def test_acktr(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) torch.nn.init.constant_(actor.sigma_param._bias, -0.5) # 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 = KFACOptimizer(actor, critic, lr=0.25) # 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 = ACKTRPolicy(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) # 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, 'acktr') 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()}")
def test_fqf(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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 feature_net = Net(args.state_shape, args.hidden_sizes[-1], hidden_sizes=args.hidden_sizes[:-1], device=args.device, softmax=False) net = FullQuantileFunction( feature_net, args.action_shape, args.hidden_sizes, num_cosines=args.num_cosines, device=args.device ) optim = torch.optim.Adam(net.parameters(), lr=args.lr) fraction_net = FractionProposalNetwork(args.num_fractions, net.input_dim) fraction_optim = torch.optim.RMSprop( fraction_net.parameters(), lr=args.fraction_lr ) policy = FQFPolicy( net, optim, fraction_net, fraction_optim, args.gamma, args.num_fractions, args.ent_coef, args.n_step, target_update_freq=args.target_update_freq ).to(args.device) # buffer if args.prioritized_replay: buf = PrioritizedVectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), alpha=args.alpha, beta=args.beta) else: buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs)) # collector train_collector = Collector(policy, train_envs, buf, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, args.task, 'fqf') 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 def train_fn(epoch, env_step): # eps annnealing, just a demo if env_step <= 10000: policy.set_eps(args.eps_train) elif env_step <= 50000: eps = args.eps_train - (env_step - 10000) / \ 40000 * (0.9 * args.eps_train) policy.set_eps(eps) else: policy.set_eps(0.1 * args.eps_train) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step) assert stop_fn(result['best_reward']) if __name__ == '__main__': pprint.pprint(result) # Let's watch its performance! env = gym.make(args.task) policy.eval() policy.set_eps(args.eps_test) 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 train_agent( args: argparse.Namespace = get_args(), agent_learn: Optional[BasePolicy] = None, agent_opponent: Optional[BasePolicy] = None, optim: Optional[torch.optim.Optimizer] = None, ) -> Tuple[dict, BasePolicy]: def env_func(): return TicTacToeEnv(args.board_size, args.win_size) train_envs = DummyVectorEnv([env_func for _ in range(args.training_num)]) test_envs = DummyVectorEnv([env_func 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) policy, optim = get_agents( args, agent_learn=agent_learn, agent_opponent=agent_opponent, optim=optim) # collector train_collector = Collector( policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, 'tic_tac_toe', 'dqn') writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = BasicLogger(writer) def save_fn(policy): if hasattr(args, 'model_save_path'): model_save_path = args.model_save_path else: model_save_path = os.path.join( args.logdir, 'tic_tac_toe', 'dqn', 'policy.pth') torch.save( policy.policies[args.agent_id - 1].state_dict(), model_save_path) def stop_fn(mean_rewards): return mean_rewards >= args.win_rate def train_fn(epoch, env_step): policy.policies[args.agent_id - 1].set_eps(args.eps_train) def test_fn(epoch, env_step): policy.policies[args.agent_id - 1].set_eps(args.eps_test) def reward_metric(rews): return rews[:, args.agent_id - 1] # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, update_per_step=args.update_per_step, logger=logger, test_in_train=False, reward_metric=reward_metric) return result, policy.policies[args.agent_id - 1]
def test_dqn(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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 Q_param = {"hidden_sizes": args.dueling_q_hidden_sizes} V_param = {"hidden_sizes": args.dueling_v_hidden_sizes} net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, dueling_param=(Q_param, V_param)).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) policy = DQNPolicy(net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq) # collector train_collector = Collector(policy, train_envs, VectorReplayBuffer(args.buffer_size, len(train_envs)), exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, args.task, 'dqn') 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 def train_fn(epoch, env_step): if env_step <= 100000: policy.set_eps(args.eps_train) elif env_step <= 500000: eps = args.eps_train - (env_step - 100000) / \ 400000 * (0.5 * args.eps_train) policy.set_eps(eps) else: policy.set_eps(0.5 * args.eps_train) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # 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, train_fn=train_fn, test_fn=test_fn, 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() policy.set_eps(args.eps_test) 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_ppo(args=get_args()): torch.set_num_threads(1) # for poor CPU 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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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 = Actor(net, args.action_shape, device=args.device).to(args.device) critic = Critic(net, 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 = torch.distributions.Categorical 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, gae_lambda=args.gae_lambda, reward_normalization=args.rew_norm, dual_clip=args.dual_clip, value_clip=args.value_clip, action_space=env.action_space, 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, 'ppo') 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, episode_per_collect=args.episode_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()}")
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.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(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 = 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)), exploration_noise=True) test_collector = Collector(policy, test_envs) # log log_path = os.path.join(args.logdir, args.task, 'ppo') 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, episode_per_collect=args.episode_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()}")
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) 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 = 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, action_space=env.action_space, action_scaling=True, action_bound_method="clip") 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_dqn(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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) # Q_param = V_param = {"hidden_sizes": [128]} # model net = Net( args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, # dueling=(Q_param, V_param), ).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) policy = DQNPolicy(net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq) # buffer if args.prioritized_replay: buf = PrioritizedVectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs), alpha=args.alpha, beta=args.beta) else: buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs)) # collector train_collector = Collector(policy, train_envs, buf, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, args.task, 'dqn') 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 def train_fn(epoch, env_step): # eps annnealing, just a demo if env_step <= 10000: policy.set_eps(args.eps_train) elif env_step <= 50000: eps = args.eps_train - (env_step - 10000) / \ 40000 * (0.9 * args.eps_train) policy.set_eps(eps) else: policy.set_eps(0.1 * args.eps_train) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # 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, train_fn=train_fn, test_fn=test_fn, 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() policy.set_eps(args.eps_test) 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()}") # save buffer in pickle format, for imitation learning unittest buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(test_envs)) policy.set_eps(0.2) collector = Collector(policy, test_envs, buf, exploration_noise=True) result = collector.collect(n_step=args.buffer_size) pickle.dump(buf, open(args.save_buffer_name, "wb")) print(result["rews"].mean())
def gomoku(args=get_args()): Collector._default_rew_metric = lambda x: x[args.agent_id - 1] if args.watch: watch(args) return policy, optim = get_agents(args) agent_learn = policy.policies[args.agent_id - 1] agent_opponent = policy.policies[2 - args.agent_id] # log log_path = os.path.join(args.logdir, 'Gomoku', 'dqn') writer = SummaryWriter(log_path) args.logger = BasicLogger(writer) opponent_pool = [agent_opponent] def env_func(): return TicTacToeEnv(args.board_size, args.win_size) test_envs = DummyVectorEnv([env_func for _ in range(args.test_num)]) for r in range(args.self_play_round): rews = [] agent_learn.set_eps(0.0) # compute the reward over previous learner for opponent in opponent_pool: policy.replace_policy(opponent, 3 - args.agent_id) test_collector = Collector(policy, test_envs) results = test_collector.collect(n_episode=100) rews.append(results['rews'].mean()) rews = np.array(rews) # weight opponent by their difficulty level rews = np.exp(-rews * 10.0) rews /= np.sum(rews) total_epoch = args.epoch args.epoch = 1 for epoch in range(total_epoch): # sample one opponent opp_id = np.random.choice(len(opponent_pool), size=1, p=rews) print(f'selection probability {rews.tolist()}') print(f'selected opponent {opp_id}') opponent = opponent_pool[opp_id.item(0)] agent = RandomPolicy() # previous learner can only be used for forward agent.forward = opponent.forward args.model_save_path = os.path.join( args.logdir, 'Gomoku', 'dqn', f'policy_round_{r}_epoch_{epoch}.pth') result, agent_learn = train_agent(args, agent_learn=agent_learn, agent_opponent=agent, optim=optim) print(f'round_{r}_epoch_{epoch}') pprint.pprint(result) learnt_agent = deepcopy(agent_learn) learnt_agent.set_eps(0.0) opponent_pool.append(learnt_agent) args.epoch = total_epoch if __name__ == '__main__': # Let's watch its performance! opponent = opponent_pool[-2] watch(args, agent_learn, opponent)
def test_c51(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, softmax=True, num_atoms=args.num_atoms) optim = torch.optim.Adam(net.parameters(), lr=args.lr) policy = C51Policy( net, optim, args.gamma, args.num_atoms, args.v_min, args.v_max, args.n_step, target_update_freq=args.target_update_freq ).to(args.device) # buffer if args.prioritized_replay: buf = PrioritizedVectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), alpha=args.alpha, beta=args.beta) else: buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs)) # collector train_collector = Collector(policy, train_envs, buf, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, args.task, 'c51') 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 train_fn(epoch, env_step): # eps annnealing, just a demo if env_step <= 10000: policy.set_eps(args.eps_train) elif env_step <= 50000: eps = args.eps_train - (env_step - 10000) / \ 40000 * (0.9 * args.eps_train) policy.set_eps(eps) else: policy.set_eps(0.1 * args.eps_train) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) 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')) pickle.dump(train_collector.buffer, open(os.path.join(log_path, 'train_buffer.pkl'), "wb")) 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']) policy.optim.load_state_dict(checkpoint['optim']) print("Successfully restore policy and optim.") else: print("Fail to restore policy and optim.") buffer_path = os.path.join(log_path, 'train_buffer.pkl') if os.path.exists(buffer_path): train_collector.buffer = pickle.load(open(buffer_path, "rb")) print("Successfully restore buffer.") else: print("Fail to restore buffer.") # 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, train_fn=train_fn, test_fn=test_fn, 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() policy.set_eps(args.eps_test) 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_ddpg(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 = Actor(net, args.action_shape, max_action=args.max_action, device=args.device).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) critic = Critic(net, device=args.device).to(args.device) critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr) policy = DDPGPolicy( actor, actor_optim, critic, critic_optim, tau=args.tau, gamma=args.gamma, exploration_noise=GaussianNoise(sigma=args.exploration_noise), 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) # log log_path = os.path.join(args.logdir, args.task, 'ddpg') 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()}")
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')) 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_dqn(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv train_envs = DummyVectorEnv( [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 Q_param = {"hidden_sizes": args.dueling_q_hidden_sizes} V_param = {"hidden_sizes": args.dueling_v_hidden_sizes} net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, dueling_param=(Q_param, V_param)).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) # prepare hyperparameters adaptive_scheme = args.adaptive_scheme adaptive_scheme[4] *= args.update_per_step adaptive_scheme[5] *= args.update_per_step reweigh_hyper = { "hard_weight": args.tper_weight, "linear": args.linear_hp, "adaptive_linear": args.adaptive_scheme, } policy = TPDQNPolicy( net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq, bk_step=args.bk_step, reweigh_type=args.reweigh_type, reweigh_hyper=reweigh_hyper) # collector train_collector = Collector( policy, train_envs, TPVectorReplayBuffer(args.buffer_size, len(train_envs)), preprocess_fn=StepPreprocess(len(train_envs), args.bk_step).get_step, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log cur_time = time.strftime('%y-%m-%d-%H-%M-%S', time.localtime()) log_path = os.path.join(args.logdir, args.task, 'tpdqn', "%s-seed%d"%(args.exp, args.seed), cur_time) 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 def train_fn(epoch, env_step): # exp decay eps = max(args.eps_train * (1 - 5e-6) ** env_step, args.eps_test) policy.set_eps(eps) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # 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, train_fn=train_fn, test_fn=test_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() policy.set_eps(args.eps_test) 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_discrete_bcq(args=get_args()): # envs env = gym.make(args.task) if args.task == 'CartPole-v0': env.spec.reward_threshold = 190 # lower the goal args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n 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 policy_net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device).to(args.device) imitation_net = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device).to(args.device) optim = torch.optim.Adam(list(policy_net.parameters()) + list(imitation_net.parameters()), lr=args.lr) policy = DiscreteBCQPolicy( policy_net, imitation_net, optim, args.gamma, args.n_step, args.target_update_freq, args.eps_test, args.unlikely_action_threshold, args.imitation_logits_penalty, ) # buffer assert os.path.exists(args.load_buffer_name), \ "Please run test_dqn.py first to get expert's data buffer." buffer = pickle.load(open(args.load_buffer_name, "rb")) # collector test_collector = Collector(policy, test_envs, exploration_noise=True) log_path = os.path.join(args.logdir, args.task, 'discrete_bcq') 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.") result = offline_trainer(policy, buffer, test_collector, args.epoch, args.update_per_epoch, args.test_num, args.batch_size, 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() policy.set_eps(args.eps_test) 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_ddpg(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] args.exploration_noise = args.exploration_noise * args.max_action 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 = Actor( net_a, args.action_shape, max_action=args.max_action, device=args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, concat=True, device=args.device) critic = Critic(net_c, device=args.device).to(args.device) critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr) policy = DDPGPolicy( actor, actor_optim, critic, critic_optim, tau=args.tau, gamma=args.gamma, exploration_noise=GaussianNoise(sigma=args.exploration_noise), 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("-", "_")}_ddpg' log_path = os.path.join(args.logdir, args.task, 'ddpg', log_file) 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_a2c_with_il(args=get_args()): torch.set_num_threads(1) # for poor CPU 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 # 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 = Actor(net, args.action_shape, device=args.device).to(args.device) critic = Critic(net, device=args.device).to(args.device) optim = torch.optim.Adam(set(actor.parameters()).union( critic.parameters()), lr=args.lr) dist = torch.distributions.Categorical policy = A2CPolicy(actor, critic, optim, dist, args.gamma, gae_lambda=args.gae_lambda, vf_coef=args.vf_coef, ent_coef=args.ent_coef, max_grad_norm=args.max_grad_norm, reward_normalization=args.rew_norm, 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) # log log_path = os.path.join(args.logdir, args.task, 'a2c') 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, episode_per_collect=args.episode_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()}") policy.eval() # here we define an imitation collector with a trivial policy if args.task == 'CartPole-v0': env.spec.reward_threshold = 190 # lower the goal net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) net = Actor(net, args.action_shape, device=args.device).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.il_lr) il_policy = ImitationPolicy(net, optim, mode='discrete') 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_dqn(args=get_args()): env = make_minigrid_env(args) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.env.action_space.shape or env.env.action_space.n # should be N_FRAMES x H x W print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) # make environments train_envs = SubprocVectorEnv( [lambda: make_minigrid_env(args) for _ in range(args.training_num)]) test_envs = SubprocVectorEnv( [lambda: make_minigrid_env_watch(args) 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) # define model net = DQN(args.state_shape[2], args.state_shape[0], args.state_shape[1], args.action_shape, args.device).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) # define policy policy = DQNPolicy(net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq) # 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) # replay buffer: `save_last_obs` and `stack_num` can be removed together # when you have enough RAM buffer = VectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), ignore_obs_next=True, ) # collector train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # log cur_time = time.strftime('%y-%m-%d-%H-%M-%S', time.localtime()) log_path = os.path.join(args.logdir, args.task, 'dqn', args.exp, str(args.seed), cur_time) 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')) def save_fn_each_epoch(policy, epoch): torch.save(policy.state_dict(), os.path.join(log_path, 'policy-%d.pth' % epoch)) def stop_fn(mean_rewards): # if env.env.spec.reward_threshold: # return mean_rewards >= env.spec.reward_threshold # elif 'Pong' in args.task: # return mean_rewards >= 20 # else: # return False return False def train_fn(epoch, env_step): # nature DQN setting, linear decay in the first 1M steps if env_step <= 1e6: eps = args.eps_train - env_step / 1e6 * \ (args.eps_train - args.eps_train_final) else: eps = args.eps_train_final policy.set_eps(eps) if env_step % 1e4 == 0 and env_step != 0 and args.dense_save_ckpt: save_fn_each_epoch(policy, env_step / 1e4) logger.write('train/eps', env_step, eps) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # watch agent's performance def watch(): print("Setup test envs ...") policy.eval() policy.set_eps(args.eps_test) test_envs.seed(args.seed) if args.save_buffer_name: print(f"Generate buffer with size {args.buffer_size}") buffer = VectorReplayBuffer(args.buffer_size, buffer_num=len(test_envs), ignore_obs_next=True, save_only_last_obs=True, stack_num=args.frames_stack) collector = Collector(policy, test_envs, buffer) result = collector.collect(n_step=args.buffer_size) print(f"Save buffer into {args.save_buffer_name}") # Unfortunately, pickle will cause oom with 1M buffer size buffer.save_hdf5(args.save_buffer_name) else: print("Testing agent ...") test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) pprint.pprint(result) if args.watch: watch() exit(0) # test train_collector and start filling replay buffer train_collector.collect(n_step=args.batch_size * args.training_num) # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False) pprint.pprint(result) watch()
def test_discrete_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 train_envs = SubprocVectorEnv( [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 net = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) actor = Actor(net, args.action_shape, softmax_output=False, device=args.device).to(args.device) actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr) net_c1 = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) critic1 = Critic(net_c1, last_size=args.action_shape, device=args.device).to(args.device) critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr) net_c2 = Net(args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device) critic2 = Critic(net_c2, last_size=args.action_shape, device=args.device).to(args.device) critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr) # better not to use auto alpha in CartPole if args.auto_alpha: target_entropy = 0.98 * np.log(np.prod(args.action_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 = DiscreteSACPolicy( actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim, args.tau, args.gamma, args.alpha, estimation_step=args.n_step, reward_normalization=args.rew_norm) # 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, 'discrete_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, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False) 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_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_c51(args=get_args()): args.cfg_path = f"maps/{args.task}.cfg" args.wad_path = f"maps/{args.task}.wad" args.res = (args.skip_num, 84, 84) env = Env(args.cfg_path, args.frames_stack, args.res) args.state_shape = args.res args.action_shape = env.action_space.shape or env.action_space.n # should be N_FRAMES x H x W print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) # make environments train_envs = SubprocVectorEnv([ lambda: Env(args.cfg_path, args.frames_stack, args.res) for _ in range(args.training_num) ]) test_envs = SubprocVectorEnv([ lambda: Env(args.cfg_path, args.frames_stack, args.res, args.save_lmp) for _ in range(min(os.cpu_count() - 1, args.test_num)) ]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) train_envs.seed(args.seed) test_envs.seed(args.seed) # define model net = C51(*args.state_shape, args.action_shape, args.num_atoms, args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) # define policy policy = C51Policy(net, optim, args.gamma, args.num_atoms, args.v_min, args.v_max, args.n_step, target_update_freq=args.target_update_freq).to( 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) # replay buffer: `save_last_obs` and `stack_num` can be removed together # when you have enough RAM buffer = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs), ignore_obs_next=True, save_only_last_obs=True, stack_num=args.frames_stack) # collector train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # log log_path = os.path.join(args.logdir, args.task, 'c51') 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')) def stop_fn(mean_rewards): if env.spec.reward_threshold: return mean_rewards >= env.spec.reward_threshold elif 'Pong' in args.task: return mean_rewards >= 20 else: return False def train_fn(epoch, env_step): # nature DQN setting, linear decay in the first 1M steps if env_step <= 1e6: eps = args.eps_train - env_step / 1e6 * \ (args.eps_train - args.eps_train_final) else: eps = args.eps_train_final policy.set_eps(eps) logger.write('train/eps', env_step, eps) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # watch agent's performance def watch(): print("Setup test envs ...") policy.eval() policy.set_eps(args.eps_test) test_envs.seed(args.seed) if args.save_buffer_name: print(f"Generate buffer with size {args.buffer_size}") buffer = VectorReplayBuffer(args.buffer_size, buffer_num=len(test_envs), ignore_obs_next=True, save_only_last_obs=True, stack_num=args.frames_stack) collector = Collector(policy, test_envs, buffer, exploration_noise=True) result = collector.collect(n_step=args.buffer_size) print(f"Save buffer into {args.save_buffer_name}") # Unfortunately, pickle will cause oom with 1M buffer size buffer.save_hdf5(args.save_buffer_name) else: print("Testing agent ...") test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) rew = result["rews"].mean() lens = result["lens"].mean() * args.skip_num print(f'Mean reward (over {result["n/ep"]} episodes): {rew}') print(f'Mean length (over {result["n/ep"]} episodes): {lens}') if args.watch: watch() exit(0) # test train_collector and start filling replay buffer train_collector.collect(n_step=args.batch_size * args.training_num) # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False) pprint.pprint(result) watch()
def test_drqn(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 # train_envs = gym.make(args.task) # you can also use tianshou.env.SubprocVectorEnv 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 = Recurrent(args.layer_num, args.state_shape, args.action_shape, args.device).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) policy = DQNPolicy( net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq) # collector buffer = VectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), stack_num=args.stack_num, ignore_obs_next=True) train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) # the stack_num is for RNN training: sample framestack obs test_collector = Collector(policy, test_envs, exploration_noise=True) # policy.set_eps(1) train_collector.collect(n_step=args.batch_size * args.training_num) # log log_path = os.path.join(args.logdir, args.task, 'drqn') 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 def train_fn(epoch, env_step): policy.set_eps(args.eps_train) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # 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, train_fn=train_fn, test_fn=test_fn, 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()}")
def test_dqn(args=get_args()): if 'ram' in args.task and 'NoFrame' not in args.task: use_ram = True else: use_ram = False if use_ram: env = make_ram_env(args) make_env_fn = make_ram_env make_watch_fn = make_ram_env_watch save_only_last_obs = False else: env = make_atari_env(args) make_env_fn = make_atari_env make_watch_fn = make_atari_env_watch save_only_last_obs = True args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.env.action_space.shape or env.env.action_space.n # should be N_FRAMES x H x W print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) # make environments train_envs = SubprocVectorEnv( [lambda: make_env_fn(args) for _ in range(args.training_num)]) test_envs = SubprocVectorEnv( [lambda: make_watch_fn(args) 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) # define model if use_ram: net = RamDQN(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device).to(args.device) elif args.lfiw: net = LfiwDQN(*args.state_shape, args.action_shape, args.device).to(args.device) else: net = DQN(*args.state_shape, args.action_shape, args.device).to(args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) # possible TODO: lfiw_optim over non-cnn parameters # prepare hyperparameters adaptive_scheme = args.adaptive_scheme adaptive_scheme[4] *= args.update_per_step adaptive_scheme[5] *= args.update_per_step reweigh_hyper = { "hard_weight": args.tper_weight, "linear": args.linear_hp, "adaptive_linear": args.adaptive_scheme, } # define policy if args.lfiw: policy = LfiwTPDQNPolicy(net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq, bk_step=args.bk_step, reweigh_type=args.reweigh_type, reweigh_hyper=reweigh_hyper, opd_temperature=args.lfiw_temp, opd_loss_coeff=args.lfiw_loss_coeff) else: policy = TPDQNPolicy(net, optim, args.gamma, args.n_step, target_update_freq=args.target_update_freq, bk_step=args.bk_step, reweigh_type=args.reweigh_type, reweigh_hyper=reweigh_hyper) # 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) # replay buffer: `save_last_obs` and `stack_num` can be removed together # when you have enough RAM if args.lfiw: buffer = TPDoubleVectorReplayBuffer( args.buffer_size, buffer_num=len(train_envs), bk_step=args.bk_step, ignore_obs_next=True, save_only_last_obs=save_only_last_obs, stack_num=args.frames_stack, fast_buffer_size=args.fast_buffer_size) else: buffer = TPVectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs), bk_step=args.bk_step, ignore_obs_next=True, save_only_last_obs=save_only_last_obs, stack_num=args.frames_stack) # collector train_collector = Collector(policy, train_envs, buffer, preprocess_fn=StepPreprocess( len(train_envs), args.bk_step).get_step, exploration_noise=True) # print(len(test_envs)) test_collector = Collector( policy, test_envs, exploration_noise=True, ) # log cur_time = time.strftime('%y-%m-%d-%H-%M-%S', time.localtime()) log_path = os.path.join(args.logdir, args.task, 'tpdqn', "%s-seed%d" % (args.exp, args.seed), cur_time) 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')) def stop_fn(mean_rewards): # if env.env.spec.reward_threshold: # return mean_rewards >= env.spec.reward_threshold # elif 'Pong' in args.task: # return mean_rewards >= 20 # else: # return False return False def train_fn(epoch, env_step): # nature DQN setting, linear decay in the first 1M steps if env_step <= 1e6: eps = args.eps_train - env_step / 1e6 * \ (args.eps_train - args.eps_train_final) else: eps = args.eps_train_final policy.set_eps(eps) logger.write('train/eps', env_step, eps) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # watch agent's performance def watch(): print("Setup test envs ...") policy.eval() policy.set_eps(args.eps_test) test_envs.seed(args.seed) if args.save_buffer_name: print(f"Generate buffer with size {args.buffer_size}") buffer = TPVectorReplayBuffer(args.buffer_size, buffer_num=len(test_envs), ignore_obs_next=True, save_only_last_obs=True, stack_num=args.frames_stack) collector = Collector(policy, test_envs, buffer) result = collector.collect(n_step=args.buffer_size) print(f"Save buffer into {args.save_buffer_name}") # Unfortunately, pickle will cause oom with 1M buffer size buffer.save_hdf5(args.save_buffer_name) else: print("Testing agent ...") test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) pprint.pprint(result) if args.watch: watch() exit(0) # test train_collector and start filling replay buffer train_collector.collect(n_step=args.batch_size * args.training_num) # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False) pprint.pprint(result) watch()
def test_discrete_crr(args=get_args()): # envs env = gym.make(args.task) if args.task == 'CartPole-v0': env.spec.reward_threshold = 190 # lower the goal args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n 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 = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, softmax=False) critic = Net(args.state_shape, args.action_shape, hidden_sizes=args.hidden_sizes, device=args.device, softmax=False) optim = torch.optim.Adam(list(actor.parameters()) + list(critic.parameters()), lr=args.lr) policy = DiscreteCRRPolicy( actor, critic, optim, args.gamma, target_update_freq=args.target_update_freq, ).to(args.device) # buffer assert os.path.exists(args.load_buffer_name), \ "Please run test_dqn.py first to get expert's data buffer." buffer = pickle.load(open(args.load_buffer_name, "rb")) # collector test_collector = Collector(policy, test_envs, exploration_noise=True) log_path = os.path.join(args.logdir, args.task, 'discrete_cql') 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 result = offline_trainer(policy, buffer, test_collector, args.epoch, args.update_per_epoch, 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) 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_discrete_bcq(args=get_args()): # envs env = make_atari_env(args) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.action_space.shape or env.action_space.n # should be N_FRAMES x H x W print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) # make environments test_envs = SubprocVectorEnv( [lambda: make_atari_env_watch(args) for _ in range(args.test_num)]) # seed np.random.seed(args.seed) torch.manual_seed(args.seed) test_envs.seed(args.seed) # model feature_net = DQN(*args.state_shape, args.action_shape, device=args.device, features_only=True).to(args.device) policy_net = Actor(feature_net, args.action_shape, device=args.device, hidden_sizes=args.hidden_sizes, softmax_output=False).to(args.device) imitation_net = Actor(feature_net, args.action_shape, device=args.device, hidden_sizes=args.hidden_sizes, softmax_output=False).to(args.device) optim = torch.optim.Adam(list(policy_net.parameters()) + list(imitation_net.parameters()), lr=args.lr) # define policy policy = DiscreteBCQPolicy(policy_net, imitation_net, optim, args.gamma, args.n_step, args.target_update_freq, args.eps_test, args.unlikely_action_threshold, args.imitation_logits_penalty) # 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) # buffer assert os.path.exists(args.load_buffer_name), \ "Please run atari_dqn.py first to get expert's data buffer." if args.load_buffer_name.endswith('.pkl'): buffer = pickle.load(open(args.load_buffer_name, "rb")) elif args.load_buffer_name.endswith('.hdf5'): buffer = VectorReplayBuffer.load_hdf5(args.load_buffer_name) else: print(f"Unknown buffer format: {args.load_buffer_name}") exit(0) # collector test_collector = Collector(policy, test_envs, exploration_noise=True) # log log_path = os.path.join( args.logdir, args.task, 'bcq', f'seed_{args.seed}_{datetime.datetime.now().strftime("%m%d-%H%M%S")}') writer = SummaryWriter(log_path) writer.add_text("args", str(args)) logger = BasicLogger(writer, update_interval=args.log_interval) def save_fn(policy): torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth')) def stop_fn(mean_rewards): return False # watch agent's performance def watch(): print("Setup test envs ...") policy.eval() policy.set_eps(args.eps_test) test_envs.seed(args.seed) print("Testing agent ...") test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) pprint.pprint(result) rew = result["rews"].mean() print(f'Mean reward (over {result["n/ep"]} episodes): {rew}') if args.watch: watch() exit(0) result = offline_trainer(policy, buffer, test_collector, args.epoch, args.update_per_epoch, args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn, logger=logger) pprint.pprint(result) watch()
def test_a2c(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._bias, -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: 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("-", "_")}_a2c' log_path = os.path.join(args.logdir, args.task, 'a2c', 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_qrdqn(args=get_args()): env = make_atari_env(args) args.state_shape = env.observation_space.shape or env.observation_space.n args.action_shape = env.env.action_space.shape or env.env.action_space.n # should be N_FRAMES x H x W print("Observations shape:", args.state_shape) print("Actions shape:", args.action_shape) # make environments train_envs = SubprocVectorEnv( [lambda: make_atari_env(args) for _ in range(args.training_num)]) test_envs = SubprocVectorEnv( [lambda: make_atari_env_watch(args) 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) # define model net = QRDQN(*args.state_shape, args.action_shape, args.num_quantiles, args.device) optim = torch.optim.Adam(net.parameters(), lr=args.lr) # define policy policy = QRDQNPolicy(net, optim, args.gamma, args.num_quantiles, args.n_step, target_update_freq=args.target_update_freq).to( 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) # replay buffer: `save_last_obs` and `stack_num` can be removed together # when you have enough RAM buffer = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs), ignore_obs_next=True, save_only_last_obs=True, stack_num=args.frames_stack) # collector train_collector = Collector(policy, train_envs, buffer, exploration_noise=True) test_collector = Collector(policy, test_envs, exploration_noise=True) # log log_path = os.path.join(args.logdir, args.task, 'qrdqn') 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')) def stop_fn(mean_rewards): if env.env.spec.reward_threshold: return mean_rewards >= env.spec.reward_threshold elif 'Pong' in args.task: return mean_rewards >= 20 else: return False def train_fn(epoch, env_step): # nature DQN setting, linear decay in the first 1M steps if env_step <= 1e6: eps = args.eps_train - env_step / 1e6 * \ (args.eps_train - args.eps_train_final) else: eps = args.eps_train_final policy.set_eps(eps) logger.write('train/eps', env_step, eps) def test_fn(epoch, env_step): policy.set_eps(args.eps_test) # watch agent's performance def watch(): print("Testing agent ...") policy.eval() policy.set_eps(args.eps_test) test_envs.seed(args.seed) test_collector.reset() result = test_collector.collect(n_episode=args.test_num, render=args.render) pprint.pprint(result) if args.watch: watch() exit(0) # test train_collector and start filling replay buffer train_collector.collect(n_step=args.batch_size * args.training_num) # 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, train_fn=train_fn, test_fn=test_fn, stop_fn=stop_fn, save_fn=save_fn, logger=logger, update_per_step=args.update_per_step, test_in_train=False) pprint.pprint(result) watch()