def test_sac_bipedal(args=get_args()): 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) 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')) # 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, 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(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.layer_num, args.state_shape, device=args.device) actor = ActorProb(net, 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 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic1 = Critic(net, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(net, 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=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(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) 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()): 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.layer_num, args.state_shape, device=args.device) actor = ActorProb(net, 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 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic1 = Critic(net, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(net, 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(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, 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_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.layer_num, args.state_shape, device=args.device) actor = ActorProb( net, 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 = Net(args.layer_num, args.state_shape, args.action_shape, concat=True, device=args.device) critic1 = Critic(net, args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) critic2 = Critic(net, 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, 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')) 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) 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) print(f'Final reward: {result["rew"]}, length: {result["len"]}') # 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(1, args.state_shape, device=args.device), args.action_shape, args.max_action, 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.step_per_epoch // 5, 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']) 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) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_sac_with_il(args=get_args()): 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_npg(args=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 = 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, activation=nn.Tanh, device=args.device) actor = ActorProb(net, args.action_shape, max_action=args.max_action, unbounded=True, device=args.device).to(args.device) critic = Critic(Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device, activation=nn.Tanh), device=args.device).to(args.device) # orthogonal initialization for m in list(actor.modules()) + list(critic.modules()): if isinstance(m, torch.nn.Linear): torch.nn.init.orthogonal_(m.weight) torch.nn.init.zeros_(m.bias) optim = torch.optim.Adam(set(actor.parameters()).union( critic.parameters()), lr=args.lr) # replace DiagGuassian with Independent(Normal) which is equivalent # pass *logits to be consistent with policy.forward def dist(*logits): return Independent(Normal(*logits), 1) policy = NPGPolicy(actor, critic, optim, dist, discount_factor=args.gamma, reward_normalization=args.rew_norm, advantage_normalization=args.norm_adv, gae_lambda=args.gae_lambda, action_space=env.action_space, optim_critic_iters=args.optim_critic_iters, actor_step_size=args.actor_step_size, deterministic_eval=True) # collector train_collector = Collector( policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs))) test_collector = Collector(policy, test_envs) # log log_path = os.path.join(args.logdir, args.task, 'npg') 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 = onpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size, step_per_collect=args.step_per_collect, 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()}")
ix, iy = load_appli_data() dx, dy = load_appld_data() x = np.concatenate([ix, dx]) y = np.concatenate([iy, dy]) xx, yy = preprocessing(x, y, 4) yy = action_to_output(yy) net = Net(training_config['num_layers'], state_shape, device=device, hidden_layer_size=training_config['hidden_size']) if config['section'] == 'SAC': actor = ActorProb( net, action_shape, 1, device, hidden_layer_size=training_config['hidden_size']).to(device) else: actor = Actor(net, action_shape, 1, device, hidden_layer_size=training_config['hidden_size']).to(device) optimizer = torch.optim.Adam(actor.parameters(), lr=0.000001) loss_func = torch.nn.MSELoss() xx = to_torch(xx, device=device, dtype=torch.float32) yy = to_torch(yy, device=device, dtype=torch.float32) idx = list(range(xx.shape[0]))
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 = 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.layer_num, args.state_shape, device=args.device) actor = ActorProb(net, 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.model == 'ODEGBM': model = ODEGBM(args).to(args.device) elif args.model == 'PriorGBM': model = PriorGBM(args).to(args.device) elif args.model == 'NODAE': model = NODAE(args).to(args.device) else: assert args.model == 'ODENet' model = ODENet(args).to(args.device) policy = SSACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, model, args, 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) # 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') if args.baseline: if not os.path.exists(log_path + '/baseline/'): os.makedirs(log_path + '/baseline/') writer = SummaryWriter(log_path + '/baseline') else: writer = SummaryWriter(log_path) def save_fn(policy): # torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) pass def stop_fn(mean_rewards): # return mean_rewards >= env.spec.reward_threshold return False # 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']) 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) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
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.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, 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) 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, 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, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True) test_collector = Collector(policy, test_envs) # log log_path = os.path.join( args.logdir, args.task, 'sac', 'seed_' + str(args.seed) + '_' + datetime.datetime.now().strftime('%m%d-%H%M%S')) writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = BasicLogger(writer, train_interval=args.log_interval) 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=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.step_per_collect, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step) pprint.pprint(result) watch()
def test_ppo(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.layer_num, args.state_shape, device=args.device) actor = ActorProb(net, args.action_shape, args.max_action, args.device).to(args.device) critic = Critic(Net(args.layer_num, args.state_shape, device=args.device), device=args.device).to(args.device) # orthogonal initialization for m in list(actor.modules()) + list(critic.modules()): if isinstance(m, torch.nn.Linear): torch.nn.init.orthogonal_(m.weight) torch.nn.init.zeros_(m.bias) optim = torch.optim.Adam(list(actor.parameters()) + list(critic.parameters()), lr=args.lr) dist = DiagGaussian policy = PPOPolicy( actor, critic, optim, dist, args.gamma, max_grad_norm=args.max_grad_norm, eps_clip=args.eps_clip, vf_coef=args.vf_coef, ent_coef=args.ent_coef, reward_normalization=args.rew_norm, # dual_clip=args.dual_clip, # dual clip cause monotonically increasing log_std :) value_clip=args.value_clip, # action_range=[env.action_space.low[0], env.action_space.high[0]],) # if clip the action, ppo would not converge :) gae_lambda=args.gae_lambda) # 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, 'ppo') 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 = onpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.collect_per_step, args.repeat_per_collect, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer) 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) print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_ppo(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)], norm_obs=True) # test_envs = gym.make(args.task) test_envs = SubprocVectorEnv( [lambda: gym.make(args.task) for _ in range(args.test_num)], norm_obs=True, obs_rms=train_envs.obs_rms, update_obs_rms=False) # 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, activation=nn.Tanh, device=args.device) actor = ActorProb(net_a, args.action_shape, max_action=args.max_action, unbounded=True, device=args.device).to(args.device) net_c = Net(args.state_shape, hidden_sizes=args.hidden_sizes, activation=nn.Tanh, device=args.device) critic = Critic(net_c, device=args.device).to(args.device) torch.nn.init.constant_(actor.sigma_param, -0.5) for m in list(actor.modules()) + list(critic.modules()): if isinstance(m, torch.nn.Linear): # orthogonal initialization torch.nn.init.orthogonal_(m.weight, gain=np.sqrt(2)) torch.nn.init.zeros_(m.bias) # do last policy layer scaling, this will make initial actions have (close to) # 0 mean and std, and will help boost performances, # see https://arxiv.org/abs/2006.05990, Fig.24 for details for m in actor.mu.modules(): if isinstance(m, torch.nn.Linear): torch.nn.init.zeros_(m.bias) m.weight.data.copy_(0.01 * m.weight.data) optim = torch.optim.Adam(set( actor.parameters()).union(critic.parameters()), lr=args.lr) lr_scheduler = None if args.lr_decay: # decay learning rate to 0 linearly max_update_num = np.ceil( args.step_per_epoch / args.step_per_collect) * args.epoch lr_scheduler = LambdaLR( optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num) def dist(*logits): return Independent(Normal(*logits), 1) policy = PPOPolicy(actor, critic, optim, dist, discount_factor=args.gamma, gae_lambda=args.gae_lambda, max_grad_norm=args.max_grad_norm, vf_coef=args.vf_coef, ent_coef=args.ent_coef, reward_normalization=args.rew_norm, action_scaling=True, action_bound_method=args.bound_action_method, lr_scheduler=lr_scheduler, action_space=env.action_space, eps_clip=args.eps_clip, value_clip=args.value_clip, dual_clip=args.dual_clip, advantage_normalization=args.norm_adv, recompute_advantage=args.recompute_adv) # 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) # log t0 = datetime.datetime.now().strftime("%m%d_%H%M%S") log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_ppo' log_path = os.path.join(args.logdir, args.task, 'ppo', log_file) writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = BasicLogger(writer, update_interval=100, train_interval=100) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) if not args.watch: # trainer result = onpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size, step_per_collect=args.step_per_collect, save_fn=save_fn, logger=logger, 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_ppo(args=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 = 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).to(args.device) critic = Critic(Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device), device=args.device).to(args.device) # orthogonal initialization for m in list(actor.modules()) + list(critic.modules()): if isinstance(m, torch.nn.Linear): torch.nn.init.orthogonal_(m.weight) torch.nn.init.zeros_(m.bias) optim = torch.optim.Adam(list(actor.parameters()) + list(critic.parameters()), lr=args.lr) # replace DiagGuassian with Independent(Normal) which is equivalent # pass *logits to be consistent with policy.forward def dist(*logits): return Independent(Normal(*logits), 1) policy = PPOPolicy( actor, critic, optim, dist, discount_factor=args.gamma, max_grad_norm=args.max_grad_norm, eps_clip=args.eps_clip, vf_coef=args.vf_coef, ent_coef=args.ent_coef, reward_normalization=args.rew_norm, advantage_normalization=args.norm_adv, recompute_advantage=args.recompute_adv, # dual_clip=args.dual_clip, # dual clip cause monotonically increasing log_std :) value_clip=args.value_clip, gae_lambda=args.gae_lambda, action_space=env.action_space) # collector train_collector = Collector( policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs))) test_collector = Collector(policy, test_envs) # log log_path = os.path.join(args.logdir, args.task, 'ppo') writer = SummaryWriter(log_path) logger = BasicLogger(writer, save_interval=args.save_interval) 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 def save_checkpoint_fn(epoch, env_step, gradient_step): # see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html torch.save( { 'model': policy.state_dict(), 'optim': optim.state_dict(), }, os.path.join(log_path, 'checkpoint.pth')) if args.resume: # load from existing checkpoint print(f"Loading agent under {log_path}") ckpt_path = os.path.join(log_path, 'checkpoint.pth') if os.path.exists(ckpt_path): checkpoint = torch.load(ckpt_path, map_location=args.device) policy.load_state_dict(checkpoint['model']) optim.load_state_dict(checkpoint['optim']) print("Successfully restore policy and optim.") else: print("Fail to restore policy and optim.") # trainer result = onpolicy_trainer(policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size, episode_per_collect=args.episode_per_collect, stop_fn=stop_fn, save_fn=save_fn, logger=logger, resume_from_log=args.resume, save_checkpoint_fn=save_checkpoint_fn) 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()}")
def test_cql(args=get_args()): if os.path.exists(args.load_buffer_name) and os.path.isfile( args.load_buffer_name): if args.load_buffer_name.endswith(".hdf5"): buffer = VectorReplayBuffer.load_hdf5(args.load_buffer_name) else: buffer = pickle.load(open(args.load_buffer_name, "rb")) else: buffer = gather_data() 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] # float if args.reward_threshold is None: # too low? default_reward_threshold = {"Pendulum-v0": -1200, "Pendulum-v1": -1200} args.reward_threshold = default_reward_threshold.get( args.task, env.spec.reward_threshold) args.state_dim = args.state_shape[0] args.action_dim = args.action_shape[0] # 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) test_envs.seed(args.seed) # model # actor network net_a = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, ) actor = ActorProb( net_a, action_shape=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) # critic network 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 = CQLPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, cql_alpha_lr=args.cql_alpha_lr, cql_weight=args.cql_weight, tau=args.tau, gamma=args.gamma, alpha=args.alpha, temperature=args.temperature, with_lagrange=args.with_lagrange, lagrange_threshold=args.lagrange_threshold, min_action=np.min(env.action_space.low), max_action=np.max(env.action_space.high), device=args.device, ) # 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 # buffer has been gathered # train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs) # log t0 = datetime.datetime.now().strftime("%m%d_%H%M%S") log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_cql' log_path = os.path.join(args.logdir, args.task, 'cql', log_file) writer = SummaryWriter(log_path) writer.add_text("args", str(args)) 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 def watch(): policy.load_state_dict( torch.load(os.path.join(log_path, 'policy.pth'), map_location=torch.device('cpu'))) policy.eval() collector = Collector(policy, env) collector.collect(n_episode=1, render=1 / 35) # trainer trainer = OfflineTrainer( policy, buffer, test_collector, args.epoch, args.step_per_epoch, args.test_num, args.batch_size, save_best_fn=save_best_fn, stop_fn=stop_fn, logger=logger, ) for epoch, epoch_stat, info in trainer: print(f"Epoch: {epoch}") print(epoch_stat) print(info) assert stop_fn(info["best_reward"]) # Let's watch its performance! if __name__ == "__main__": pprint.pprint(info) 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()}")
def test_a2c(args=get_args()): env, train_envs, test_envs = make_mujoco_env(args.task, args.seed, args.training_num, args.test_num, obs_norm=True) 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, activation=nn.Tanh, device=args.device, ) actor = ActorProb( net_a, args.action_shape, max_action=args.max_action, unbounded=True, device=args.device, ).to(args.device) net_c = Net( args.state_shape, hidden_sizes=args.hidden_sizes, activation=nn.Tanh, device=args.device, ) critic = Critic(net_c, device=args.device).to(args.device) torch.nn.init.constant_(actor.sigma_param, -0.5) for m in list(actor.modules()) + list(critic.modules()): if isinstance(m, torch.nn.Linear): # orthogonal initialization torch.nn.init.orthogonal_(m.weight, gain=np.sqrt(2)) torch.nn.init.zeros_(m.bias) # do last policy layer scaling, this will make initial actions have (close to) # 0 mean and std, and will help boost performances, # see https://arxiv.org/abs/2006.05990, Fig.24 for details for m in actor.mu.modules(): if isinstance(m, torch.nn.Linear): torch.nn.init.zeros_(m.bias) m.weight.data.copy_(0.01 * m.weight.data) optim = torch.optim.RMSprop( list(actor.parameters()) + list(critic.parameters()), lr=args.lr, eps=1e-5, alpha=0.99, ) lr_scheduler = None if args.lr_decay: # decay learning rate to 0 linearly max_update_num = np.ceil( args.step_per_epoch / args.step_per_collect) * args.epoch lr_scheduler = LambdaLR( optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num) def dist(*logits): return Independent(Normal(*logits), 1) policy = A2CPolicy( actor, critic, optim, dist, discount_factor=args.gamma, gae_lambda=args.gae_lambda, max_grad_norm=args.max_grad_norm, vf_coef=args.vf_coef, ent_coef=args.ent_coef, reward_normalization=args.rew_norm, action_scaling=True, action_bound_method=args.bound_action_method, lr_scheduler=lr_scheduler, action_space=env.action_space, ) # load a previous policy if args.resume_path: ckpt = torch.load(args.resume_path, map_location=args.device) policy.load_state_dict(ckpt["model"]) train_envs.set_obs_rms(ckpt["obs_rms"]) test_envs.set_obs_rms(ckpt["obs_rms"]) 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) # log now = datetime.datetime.now().strftime("%y%m%d-%H%M%S") args.algo_name = "a2c" 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): state = { "model": policy.state_dict(), "obs_rms": train_envs.get_obs_rms() } torch.save(state, os.path.join(log_path, "policy.pth")) if not args.watch: # trainer result = onpolicy_trainer( policy, train_collector, test_collector, args.epoch, args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size, step_per_collect=args.step_per_collect, save_best_fn=save_best_fn, logger=logger, 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) 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 log_path = os.path.join( args.logdir, args.task, 'sac', 'seed_' + str(args.seed) + '_' + datetime.datetime.now().strftime('%m%d_%H%M%S') + '-' + args.task.replace('-', '_') + '_sac') writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = BasicLogger(writer) 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.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) # 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()): # initialize environment 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 = VectorEnv( [lambda: gym.make(args.task) for _ in range(args.training_num)]) 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, hidden_layer_size=args.hidden_size ).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, hidden_layer_size=args.hidden_size ).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, hidden_layer_size=args.hidden_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) 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=False, ignore_done=False) # collector if args.training_num == 0: max_episode_steps = train_envs._max_episode_steps else: max_episode_steps = train_envs.envs[0]._max_episode_steps train_collector = Collector( policy, train_envs, ReplayBuffer(args.buffer_size, max_ep_len=max_episode_steps)) test_collector = Collector(policy, test_envs, mode='test') # log log_path = os.path.join(args.logdir, args.task, 'sac', str(args.seed)) writer = SummaryWriter(log_path) def save_fn(policy, name='policy.pth'): torch.save(policy.state_dict(), os.path.join(log_path, name)) env.spec.reward_threshold = 100000 def stop_fn(x): return x >= env.spec.reward_threshold # trainer result = offpolicy_exact_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, epochs_to_save=[1, 50, 100, 150, 200]) 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()