def env_creator(env_cls, env_config):
env = env_cls(**env_config)
if isinstance(env.observation_space, (Box, FloatBox)):
obs_min = env.observation_space.low
obs_max = env.observation_space.high
env = TransformObservation(env, lambda obs: 2 * (obs - obs_min) / (obs_max - obs_min) - 1)
elif isinstance(env.observation_space, IntBox):
obs_min = env.observation_space.low
obs_max = env.observation_space.high
env = TransformObservation(env, lambda obs: (obs - obs_min) / (obs_max - obs_min))
return env
def test_transform_observation(env_id):
affine_transform = lambda x: 3 * x + 2
env = gym.make(env_id)
wrapped_env = TransformObservation(gym.make(env_id),
lambda obs: affine_transform(obs))
obs = env.reset(seed=0)
wrapped_obs = wrapped_env.reset(seed=0)
assert np.allclose(wrapped_obs, affine_transform(obs))
action = env.action_space.sample()
obs, reward, done, _ = env.step(action)
wrapped_obs, wrapped_reward, wrapped_done, _ = wrapped_env.step(action)
assert np.allclose(wrapped_obs, affine_transform(obs))
assert np.allclose(wrapped_reward, reward)
assert wrapped_done == done
def make_atari_env(env_name):
env = gym.make(env_name)
env = FireReset(env)
env = AtariPreprocessing(env,
noop_max=30,
frame_skip=2,
screen_size=84,
terminal_on_life_loss=False,
grayscale_obs=True,
scale_obs=False)
env = PyTorchImageWrapper(env)
env = FrameStack(env, num_stack=4)
env = TransformObservation(env, f=np.array)
env = ConcatWrapper(env, axis=0)
env = TransformObservation(
env, f=lambda obs: np.asarray(obs, dtype=np.float32) / 255.0)
return env
def get_env(*args, **kwargs):
return GymEnvWrapper(
TransformObservation(env=FrameStack(
num_stack=4,
env=(gym_wrapper.GymFromDMEnv(
bsuite.load_and_record_to_csv(
bsuite_id=bsuite_id,
results_dir=results_dir,
overwrite=True,
)) if not gym_id else gym.make(gym_id))),
f=lambda lazy_frames: np.reshape(
np.stack(lazy_frames._frames), -1)))
def make_env(seed, max_episode_steps=1500):
env = WestWorld(
seed=seed,
room_size=2,
obs_width=64,
obs_height=64,
max_episode_steps=max_episode_steps,
decore_option=DecoreOption.PORTRAIT,
num_chars_on_wall=1,
)
env = TransformObservation(
env, f=lambda obs: np.asarray(obs, dtype=np.float32) / 255.0)
env = PyTorchImageWrapper(env)
return env
def create_env(env_name="SuperMarioBros-1-1-v0"):
env = gym_super_mario_bros.make(env_name)
# Restricts action space to only "right" and "jump + right"
env = JoypadSpace(env, [["right"], ["right", "A"]])
# Accumulates rewards every 4th frame
env = SkipFrame(env, skip=4)
# Transform RGB image to graycale, [240, 256]
env = GrayScaleObservation(env)
# Downsample to new size, [1, 84, 84]
env = ResizeObservation(env, shape=84)
# Add extra precision to np.array state
env = TransformObservation(env, f=lambda x: x / 255.)
# Squash 4 consecutive frames of the environment into a
# single observation point to feed to our learning model, [4, 84, 84]
env = FrameStack(env, num_stack=4)
return env
import datetime
import gym
import matplotlib.pyplot as plt
import numpy as np
from gym.wrappers import FrameStack, GrayScaleObservation, Monitor, TransformObservation
from torch.optim import Adam
import autonomous_systems_project.callbacks as cb
from autonomous_systems_project.agents import DoubleDQNAgent, DQNAgent, SimpleDQN
from autonomous_systems_project.memory import RandomReplayMemory
env = gym.make("CartPole-v1")
env = TransformObservation(env, lambda obs: np.array(obs).astype(np.float32))
policy_net = SimpleDQN(4, env.action_space.n)
target_net = SimpleDQN(4, env.action_space.n)
target_net.load_state_dict(policy_net.state_dict())
gamma = 0.9
batch_size = 32
epsilon_steps = 1000
memory_size = 100000
episodes = 500
lr = 0.001
target_update = 5
parameters = {
"env": env.unwrapped.spec.id,
"gamma": gamma,
"batch_size": batch_size,
from agent import Mario
from wrappers import ResizeObservation, SkipFrame
# Initialize Super Mario environment
env = gym_super_mario_bros.make('SuperMarioBros-1-1-v0')
# Limit the action-space to
# 0. walk right
# 1. jump right
env = JoypadSpace(env, [['right'], ['right', 'A']])
# Apply Wrappers to environment
env = SkipFrame(env, skip=4)
env = GrayScaleObservation(env, keep_dim=False)
env = ResizeObservation(env, shape=84)
env = TransformObservation(env, f=lambda x: x / 255.)
env = FrameStack(env, num_stack=4)
env.reset()
save_dir = Path('checkpoints') / datetime.datetime.now().strftime(
'%Y-%m-%dT%H-%M-%S')
save_dir.mkdir(parents=True)
checkpoint = None # Path('checkpoints/2020-10-21T18-25-27/mario.chkpt')
# Add in check to see if GPU is avaliable (BM)
if torch.cuda.is_available():
device = torch.device("cuda:0")
print("Using GPU!")
else: