def gen_avg_return(agent, env):
'''Generate average return for agent and an env'''
with util.ctx_lab_mode('eval'): # enter eval context
agent.algorithm.update() # set explore_var etc. to end_val under ctx
with torch.no_grad():
ret = gen_return(agent, env)
# exit eval context, restore variables simply by updating
agent.algorithm.update()
return ret
def gen_avg_return(agent, env, num_eval=NUM_EVAL):
'''Generate average return for agent and an env'''
with util.ctx_lab_mode('eval'): # enter eval context
agent.algorithm.update() # set explore_var etc. to end_val under ctx
with torch.no_grad():
returns = [gen_return(agent, env) for i in range(num_eval)]
# exit eval context, restore variables simply by updating
agent.algorithm.update()
return np.mean(returns)
def __init__(self, spec, global_nets=None):
self.spec = spec
self.index = self.spec['meta']['session']
util.set_random_seed(self.spec)
util.set_cuda_id(self.spec)
util.set_logger(self.spec, logger, 'session')
spec_util.save(spec, unit='session')
self.agent, self.env = make_agent_env(self.spec, global_nets)
with util.ctx_lab_mode('eval'): # env for eval
self.eval_env = make_env(self.spec)
logger.info(util.self_desc(self))
def run_eval_episode(self):
with util.ctx_lab_mode('eval'): # enter eval context
self.agent.algorithm.update(
) # set explore_var etc. to end_val under ctx
self.eval_env.clock.tick('epi')
logger.info(
f'Running eval episode for trial {self.info_space.get("trial")} session {self.index}'
)
total_reward = 0
reward, state, done = self.eval_env.reset()
while not done:
self.eval_env.clock.tick('t')
action = self.agent.act(state)
reward, state, done = self.eval_env.step(action)
total_reward += reward
# exit eval context, restore variables simply by updating
self.agent.algorithm.update()
# update body.eval_df
self.agent.body.eval_update(self.eval_env, total_reward)
self.agent.body.log_summary(body_df_kind='eval')
def __init__(self, spec, info_space, global_nets=None):
self.spec = spec
self.info_space = info_space
self.index = self.info_space.get('session')
util.set_logger(self.spec, self.info_space, logger, 'session')
self.data = None
# init singleton agent and env
self.env = make_env(self.spec)
util.set_rand_seed(self.info_space.get_random_seed(), self.env)
with util.ctx_lab_mode('eval'): # env for eval
self.eval_env = make_env(self.spec)
util.set_rand_seed(self.info_space.get_random_seed(),
self.eval_env)
util.try_set_cuda_id(self.spec, self.info_space)
body = Body(self.env, self.spec['agent'])
self.agent = Agent(self.spec,
self.info_space,
body=body,
global_nets=global_nets)
enable_aeb_space(self) # to use lab's data analysis framework
logger.info(util.self_desc(self))
logger.info(f'Initialized session {self.index}')