def make_vec_env(self,dataset, env_args):
env_args["df"]= dataset
env = make_vec_env('crypt-v001', env_kwargs=env_args)
env = VecCheckNan(env, raise_exception=True)
env = VecNormalize(
env, norm_obs=True, norm_reward=False, clip_obs=10.0, gamma=0.95
)
return env
def make_dummy_env(self, dataset, env_args):
env = gym.make("crypt-v001", df=dataset, **env_args)
check_env(env)
env = DummyVecEnv([lambda: env])
env = VecCheckNan(env, raise_exception=True)
env = VecNormalize(
env, norm_obs=True, norm_reward=False, clip_obs=10.0, gamma=0.95
)
return env
def test_check_nan():
"""Test VecCheckNan Object"""
env = DummyVecEnv([NanAndInfEnv])
env = VecCheckNan(env, raise_exception=True)
env.step([[0]])
with pytest.raises(ValueError):
env.step([[float('NaN')]])
with pytest.raises(ValueError):
env.step([[float('inf')]])
with pytest.raises(ValueError):
env.step([[-1]])
with pytest.raises(ValueError):
env.step([[1]])
env.step(np.array([[0, 1], [0, 1]]))
env.reset()
def test_check_nan():
"""Test VecCheckNan Object"""
env = DummyVecEnv([NanAndInfEnv])
env = VecCheckNan(env, raise_exception=True)
env.step([[0]])
try:
env.step([[float('NaN')]])
except ValueError:
pass
else:
assert False
try:
env.step([[float('inf')]])
except ValueError:
pass
else:
assert False
try:
env.step([[-1]])
except ValueError:
pass
else:
assert False
try:
env.step([[1]])
except ValueError:
pass
else:
assert False
env.step(np.array([[0, 1], [0, 1]]))
def _check_nan(env: gym.Env) -> None:
"""Check for Inf and NaN using the VecWrapper."""
vec_env = VecCheckNan(DummyVecEnv([lambda: env]))
for _ in range(10):
action = np.array([env.action_space.sample()])
_, _, _, _ = vec_env.step(action)
def make_env(
args,
num_envs=None,
include_norm=False,
norm_reward=True,
**kwargs,
):
"""Return a vectorized environment containing `num_envs` or `args.num_envs`
environments (depending on whether `num_envs is None`).
`args`, the command line arguments, specify several values. See `kwargs`
for a more detailed explanation on their interaction.
`include_norm` specifies whether the environment is wrapped in a
normalizing environment.
`norm_reward` indicates whether the rewards are normalized (only
revelant if `include_norm is True`).
`kwargs` are passed directly to the environment creation function. Any
value given via `kwargs` has priority over the one given by `args`.
"""
if num_envs is None:
num_envs = args.num_envs
# `kwargs` given via `args`
args_kwargs = {}
for arg in [
'M',
'dt',
'restol',
'lambda_real_interval',
'lambda_imag_interval',
'lambda_real_interpolation_interval',
'norm_factor',
'residual_weight',
'step_penalty',
'reward_iteration_only',
'reward_strategy',
'collect_states',
'example',
]:
args_kwargs[arg] = kwargs.pop(arg, getattr(args, arg))
all_kwargs = {**kwargs, **args_kwargs}
# SAC does not support float64
if args.model_class == 'SAC':
all_kwargs['use_doubles'] = False
seed = all_kwargs.pop('seed', args.seed)
def gym_make(i):
return lambda: gym.make(
args.envname,
seed=seed + i if seed is not None else None,
**all_kwargs,
)
env = DummyVecEnv([gym_make(i) for i in range(num_envs)])
if include_norm:
if hasattr(args, 'env_path') and args.env_path is not None:
env = VecNormalize.load(str(Path(args.env_path)), env)
else:
# When training, set `norm_reward = True`, I hear...
if 'gamma' in args.model_kwargs:
env = VecNormalize(
env,
norm_obs=args.norm_obs,
norm_reward=norm_reward,
gamma=args.model_kwargs['gamma'],
)
else:
env = VecNormalize(
env,
norm_obs=args.norm_obs,
norm_reward=norm_reward,
)
if debug_nans:
env = VecCheckNan(env, raise_exception=True)
return env
game = "SuperMarioKart-Snes"
scenario = "C:\\Projects\\OpenAI Games\\retro-ai-hacking\\scenarios\\SuperMarioKart-Snes\\custom_rewards.json"
state = "C:\\Users\\joncocks\\anaconda3\\envs\\retro_ai_3\\Lib\\site-packages\\retro\\data\\contrib\\SuperMarioKart-Snes\\MarioCircuit1.GP.50cc.1P.Luigi.Start.state"
# game = "Fzero-Snes"
# scenario = "C:\\Users\\joncocks\\anaconda3\\envs\\retro_ai_3\\Lib\\site-packages\\retro\\data\\contrib\\Fzero-Snes\\scenario.json"
# state = "C:\\Users\\joncocks\\anaconda3\\envs\\retro_ai_3\\Lib\\site-packages\\retro\\data\\contrib\\Fzero-Snes\\practice.mutecity.bluefalcon.norival.start.state"
experiment_id = str(uuid4())
n_cpus = 8
env = SubprocVecEnv(
[lambda: get_env(game, state, scenario) for i in range(n_cpus)])
# env = DummyVecEnv([lambda: get_env(game, state, scenario)])
# env = VecNormalize(env, norm_obs=True, norm_reward=False)
env = VecCheckNan(env, raise_exception=True)
# Create a callback to save every n timesteps
prefix = "ppo_" + game + "_" + experiment_id
checkpoint_callback = CheckpointCallback(
save_freq=100000,
save_path="C:\\Projects\\OpenAI Games\\retro-ai-hacking\\models",
name_prefix=prefix)
savefile_name = prefix + "_final"
savefile_name = os.path.join(
"C:\\Projects\\OpenAI Games\\retro-ai-hacking\\models", savefile_name)
model = PPO(
CnnPolicy,