warning_str = (
f"WARN: Inference could not be run for agent `{name}` as it was not "
f"found for the `{args.env_name}` environment.")
print(colorize(warning_str, "yellow"))
continue
# Get environment action and observation space dimensions
a_lowerbound = env.action_space.low
a_upperbound = env.action_space.high
s_dim = env.observation_space.shape[0]
a_dim = env.action_space.shape[0]
# Initiate the LAC policy
policy = LAC(a_dim,
s_dim,
act_limits={
"low": a_lowerbound,
"high": a_upperbound
})
# Retrieve all trained policies for a given agent
print("Looking for policies...")
policy_list = os.listdir(model_path)
policy_list = [
policy_name for policy_name in policy_list if osp.exists(
osp.abspath(
osp.join(model_path, policy_name, "policy/checkpoint")))
]
policy_list = [int(item) for item in policy_list if item.isnumeric()]
policy_list.sort()
# Check if a given policy exists for the current agent
# Check if model exists
if not os.path.exists(MODEL_PATH):
print(
f"Shutting down robustness eval since model `{args.model_name}` was "
f"not found for the `{args.env_name}` environment.")
sys.exit(0)
# Get environment action and observation space dimensions
a_lowerbound = env.action_space.low
a_upperbound = env.action_space.high
s_dim = env.observation_space.shape[0]
a_dim = env.action_space.shape[0]
# Create policy
policy = LAC(a_dim, s_dim)
# Retrieve agents
print("Looking for policies (rollouts)...")
rollout_list = os.listdir(MODEL_PATH)
rollout_list = [
rollout_name for rollout_name in rollout_list if os.path.exists(
os.path.abspath(MODEL_PATH + "/" + rollout_name +
"/policy/checkpoint"))
]
rollout_list = [int(item) for item in rollout_list if item.isnumeric()]
rollout_list.sort() # Sort rollouts_list
# Check if model exists
if not rollout_list:
print(
def dynamic(policy_path, env_name, env_params, alg_params, eval_params):
"""Performs dynamic robustness evaluation.
Args:
policy_path (str): Log path.
env_name (str): The gym environment you want to use.
alg_params (dict): Dictionary containing the algorithm parameters.
"""
# Retrieve environment
env = get_env_from_name(env_name, ENV_SEED)
# Get trained policy
s_dim = env.observation_space.shape[0]
a_dim = env.action_space.shape[0]
policy = LAC(a_dim, s_dim)
# Configure logger
log_path = policy_path + "/eval/dynamic/" + eval_params[
"additional_description"]
eval_params.update({"magnitude": 0})
logger.configure(dir=log_path, format_strs=["csv"])
# Evaluate policy results
_, paths = evaluation(policy_path, env_name, env, env_params, eval_params,
policy)
max_len = 0
print(len(paths))
for path in paths["s"]:
path_length = len(path)
if path_length > max_len:
max_len = path_length
average_path = np.average(np.array(paths["s"]), axis=0)
std_path = np.std(np.array(paths["s"]), axis=0)
for i in range(max_len):
logger.logkv("average_path", average_path[i])
logger.logkv("std_path", std_path[i])
logger.logkv("reference", paths["reference"][0][i])
logger.dumpkvs()
if eval_params["directly_show"]:
fig = plt.figure(figsize=(9, 6))
ax = fig.add_subplot(111)
if eval_params["plot_average"]:
t = range(max_len)
ax.plot(t, average_path, color="red")
ax.fill_between(
t,
average_path - std_path,
average_path + std_path,
color="red",
alpha=0.1,
)
plt.show()
else:
for path in paths["s"]:
path_length = len(path)
print(path_length)
t = range(path_length)
path = np.array(path)
# Ex3_EKF
fig = plt.figure()
ax = fig.add_subplot(111)
ax.plot(t, path, color="red")
plt.show()
ax.plot(t, np.array(path), color="blue", label="0.1")
plt.show()