def policyCrossEval(log_dir,
task,
episode,
model_path,
num_timesteps=2000,
num_cpu=1,
seed=0):
train_args, algo_name, algo_class, srl_model_path, env_kwargs = loadConfigAndSetup(
log_dir)
env_kwargs = EnvsKwargs(task, env_kwargs)
OK = True
if (not OK):
# no latest model saved yet
return None, False
else:
pass
printGreen(
"Evaluation from the model saved at: {}, with evaluation time steps: {}"
.format(model_path, num_timesteps))
log_dir, environment, algo_args = createEnv(log_dir,
train_args,
algo_name,
algo_class,
env_kwargs,
num_cpu=num_cpu,
seed=seed)
reward = policyEval(environment, model_path, log_dir, algo_class,
algo_args, num_timesteps, num_cpu)
# Just a trick to save the episode number of the reward,but need a little bit more space to store
reward = np.append(episode, reward)
return reward, True
def policyCrossEval(log_dir,
task,
episode,
model_path,
num_timesteps=2000,
num_cpu=1):
"""
To do a cross evaluation for a certain policy for different tasks
A version of real time evaluation but with some bugs to fix
:param log_dir:
:param task:
:param episode:
:param model_path:
:param num_timesteps: How many timesteps to evaluate the policy
:param num_cpu:
:return:
"""
train_args, algo_name, algo_class, srl_model_path, env_kwargs = loadConfigAndSetup(
log_dir)
env_kwargs = EnvsKwargs(task, env_kwargs)
OK = True
if (not OK):
# no latest model saved yet
return None, False
else:
pass
printGreen(
"Evaluation from the model saved at: {}, with evaluation time steps: {}"
.format(model_path, num_timesteps))
log_dir, environment, algo_args = createEnv(log_dir,
train_args,
algo_name,
algo_class,
env_kwargs,
num_cpu=num_cpu)
reward = policyEval(environment, model_path, log_dir, algo_class,
algo_args, num_timesteps, num_cpu)
# Just a trick to save the episode number of the reward,but need a little bit more space to store
reward = np.append(episode, reward)
return reward, True