def __init__(self, args, mask=None, name='full'): env = gym.make(args.task) if args.task == 'Pendulum-v0': env.spec.reward_threshold = -250 self.state_shape = env.observation_space.shape or env.observation_space.n self.action_shape = env.action_space.shape or env.action_space.n self.max_action = env.action_space.high[0] self.stop_fn = lambda x: x >= env.spec.reward_threshold # env self.train_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) self.test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # mask state_dim = int(np.prod(self.state_shape)) self._view_mask = torch.ones(state_dim) if mask == 'even': for i in range(0, state_dim, 2): self._view_mask[i] = 0 elif mask == "odd": for i in range(1, state_dim, 2): self._view_mask[i] = 0 elif type(mask) == int: self._view_mask[mask] = 0 # policy self.actor = ActorProbWithView( args.layer_num, self.state_shape, self.action_shape, self.max_action, self._view_mask, args.device ).to(args.device) self.actor_optim = torch.optim.Adam(self.actor.parameters(), lr=args.actor_lr) self.critic1 = CriticWithView( args.layer_num, self.state_shape, self._view_mask, self.action_shape, args.device ).to(args.device) self.critic1_optim = torch.optim.Adam(self.critic1.parameters(), lr=args.critic_lr) self.critic2 = CriticWithView( args.layer_num, self.state_shape, self._view_mask, self.action_shape, args.device ).to(args.device) self.critic2_optim = torch.optim.Adam(self.critic2.parameters(), lr=args.critic_lr) self.policy = SACPolicy( self.actor, self.actor_optim, self.critic1, self.critic1_optim, self.critic2, self.critic2_optim, args.tau, args.gamma, args.alpha, [env.action_space.low[0], env.action_space.high[0]], reward_normalization=True, ignore_done=True) # collector self.train_collector = Collector(self.policy, self.train_envs, ReplayBuffer(args.buffer_size)) self.test_collector = Collector(self.policy, self.test_envs) # log self.writer = SummaryWriter(f"{args.logdir}/{args.task}/sac/{args.note}/{name}")
def test_sac(args=get_args()): env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) print("Action range:", np.min(env.action_space.low), np.max(env.action_space.high)) # train_envs = gym.make(args.task) if args.training_num > 1: train_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)] ) else: train_envs = gym.make(args.task) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)] ) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net_a = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb( net_a, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True, conditioned_sigma=True ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device ) net_c2 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device ) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, tau=args.tau, gamma=args.gamma, alpha=args.alpha, estimation_step=args.n_step, action_space=env.action_space ) # load a previous policy if args.resume_path: policy.load_state_dict(torch.load(args.resume_path, map_location=args.device)) print("Loaded agent from: ", args.resume_path) # collector if args.training_num > 1: buffer = VectorReplayBuffer(args.buffer_size, len(train_envs)) else: buffer = ReplayBuffer(args.buffer_size) train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs) train_collector.collect(n_step=args.start_timesteps, random=True) # log t0 = datetime.datetime.now().strftime("%m%d_%H%M%S") log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_sac' log_path = os.path.join(args.logdir, args.task, 'sac', log_file) writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = TensorboardLogger(writer) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) if not args.watch: # trainer result = offpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False ) pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) print(f'Final reward: {result["rews"].mean()}, length: {result["lens"].mean()}')
def test_sac(args=get_args()): torch.set_num_threads(1) # we just need only one thread for NN env = gym.make(args.task) if args.task == 'Pendulum-v0': env.spec.reward_threshold = -250 args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) train_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model actor = ActorProb(args.layer_num, args.state_shape, args.action_shape, args.max_action, args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) critic1 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) policy = SACPolicy(actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, args.tau, args.gamma, args.alpha, [env.action_space.low[0], env.action_space.high[0]], reward_normalization=True, ignore_done=True) # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(x): return x >= env.spec.reward_threshold # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer) assert stop_fn(result['best_reward']) train_collector.close() test_collector.close() if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) collector = Collector(policy, env) result = collector.collect(n_episode=1, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}') collector.close()
def test_sac_with_il(args=get_args()): torch.set_num_threads(1) # we just need only one thread for NN env = gym.make(args.task) if args.task == 'Pendulum-v0': env.spec.reward_threshold = -250 args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) train_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb(net, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, action_range=[env.action_space.low[0], env.action_space.high[0]], tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=args.rew_norm, estimation_step=args.n_step) # collector train_collector = Collector(policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) logger = BasicLogger(writer) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return mean_rewards >= env.spec.reward_threshold # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, update_per_step=args.update_per_step, stop_fn=stop_fn, save_fn=save_fn, logger=logger) assert stop_fn(result['best_reward']) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) policy.eval() collector = Collector(policy, env) result = collector.collect(n_episode=1, render=args.render) rews, lens = result["rews"], result["lens"] print(f"Final reward: {rews.mean()}, length: {lens.mean()}") # here we define an imitation collector with a trivial policy policy.eval() if args.task == 'Pendulum-v0': env.spec.reward_threshold = -300 # lower the goal net = Actor(Net(args.state_shape, hidden_sizes=args.imitation_hidden_sizes, device=args.device), args.action_shape, max_action=args.max_action, device=args.device).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.il_lr) il_policy = ImitationPolicy(net, optim, mode='continuous') il_test_collector = Collector( il_policy, DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)])) train_collector.reset() result = offpolicy_trainer(il_policy, train_collector, il_test_collector, args.epoch, args.il_step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, logger=logger) assert stop_fn(result['best_reward']) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) il_policy.eval() collector = Collector(il_policy, env) result = collector.collect(n_episode=1, render=args.render) rews, lens = result["rews"], result["lens"] print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_sac_with_il(args=get_args()): # if you want to use python vector env, please refer to other test scripts train_envs = env = envpool.make_gym( args.task, num_envs=args.training_num, seed=args.seed ) test_envs = envpool.make_gym(args.task, num_envs=args.test_num, seed=args.seed) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] if args.reward_threshold is None: default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250} args.reward_threshold = default_reward_threshold.get( args.task, env.spec.reward_threshold ) # you can also use tianshou.env.SubprocVectorEnv # seed np.random.seed(args.seed) torch.manual_seed(args.seed) # model net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb( net, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device ) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device ) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=args.rew_norm, estimation_step=args.n_step, action_space=env.action_space ) # collector train_collector = Collector( policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True ) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) logger = TensorboardLogger(writer) def save_best_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return mean_rewards >= args.reward_threshold # trainer result = offpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, update_per_step=args.update_per_step, stop_fn=stop_fn, save_best_fn=save_best_fn, logger=logger ) assert stop_fn(result['best_reward']) # here we define an imitation collector with a trivial policy policy.eval() if args.task.startswith("Pendulum"): args.reward_threshold -= 50 # lower the goal net = Actor( Net( args.state_shape, hidden_sizes=args.imitation_hidden_sizes, device=args.device ), args.action_shape, max_action=args.max_action, device=args.device ).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.il_lr) il_policy = ImitationPolicy( net, optim, action_space=env.action_space, action_scaling=True, action_bound_method="clip" ) il_test_collector = Collector( il_policy, envpool.make_gym(args.task, num_envs=args.test_num, seed=args.seed), ) train_collector.reset() result = offpolicy_trainer( il_policy, train_collector, il_test_collector, args.epoch, args.il_step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, stop_fn=stop_fn, save_best_fn=save_best_fn, logger=logger ) assert stop_fn(result['best_reward'])
def test_sac(): args, log_path, writer = get_args() env = gym.make(args.task) if args.task == 'Pendulum-v0': env.spec.reward_threshold = -250 args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) train_envs = ShmPipeVecEnv([ lambda: TransformReward(BipedalWrapper(gym.make(args.task)), lambda reward: 5 * reward) for _ in range(args.training_num) ]) # test_envs = gym.make(args.task) test_envs = ShmPipeVecEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed + 1) # model actor = ActorProb(args.layer_num, args.state_shape, args.action_shape, args.max_action, args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) critic = DQCritic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic_target = DQCritic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr) policy = SACPolicy(actor, actor_optim, critic, critic_optim, critic_target, env.action_space, args.device, args.tau, args.gamma, args.alpha, reward_normalization=args.rew_norm, ignore_done=False) if args.mode == 'test': policy.load_state_dict( torch.load("{}/{}/{}/policy.pth".format(args.logdir, args.task, args.comment), map_location=args.device)) env = gym.make(args.task) collector = Collector(policy, env # Monitor(env, 'video', force=True) ) result = collector.collect(n_episode=10, render=args.render) print( f'Final reward: {result["ep/reward"]}, length: {result["ep/len"]}') collector.close() exit() # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) train_collector.collect(10000, sampling=True) test_collector = Collector(policy, test_envs) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(x): return x >= env.spec.reward_threshold + 5 # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_episode, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer) assert stop_fn(result['best_reward']) pprint.pprint(result)
def test_sac(args=get_args()): env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # train_envs = gym.make(args.task) train_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb(net, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, action_range=[env.action_space.low[0], env.action_space.high[0]], tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=args.rew_norm, exploration_noise=OUNoise(0.0, args.noise_std)) # collector train_collector = Collector(policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) logger = BasicLogger(writer) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return mean_rewards >= env.spec.reward_threshold # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, update_per_step=args.update_per_step, stop_fn=stop_fn, save_fn=save_fn, logger=logger) assert stop_fn(result['best_reward']) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) rews, lens = result["rews"], result["lens"] print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_sac_bipedal(args=get_args()): env = EnvWrapper(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] train_envs = SubprocVectorEnv( [lambda: EnvWrapper(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv([lambda: EnvWrapper(args.task, reward_scale=1) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net_a = Net(args.layer_num, args.state_shape, device=args.device) actor = ActorProb( net_a, args.action_shape, args.max_action, args.device, unbounded=True ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic1 = Critic(net_c1, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic2 = Critic(net_c2, args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, action_range=[env.action_space.low[0], env.action_space.high[0]], tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=args.rew_norm, ignore_done=args.ignore_done, estimation_step=args.n_step) # load a previous policy if args.resume_path: policy.load_state_dict(torch.load(args.resume_path)) print("Loaded agent from: ", args.resume_path) # collector train_collector = Collector( policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return mean_rewards >= env.spec.reward_threshold # trainer result = offpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer, test_in_train=False) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=[1] * args.test_num, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_sac_bipedal(args=get_args()): torch.set_num_threads(1) # we just need only one thread for NN env = EnvWrapper(args.task) def IsStop(reward): return reward >= env.spec.reward_threshold args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] train_envs = SubprocVectorEnv( [lambda: EnvWrapper(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv([ lambda: EnvWrapper(args.task, reward_scale=1) for _ in range(args.test_num) ]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net_a = Net(args.layer_num, args.state_shape, device=args.device) actor = ActorProb(net_a, args.action_shape, args.max_action, args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic1 = Critic(net_c1, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic2 = Critic(net_c2, args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) policy = SACPolicy(actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, args.tau, args.gamma, args.alpha, [env.action_space.low[0], env.action_space.high[0]], reward_normalization=args.rew_norm, ignore_done=args.ignore_done, estimation_step=args.n_step) # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, stop_fn=IsStop, save_fn=save_fn, writer=writer) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=[1] * args.test_num, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def gather_data(): """Return expert buffer data.""" args = get_args() env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] if args.reward_threshold is None: default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250} args.reward_threshold = default_reward_threshold.get( args.task, env.spec.reward_threshold) # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) train_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = DummyVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb( net, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True, ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device, ) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device, ) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=args.rew_norm, estimation_step=args.n_step, action_space=env.action_space, ) # collector buffer = VectorReplayBuffer(args.buffer_size, len(train_envs)) train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) logger = TensorboardLogger(writer) def save_best_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return mean_rewards >= args.reward_threshold # trainer offpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, update_per_step=args.update_per_step, save_best_fn=save_best_fn, stop_fn=stop_fn, logger=logger, ) train_collector.reset() result = train_collector.collect(n_step=args.buffer_size) rews, lens = result["rews"], result["lens"] print(f"Final reward: {rews.mean()}, length: {lens.mean()}") if args.save_buffer_name.endswith(".hdf5"): buffer.save_hdf5(args.save_buffer_name) else: pickle.dump(buffer, open(args.save_buffer_name, "wb")) return buffer
def test_sac(args=get_args()): env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # train_envs = gym.make(args.task) train_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model actor = ActorProb(args.layer_num, args.state_shape, args.action_shape, args.max_action, args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) critic1 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) alpha = (target_entropy, log_alpha, alpha_optim) else: alpha = args.alpha policy = SACPolicy(actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, args.tau, args.gamma, alpha, [env.action_space.low[0], env.action_space.high[0]], reward_normalization=args.rew_norm, ignore_done=True, exploration_noise=OUNoise(0.0, args.noise_std)) # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(x): return x >= env.spec.reward_threshold # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer) assert stop_fn(result['best_reward']) train_collector.close() test_collector.close() if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) collector = Collector(policy, env) result = collector.collect(n_episode=1, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}') collector.close()
def test_sac(args=get_args()): torch.set_num_threads(1) env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # you can also use tianshou.env.SubprocVectorEnv # train_envs = gym.make(args.task) train_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb(net, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic1 = Critic(net, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic2 = Critic(net, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, action_range=[env.action_space.low[0], env.action_space.high[0]], tau=args.tau, gamma=args.gamma, alpha=args.alpha, reward_normalization=True, ignore_done=True) # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log log_path = os.path.join(args.logdir, args.task, 'sac', args.run_id) writer = SummaryWriter(log_path) def stop_fn(mean_rewards): return mean_rewards >= env.spec.reward_threshold # trainer result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, stop_fn=stop_fn, writer=writer, log_interval=args.log_interval) assert stop_fn(result['best_reward']) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=[1] * args.test_num, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_sac(args=get_args()): env, train_envs, test_envs = make_mujoco_env(args.task, args.seed, args.training_num, args.test_num, obs_norm=False) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) print("Action range:", np.min(env.action_space.low), np.max(env.action_space.high)) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) # model net_a = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = ActorProb( net_a, args.action_shape, max_action=args.max_action, device=args.device, unbounded=True, conditioned_sigma=True, ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device, ) net_c2 = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device, ) critic1 = Critic(net_c1, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(net_c2, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, tau=args.tau, gamma=args.gamma, alpha=args.alpha, estimation_step=args.n_step, action_space=env.action_space, ) # load a previous policy if args.resume_path: policy.load_state_dict( torch.load(args.resume_path, map_location=args.device)) print("Loaded agent from: ", args.resume_path) # collector if args.training_num > 1: buffer = VectorReplayBuffer(args.buffer_size, len(train_envs)) else: buffer = ReplayBuffer(args.buffer_size) train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs) train_collector.collect(n_step=args.start_timesteps, random=True) # log now = datetime.datetime.now().strftime("%y%m%d-%H%M%S") args.algo_name = "sac" log_name = os.path.join(args.task, args.algo_name, str(args.seed), now) log_path = os.path.join(args.logdir, log_name) # logger if args.logger == "wandb": logger = WandbLogger( save_interval=1, name=log_name.replace(os.path.sep, "__"), run_id=args.resume_id, config=args, project=args.wandb_project, ) writer = SummaryWriter(log_path) writer.add_text("args", str(args)) if args.logger == "tensorboard": logger = TensorboardLogger(writer) else: # wandb logger.load(writer) def save_best_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth")) if not args.watch: # trainer result = offpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.step_per_collect, args.test_num, args.batch_size, save_best_fn=save_best_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False, ) pprint.pprint(result) # Let's watch its performance! policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) print( f'Final reward: {result["rews"].mean()}, length: {result["lens"].mean()}' )
def test_sac(args=get_args()): env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) print("Action range:", np.min(env.action_space.low), np.max(env.action_space.high)) # train_envs = gym.make(args.task) train_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model net = Net(args.layer_num, args.state_shape, device=args.device, hidden_layer_size=args.hidden_layer_size) actor = ActorProb( net, args.action_shape, args.max_action, args.device, unbounded=True, hidden_layer_size=args.hidden_layer_size, conditioned_sigma=True, ).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device, hidden_layer_size=args.hidden_layer_size) critic1 = Critic(net_c1, args.device, hidden_layer_size=args.hidden_layer_size).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device, hidden_layer_size=args.hidden_layer_size) critic2 = Critic(net_c2, args.device, hidden_layer_size=args.hidden_layer_size).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) if args.auto_alpha: target_entropy = -np.prod(env.action_space.shape) log_alpha = torch.zeros(1, requires_grad=True, device=args.device) alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr) args.alpha = (target_entropy, log_alpha, alpha_optim) policy = SACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, action_range=[env.action_space.low[0], env.action_space.high[0]], tau=args.tau, gamma=args.gamma, alpha=args.alpha, estimation_step=args.n_step) # load a previous policy if args.resume_path: policy.load_state_dict( torch.load(args.resume_path, map_location=args.device)) print("Loaded agent from: ", args.resume_path) # collector train_collector = Collector(policy, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # log log_path = os.path.join(args.logdir, args.task, 'sac') writer = SummaryWriter(log_path) def watch(): # watch agent's performance print("Testing agent ...") policy.eval() test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=[1] * args.test_num, render=args.render) pprint.pprint(result) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return False if args.watch: watch() exit(0) # trainer train_collector.collect(n_step=args.pre_collect_step, random=True) result = offpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.test_num, args.batch_size, args.update_per_step, stop_fn=stop_fn, save_fn=save_fn, writer=writer, log_interval=args.log_interval) pprint.pprint(result) watch()
def test_sac(args=get_args()): env = gym.make(args.task) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n args.max_action = env.action_space.high[0] # train_envs = gym.make(args.task) train_envs = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # model actor = ActorProb(args.layer_num, args.state_shape, args.action_shape, args.max_action, args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) critic1 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) policy = SACPolicy(actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, args.tau, args.gamma, args.alpha, [env.action_space.low[0], env.action_space.high[0]], reward_normalization=True, ignore_done=True) # Load expert model. assert args.load is not None, 'args.load should not be None' expert = deepcopy(policy) expert.load_state_dict( torch.load(f'{args.logdir}/{args.task}/sac/{args.load}/policy.pth')) expert.eval() # collector expert_collector = Collector(expert, train_envs, ReplayBuffer(args.buffer_size)) test_collector = Collector(policy, test_envs) # train_collector.collect(n_step=args.buffer_size) # log writer = SummaryWriter(f'{args.logdir}/{args.task}/imitation/{args.note}') def stop_fn(x): return x >= (args.reward_threshold or env.spec.reward_threshold) def learner(pol, batch, batch_size, repeat, peer=0.): losses, peer_terms, ent_losses = [], [], [] for _ in range(repeat): for b in batch.split(batch_size): acts = pol(b).act demo = torch.tensor(b.act, dtype=torch.float) loss = F.mse_loss(acts, demo) if peer != 0: peer_demo = demo[torch.randperm(len(demo))] peer_term = peer * F.mse_loss(acts, peer_demo) loss -= peer_term peer_terms.append(peer_term.detach().cpu.numpy()) pol.actor_optim.zero_grad() loss.backward() pol.actor_optim.step() losses.append(loss.detach().cpu().numpy()) return { 'loss': losses, 'loss/ent': ent_losses, 'loss/peer': peer_terms if peer else None, 'peer': peer, } # trainer result = imitation_trainer(policy, learner, expert_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, 1, args.test_num, args.batch_size, stop_fn=stop_fn, writer=writer, task=args.task, peer=args.peer, peer_decay_steps=args.peer_decay_steps) assert stop_fn(result['best_reward']) expert_collector.close() test_collector.close() if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) collector = Collector(policy, env) result = collector.collect(n_episode=1, render=args.render) print(f'Final reward: {result["rew"]}, length: {result["len"]}') collector.close()