieldHandler |
|||
|
|||
|
|||
class OneHotShieldingWrapper(gym.core.ObservationWrapper): |
|||
def __init__(self, env, vector_index, framestack): |
|||
super().__init__(env) |
|||
self.framestack = framestack |
|||
# 49=7x7 field of vision; 16=object types; 6=colors; 3=state types. |
|||
# +4: Direction. |
|||
self.single_frame_dim = 49 * (len(OBJECT_TO_IDX) + len(COLORS) + len(STATE_TO_IDX)) + 4 |
|||
self.init_x = None |
|||
self.init_y = None |
|||
self.x_positions = [] |
|||
self.y_positions = [] |
|||
self.x_y_delta_buffer = deque(maxlen=100) |
|||
self.vector_index = vector_index |
|||
self.frame_buffer = deque(maxlen=self.framestack) |
|||
for _ in range(self.framestack): |
|||
self.frame_buffer.append(np.zeros((self.single_frame_dim,))) |
|||
|
|||
self.observation_space = Dict( |
|||
{ |
|||
"data": gym.spaces.Box(0.0, 1.0, shape=(self.single_frame_dim * self.framestack,), dtype=np.float32), |
|||
"action_mask": gym.spaces.Box(0, 10, shape=(env.action_space.n,), dtype=int), |
|||
} |
|||
) |
|||
|
|||
def observation(self, obs): |
|||
# Debug output: max-x/y positions to watch exploration progress. |
|||
# print(F"Initial observation in Wrapper {obs}") |
|||
if self.step_count == 0: |
|||
for _ in range(self.framestack): |
|||
self.frame_buffer.append(np.zeros((self.single_frame_dim,))) |
|||
if self.vector_index == 0: |
|||
if self.x_positions: |
|||
max_diff = max( |
|||
np.sqrt( |
|||
(np.array(self.x_positions) - self.init_x) ** 2 |
|||
+ (np.array(self.y_positions) - self.init_y) ** 2 |
|||
) |
|||
) |
|||
self.x_y_delta_buffer.append(max_diff) |
|||
print( |
|||
"100-average dist travelled={}".format( |
|||
np.mean(self.x_y_delta_buffer) |
|||
) |
|||
) |
|||
self.x_positions = [] |
|||
self.y_positions = [] |
|||
self.init_x = self.agent_pos[0] |
|||
self.init_y = self.agent_pos[1] |
|||
|
|||
|
|||
self.x_positions.append(self.agent_pos[0]) |
|||
self.y_positions.append(self.agent_pos[1]) |
|||
|
|||
image = obs["data"] |
|||
# One-hot the last dim into 16, 6, 3 one-hot vectors, then flatten. |
|||
objects = one_hot(image[:, :, 0], depth=len(OBJECT_TO_IDX)) |
|||
colors = one_hot(image[:, :, 1], depth=len(COLORS)) |
|||
states = one_hot(image[:, :, 2], depth=len(STATE_TO_IDX)) |
|||
|
|||
all_ = np.concatenate([objects, colors, states], -1) |
|||
all_flat = np.reshape(all_, (-1,)) |
|||
direction = one_hot(np.array(self.agent_dir), depth=4).astype(np.float32) |
|||
single_frame = np.concatenate([all_flat, direction]) |
|||
self.frame_buffer.append(single_frame) |
|||
|
|||
tmp = {"data": np.concatenate(self.frame_buffer), "action_mask": obs["action_mask"] } |
|||
return tmp |
|||
|
|||
|
|||
class MiniGridShieldingWrapper(gym.core.Wrapper): |
|||
def __init__(self, |
|||
env, |
|||
shield_creator : ShieldHandler, |
|||
shield_query_creator, |
|||
create_shield_at_reset=True, |
|||
mask_actions=True): |
|||
super(MiniGridShieldingWrapper, self).__init__(env) |
|||
self.max_available_actions = env.action_space.n |
|||
self.observation_space = Dict( |
|||
{ |
|||
"data": env.observation_space.spaces["image"], |
|||
"action_mask" : Box(0, 10, shape=(self.max_available_actions,), dtype=np.int8), |
|||
} |
|||
) |
|||
self.shield_creator = shield_creator |
|||
self.create_shield_at_reset = create_shield_at_reset |
|||
self.shield = shield_creator.create_shield(env=self.env) |
|||
self.mask_actions = mask_actions |
|||
self.shield_query_creator = shield_query_creator |
|||
print(F"Shielding is {self.mask_actions}") |
|||
|
|||
def create_action_mask(self): |
|||
if not self.mask_actions: |
|||
ret = np.array([1.0] * self.max_available_actions, dtype=np.int8) |
|||
return ret |
|||
|
|||
cur_pos_str = self.shield_query_creator(self.env) |
|||
|
|||
# Create the mask |
|||
# If shield restricts action mask only valid with 1.0 |
|||
# else set all actions as valid |
|||
allowed_actions = [] |
|||
mask = np.array([0.0] * self.max_available_actions, dtype=np.int8) |
|||
|
|||
if cur_pos_str in self.shield and self.shield[cur_pos_str]: |
|||
allowed_actions = self.shield[cur_pos_str] |
|||
zeroes = np.array([0.0] * len(allowed_actions), dtype=np.int8) |
|||
has_allowed_actions = False |
|||
|
|||
for allowed_action in allowed_actions: |
|||
index = get_action_index_mapping(allowed_action.labels) # Allowed_action is a set |
|||
if index is None: |
|||
assert(False) |
|||
|
|||
allowed = 1.0 |
|||
has_allowed_actions = True |
|||
mask[index] = allowed |
|||
else: |
|||
for index, x in enumerate(mask): |
|||
mask[index] = 1.0 |
|||
|
|||
front_tile = self.env.grid.get(self.env.front_pos[0], self.env.front_pos[1]) |
|||
|
|||
if front_tile is not None and front_tile.type == "key": |
|||
mask[Actions.pickup] = 1.0 |
|||
|
|||
|
|||
if front_tile and front_tile.type == "door": |
|||
mask[Actions.toggle] = 1.0 |
|||
# print(F"Mask is {mask} State: {cur_pos_str}") |
|||
return mask |
|||
|
|||
def reset(self, *, seed=None, options=None): |
|||
obs, infos = self.env.reset(seed=seed, options=options) |
|||
|
|||
if self.create_shield_at_reset and self.mask_actions: |
|||
self.shield = self.shield_creator.create_shield(env=self.env) |
|||
|
|||
mask = self.create_action_mask() |
|||
return { |
|||
"data": obs["image"], |
|||
"action_mask": mask |
|||
}, infos |
|||
|
|||
def step(self, action): |
|||
orig_obs, rew, done, truncated, info = self.env.step(action) |
|||
|
|||
mask = self.create_action_mask() |
|||
obs = { |
|||
"data": orig_obs["image"], |
|||
"action_mask": mask, |
|||
} |
|||
|
|||
return obs, rew, done, truncated, info |
|||
|
|||
|
|||
def shielding_env_creater(config): |
|||
name = config.get("name", "MiniGrid-LavaCrossingS9N3-v0") |
|||
framestack = config.get("framestack", 4) |
|||
args = config.get("args", None) |
|||
args.grid_path = F"{args.expname}_{args.grid_path}_{config.worker_index}.txt" |
|||
args.prism_path = F"{args.expname}_{args.prism_path}_{config.worker_index}.prism" |
|||
shielding = config.get("shielding", False) |
|||
shield_creator = MiniGridShieldHandler(grid_file=args.grid_path, |
|||
grid_to_prism_path=args.grid_to_prism_binary_path, |
|||
prism_path=args.prism_path, |
|||
formula=args.formula, |
|||
shield_value=args.shield_value, |
|||
prism_config=args.prism_config, |
|||
shield_comparision=args.shield_comparision) |
|||
|
|||
probability_intended = args.probability_intended |
|||
probability_displacement = args.probability_displacement |
|||
|
|||
env = gym.make(name, randomize_start=True,probability_intended=probability_intended, probability_displacement=probability_displacement) |
|||
env = MiniGridShieldingWrapper(env, shield_creator=shield_creator, shield_query_creator=create_shield_query ,mask_actions=shielding != ShieldingConfig.Disabled) |
|||
|
|||
env = OneHotShieldingWrapper(env, |
|||
config.vector_index if hasattr(config, "vector_index") else 0, |
|||
framestack=framestack |
|||
) |
|||
|
|||
|
|||
return env |
|||
|
|||
|
|||
def register_minigrid_shielding_env(args): |
|||
env_name = "mini-grid-shielding" |
|||
register_env(env_name, shielding_env_creater) |
|||
|
|||
ModelCatalog.register_custom_model( |
|||
"shielding_model", |
|||
TorchActionMaskModel |
|||
) |
@ -0,0 +1,68 @@ |
|||
import gymnasium as gym |
|||
import numpy as np |
|||
import random |
|||
|
|||
class MiniGridSbShieldingWrapper(gym.core.Wrapper): |
|||
def __init__(self, |
|||
env, |
|||
shield_creator : ShieldHandler, |
|||
shield_query_creator, |
|||
create_shield_at_reset = True, |
|||
mask_actions=True, |
|||
): |
|||
super(MiniGridSbShieldingWrapper, self).__init__(env) |
|||
self.max_available_actions = env.action_space.n |
|||
self.observation_space = env.observation_space.spaces["image"] |
|||
|
|||
self.shield_creator = shield_creator |
|||
self.mask_actions = mask_actions |
|||
self.shield_query_creator = shield_query_creator |
|||
|
|||
def create_action_mask(self): |
|||
if not self.mask_actions: |
|||
return np.array([1.0] * self.max_available_actions, dtype=np.int8) |
|||
|
|||
cur_pos_str = self.shield_query_creator(self.env) |
|||
|
|||
allowed_actions = [] |
|||
|
|||
# Create the mask |
|||
# If shield restricts actions, mask only valid actions with 1.0 |
|||
# else set all actions valid |
|||
mask = np.array([0.0] * self.max_available_actions, dtype=np.int8) |
|||
|
|||
if cur_pos_str in self.shield and self.shield[cur_pos_str]: |
|||
allowed_actions = self.shield[cur_pos_str] |
|||
for allowed_action in allowed_actions: |
|||
index = get_action_index_mapping(allowed_action.labels) |
|||
if index is None: |
|||
assert(False) |
|||
|
|||
mask[index] = random.choices([0.0, 1.0], weights=(1 - allowed_action.prob, allowed_action.prob))[0] |
|||
else: |
|||
for index, x in enumerate(mask): |
|||
mask[index] = 1.0 |
|||
|
|||
front_tile = self.env.grid.get(self.env.front_pos[0], self.env.front_pos[1]) |
|||
|
|||
|
|||
if front_tile and front_tile.type == "door": |
|||
mask[Actions.toggle] = 1.0 |
|||
|
|||
return mask |
|||
|
|||
|
|||
def reset(self, *, seed=None, options=None): |
|||
obs, infos = self.env.reset(seed=seed, options=options) |
|||
|
|||
shield = self.shield_creator.create_shield(env=self.env) |
|||
|
|||
self.shield = shield |
|||
return obs["image"], infos |
|||
|
|||
def step(self, action): |
|||
orig_obs, rew, done, truncated, info = self.env.step(action) |
|||
obs = orig_obs["image"] |
|||
|
|||
return obs, rew, done, truncated, info |
|||
|
Write
Preview
Loading…
Cancel
Save
Reference in new issue