def evaluate(self):
"""Evaluate model."""
self.pi.eval()
misc.set_env_to_eval_mode(self.env)
# Eval policy
os.makedirs(os.path.join(self.logdir, 'eval'), exist_ok=True)
outfile = os.path.join(self.logdir, 'eval',
self.ckptr.format.format(self.t) + '.json')
stats = rl_evaluate(self.env, self.pi, self.eval_num_episodes, outfile,
self.device)
logger.add_scalar('eval/mean_episode_reward', stats['mean_reward'],
self.t, time.time())
logger.add_scalar('eval/mean_episode_length', stats['mean_length'],
self.t, time.time())
# Record policy
# os.makedirs(os.path.join(self.logdir, 'video'), exist_ok=True)
# outfile = os.path.join(self.logdir, 'video',
# self.ckptr.format.format(self.t) + '.mp4')
# rl_record(self.env, self.pi, self.record_num_episodes, outfile,
# self.device)
self.pi.train()
misc.set_env_to_train_mode(self.env)
def evaluate(self):
"""Evaluate."""
eval_env = VecEpsilonGreedy(VecFrameStack(self.env, self.frame_stack),
self.eval_eps)
self.qf.eval()
misc.set_env_to_eval_mode(eval_env)
# Eval policy
os.makedirs(os.path.join(self.logdir, 'eval'), exist_ok=True)
outfile = os.path.join(self.logdir, 'eval',
self.ckptr.format.format(self.t) + '.json')
stats = rl_evaluate(eval_env, self.qf, self.eval_num_episodes, outfile,
self.device)
logger.add_scalar('eval/mean_episode_reward', stats['mean_reward'],
self.t, time.time())
logger.add_scalar('eval/mean_episode_length', stats['mean_length'],
self.t, time.time())
# Record policy
os.makedirs(os.path.join(self.logdir, 'video'), exist_ok=True)
outfile = os.path.join(self.logdir, 'video',
self.ckptr.format.format(self.t) + '.mp4')
rl_record(eval_env, self.qf, self.record_num_episodes, outfile,
self.device)
self.qf.train()
misc.set_env_to_train_mode(self.env)
self.data_manager.manual_reset()