The source code and dockerfile for the GSW2024 AI Lab.
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
This repo is archived. You can view files and clone it, but cannot push or open issues/pull-requests.
 
 
 
 
 
 

131 lines
6.3 KiB

import gymnasium as gym
import minigrid
from ray.tune import register_env
from ray.rllib.algorithms.ppo import PPOConfig
from ray.rllib.algorithms.dqn.dqn import DQNConfig
from ray.tune.logger import pretty_print
from ray.rllib.models import ModelCatalog
from ray.rllib.algorithms.algorithm import Algorithm
from torch_action_mask_model import TorchActionMaskModel
from rllibutils import OneHotShieldingWrapper, MiniGridShieldingWrapper
from utils import parse_arguments, create_log_dir, ShieldingConfig
from utils import MiniGridShieldHandler, create_shield_query
from callbacks import CustomCallback
from ray.tune.logger import TBXLogger
import imageio
import os
import matplotlib.pyplot as plt
def shielding_env_creater(config):
name = config.get("name", "MiniGrid-LavaSlipperyS12-v2")
framestack = config.get("framestack", 4)
args = config.get("args", None)
args.grid_path = F"{args.grid_path}_{config.worker_index}.txt"
args.prism_path = F"{args.prism_path}_{config.worker_index}.prism"
shield_creator = MiniGridShieldHandler(args.grid_path, args.grid_to_prism_binary_path, args.prism_path, args.formula)
env = gym.make(name, randomize_start=False)
env = MiniGridShieldingWrapper(env, shield_creator=shield_creator, shield_query_creator=create_shield_query, mask_actions=False)
# env = minigrid.wrappers.ImgObsWrapper(env)
# env = ImgObsWrapper(env)
env = OneHotShieldingWrapper(env,
config.vector_index if hasattr(config, "vector_index") else 0,
framestack=framestack
)
env.randomize_start = False
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
)
import argparse
args = parse_arguments(argparse)
register_minigrid_shielding_env(args)
# Use the Algorithm's `from_checkpoint` utility to get a new algo instance
# that has the exact same state as the old one, from which the checkpoint was
# created in the first place:
# checkpoints = [('/home/knolli/Documents/University/Thesis/log_results/sh:none-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_low.yaml/checkpoint_000030', 'No_shield'),
# ("/home/knolli/Documents/University/Thesis/log_results/Relative_06/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_high.yaml/checkpoint_000030", "Rel_06_high"),
# ("/home/knolli/Documents/University/Thesis/log_results/Relative_06/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_medium.yaml/checkpoint_000030", "Rel_06_med"),
# ("/home/knolli/Documents/University/Thesis/log_results/Relative_06/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_low.yaml/checkpoint_000030", "Rel_06_low"),
# ("/home/knolli/Documents/University/Thesis/log_results/RELATIVE_1/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_high.yaml/checkpoint_000016", "Rel_1_high"),
# ("/home/knolli/Documents/University/Thesis/log_results/RELATIVE_1/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_medium.yaml/checkpoint_000030", "Rel_1_med"),
# ("/home/knolli/Documents/University/Thesis/log_results/RELATIVE_1/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:adv_config_slippery_low.yaml/checkpoint_000030", "Rel_1_low")]
checkpoints = [
# ('/home/knolli/Documents/University/Thesis/log_results/sh:none-value:0.9-env:MiniGrid-LavaSlipperyS12-v2-conf:slippery_high_pro.yaml/checkpoint_000070', "no_shielding"),
# ('/home/knolli/Documents/University/Thesis/log_results/sh:full-value:0.9-env:MiniGrid-LavaSlipperyS12-v2-conf:slippery_high_pro.yaml/checkpoint_000070', "shielding_09"),
# ('/home/knolli/Documents/University/Thesis/log_results/sh:full-value:1.0-env:MiniGrid-LavaSlipperyS12-v2-conf:slippery_high_pro.yaml/checkpoint_000070', "shielding_1")]
('/home/knolli/Documents/University/Thesis/logresults/exp/trial_0_2024-01-09_22-39-43/checkpoint_000002', 'v3')]
# checkpoints = [('/home/knolli/Documents/University/Thesis/log_results/sh:full-env:MiniGrid-LavaSlipperyS12-v2-conf:slippery_high_prob.yaml/checkpoint_000060', "Shielded_Gif")]
for path_to_checkpoint, gif_name in checkpoints:
algo = Algorithm.from_checkpoint(path_to_checkpoint)
policy = algo.get_policy()
# Continue training.
name = "MiniGrid-LavaSlipperyS12-v0"
shield_creator = MiniGridShieldHandler(F"./{args.grid_path}_1.txt", args.grid_to_prism_binary_path, F"./{args.prism_path}_1.prism", args.formula)
env = gym.make(name, randomize_start=False, probability_forward=3/9, probability_direct_neighbour=5/9, probability_next_neighbour=7/9,)
env = MiniGridShieldingWrapper(env, shield_creator=shield_creator, shield_query_creator=create_shield_query, mask_actions=True)
# env = minigrid.wrappers.ImgObsWrapper(env)
# env = ImgObsWrapper(env)
env = OneHotShieldingWrapper(env,
0,
framestack=4
)
episode_reward = 0
terminated = truncated = False
obs, info = env.reset()
i = 0
filenames = []
while not terminated and not truncated:
action = algo.compute_single_action(obs)
policy_actions = policy.compute_single_action(obs)
# print(f'Policy actions {policy_actions}')
# print(f'Policy actions {policy_actions.logits}')
policy_action = policy_actions[2]['action_dist_inputs'].argmax()
# print(f'The action is: {action} vs policy action {policy_action}')
if policy_action != action:
print('policy action deviated')
action = policy_action
obs, reward, terminated, truncated, info = env.step(action)
episode_reward += reward
filename = F"./frames/{i}.jpg"
img = env.get_frame()
plt.imsave(filename, img)
filenames.append(filename)
i = i + 1
import imageio
images = []
for filename in filenames:
images.append(imageio.imread(filename))
imageio.mimsave(F'./{gif_name}.gif', images)
for filename in filenames:
os.remove(filename)