def load_act(path): with open(path, "rb") as f: model_data, act_params = cloudpickle.load(f) act = deepq.build_act(**act_params) sess = tf.Session() sess.__enter__() with tempfile.TemporaryDirectory() as td: arc_path = os.path.join(td, "packed.zip") with open(arc_path, "wb") as f: f.write(model_data) zipfile.ZipFile(arc_path, 'r', zipfile.ZIP_DEFLATED).extractall(td) load_variables(os.path.join(td, "model")) return ActWrapper(act, act_params)
def load(self, path): U.load_variables(path, sess=self.sess)
def learn(env, network, seed=None, pool=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_initial_eps=1.0, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=1, checkpoint_freq=100, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, experiment_name='unnamed', load_path=None, **network_kwargs): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. experiment_name: str name of the experiment (default: trial) load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=make_obs_ph, q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=exploration_initial_eps, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() reward_shaper = ActionAdviceRewardShaper('../completed-observations') reward_shaper.load() full_exp_name = '{}-{}'.format(date.today().isoformat(), experiment_name) experiment_dir = os.path.join('experiments', full_exp_name) if not os.path.exists(experiment_dir): os.makedirs(experiment_dir) summary_dir = os.path.join(experiment_dir, 'summaries') os.makedirs(summary_dir, exist_ok=True) summary_writer = tf.summary.FileWriter(summary_dir) checkpoint_dir = os.path.join(experiment_dir, 'checkpoints') os.makedirs(checkpoint_dir, exist_ok=True) with tempfile.TemporaryDirectory() as td: td = checkpoint_dir or td os.makedirs(td, exist_ok=True) model_file = os.path.join(td, "best_model") model_saved = False saved_mean_reward = None if os.path.exists(model_file): print('Model is loading') load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) episode_rewards = [] update_step_t = 0 while update_step_t < total_timesteps: # Reset the environment obs = env.reset() obs = StatePreprocessor.process(obs) episode_rewards.append(0.0) reset = True done = False # Sample the episode until it is completed act_step_t = update_step_t while not done: if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(act_step_t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(act_step_t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log( 1. - exploration.value(act_step_t) + exploration.value(act_step_t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True biases = reward_shaper.get_action_potentials(obs) action = act(np.array(obs)[None], biases, update_eps=update_eps, **kwargs)[0] reset = False pairs = env.step(action) action, (new_obs, rew, done, _) = pairs[-1] # Write down the real reward but learn from normalized version episode_rewards[-1] += rew rew = np.sign(rew) * np.log(1 + np.abs(rew)) new_obs = StatePreprocessor.process(new_obs) logger.log('{}/{} obs {} action {}'.format( act_step_t, total_timesteps, obs, action)) act_step_t += 1 if len(new_obs) == 0: done = True else: replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs # Post episode logging summary = tf.Summary(value=[ tf.Summary.Value(tag="rewards", simple_value=episode_rewards[-1]) ]) summary_writer.add_summary(summary, act_step_t) summary = tf.Summary( value=[tf.Summary.Value(tag="eps", simple_value=update_eps)]) summary_writer.add_summary(summary, act_step_t) summary = tf.Summary(value=[ tf.Summary.Value(tag="episode_steps", simple_value=act_step_t - update_step_t) ]) summary_writer.add_summary(summary, act_step_t) mean_5ep_reward = round(np.mean(episode_rewards[-5:]), 1) num_episodes = len(episode_rewards) if print_freq is not None and num_episodes % print_freq == 0: logger.record_tabular("steps", act_step_t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 5 episode reward", mean_5ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(act_step_t))) logger.dump_tabular() # Do the learning start = time.time() while update_step_t < min(act_step_t, total_timesteps): if update_step_t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample( batch_size, beta=beta_schedule.value(update_step_t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( batch_size) weights, batch_idxes = np.ones_like(rewards), None biases_t = pool.map(reward_shaper.get_action_potentials, obses_t) biases_tp1 = pool.map(reward_shaper.get_action_potentials, obses_tp1) td_errors, weighted_error = train(obses_t, biases_t, actions, rewards, obses_tp1, biases_tp1, dones, weights) # Loss logging summary = tf.Summary(value=[ tf.Summary.Value(tag='weighted_error', simple_value=weighted_error) ]) summary_writer.add_summary(summary, update_step_t) if prioritized_replay: new_priorities = np.abs( td_errors) + prioritized_replay_eps replay_buffer.update_priorities( batch_idxes, new_priorities) if update_step_t % target_network_update_freq == 0: # Update target network periodically. update_target() update_step_t += 1 stop = time.time() logger.log("Learning took {:.2f} seconds".format(stop - start)) if checkpoint_freq is not None and num_episodes % checkpoint_freq == 0: # Periodically save the model and the replay buffer rec_model_file = os.path.join( td, "model_{}_{:.2f}".format(num_episodes, mean_5ep_reward)) save_variables(rec_model_file) buffer_file = os.path.join( td, "buffer_{}_{}".format(num_episodes, update_step_t)) with open(buffer_file, 'wb') as foutput: cloudpickle.dump(replay_buffer, foutput) # Check whether it is best if saved_mean_reward is None or mean_5ep_reward > saved_mean_reward: if print_freq is not None: logger.log( "Saving model due to mean reward increase: {} -> {}" .format(saved_mean_reward, mean_5ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_5ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format( saved_mean_reward)) load_variables(model_file) return act
def learn(env, policy_func, reward_giver, expert_dataset, rank, pretrained, pretrained_weight, *, g_step, d_step, entcoeff, save_per_iter, ckpt_dir, log_dir, timesteps_per_batch, task_name, gamma, lam, max_kl, cg_iters, cg_damping=1e-2, vf_stepsize=1e-4, d_stepsize=1e-4, vf_iters=3, max_timesteps=0, max_episodes=0, max_iters=0, callback=None): nworkers = MPI.COMM_WORLD.Get_size() rank = MPI.COMM_WORLD.Get_rank() np.set_printoptions(precision=3) # Setup losses and stuff # ---------------------------------------- ob_space = env.observation_space ac_space = env.action_space pi = policy_func("pi", ob_space, ac_space, reuse=(pretrained_weight != None)) # oldpi = policy_func("oldpi", ob_space, ac_space) atarg = tf.placeholder( dtype=tf.float32, shape=[None]) # Target advantage function (if applicable) ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return #ob = U.get_placeholder_cached(name="ob") ob_config = U.get_placeholder_cached(name="ob") ob_target = U.get_placeholder_cached(name="goal") obs_pos = U.get_placeholder_cached(name="obs_pos") obs_ori = U.get_placeholder_cached(name="obs_ori") ac = pi.pdtype.sample_placeholder([None]) kloldnew = oldpi.pd.kl(pi.pd) ent = pi.pd.entropy() meankl = tf.reduce_mean(kloldnew) meanent = tf.reduce_mean(ent) entbonus = entcoeff * meanent vferr = tf.reduce_mean(tf.square(pi.vpred - ret)) ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # advantage * pnew / pold surrgain = tf.reduce_mean(ratio * atarg) optimgain = surrgain + entbonus losses = [optimgain, meankl, entbonus, surrgain, meanent] loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"] dist = meankl all_var_list = pi.get_trainable_variables() var_list = [ v for v in all_var_list if v.name.startswith("pi/pol") or v.name.startswith("pi/logstd") or v.name.startswith("pi/obs") ] vf_var_list = [ v for v in all_var_list if v.name.startswith("pi/vf") or v.name.startswith("pi/obs") ] # assert len(var_list) == len(vf_var_list) + 1 d_adam = MpiAdam(reward_giver.get_trainable_variables()) vfadam = MpiAdam(vf_var_list) get_flat = U.GetFlat(var_list) set_from_flat = U.SetFromFlat(var_list) klgrads = tf.gradients(dist, var_list) flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan") shapes = [var.get_shape().as_list() for var in var_list] start = 0 tangents = [] for shape in shapes: sz = U.intprod(shape) tangents.append(tf.reshape(flat_tangent[start:start + sz], shape)) start += sz gvp = tf.add_n([ tf.reduce_sum(g * tangent) for (g, tangent) in zipsame(klgrads, tangents) ]) # pylint: disable=E1111 fvp = U.flatgrad(gvp, var_list) assign_old_eq_new = U.function( [], [], updates=[ tf.assign(oldv, newv) for (oldv, newv) in zipsame(oldpi.get_variables(), pi.get_variables()) ]) compute_losses = U.function( [ob_config, ob_target, obs_pos, obs_ori, ac, atarg], losses) compute_lossandgrad = U.function( [ob_config, ob_target, obs_pos, obs_ori, ac, atarg], losses + [U.flatgrad(optimgain, var_list)]) compute_fvp = U.function( [flat_tangent, ob_config, ob_target, obs_pos, obs_ori, ac, atarg], fvp) compute_vflossandgrad = U.function( [ob_config, ob_target, obs_pos, obs_ori, ret], U.flatgrad(vferr, vf_var_list)) @contextmanager def timed(msg): if rank == 0: print(colorize(msg, color='magenta')) tstart = time.time() yield print( colorize("done in %.3f seconds" % (time.time() - tstart), color='magenta')) else: yield def allmean(x): assert isinstance(x, np.ndarray) out = np.empty_like(x) MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM) out /= nworkers return out U.initialize() th_init = get_flat() MPI.COMM_WORLD.Bcast(th_init, root=0) set_from_flat(th_init) d_adam.sync() vfadam.sync() if rank == 0: print("Init param sum", th_init.sum(), flush=True) # Prepare for rollouts # ---------------------------------------- seg_gen = traj_segment_generator(pi, env, reward_giver, timesteps_per_batch, stochastic=True) episodes_so_far = 0 timesteps_so_far = 0 iters_so_far = 0 tstart = time.time() lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards true_rewbuffer = deque(maxlen=40) assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1 g_loss_stats = stats(loss_names) d_loss_stats = stats(reward_giver.loss_name) ep_stats = stats(["True_rewards", "Rewards", "Episode_length"]) # if provide pretrained weight if pretrained_weight is not None: U.load_variables(pretrained_weight, variables=pi.get_variables()) th_afterbc = get_flat() print("param sum after bc", th_afterbc.sum(), flush=True) while True: if callback: callback(locals(), globals()) if max_timesteps and timesteps_so_far >= max_timesteps: break elif max_episodes and episodes_so_far >= max_episodes: break elif max_iters and iters_so_far >= max_iters: break # Save model if rank == 0 and iters_so_far % save_per_iter == 0 and ckpt_dir is not None: fname = os.path.join(ckpt_dir, task_name) os.makedirs(os.path.dirname(fname), exist_ok=True) saver = tf.train.Saver() saver.save(tf.get_default_session(), fname) logger.log("********** Iteration %i ************" % iters_so_far) def fisher_vector_product(p): v1 = allmean(compute_fvp(p, *fvpargs)) # print("norm(v1):%.2e, norm(p):%.2e, cg_damping:%.2e"%(np.linalg.norm(v1), np.linalg.norm(p), cg_damping)) return v1 + cg_damping * p # ------------------ Update G ------------------ logger.log("Optimizing Policy...") for _ in range(g_step): with timed("sampling"): seg = seg_gen.__next__() add_vtarg_and_adv(seg, gamma, lam) # ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets)) ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[ "tdlamret"] vpredbefore = seg[ "vpred"] # predicted value function before udpate atarg = (atarg - atarg.mean()) / atarg.std( ) # standardized advantage function estimate if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy config, goal, obstacle_pos, obstacle_ori = [], [], [], [] for o in seg["ob"]: config.append(o["joint"]) goal.append(o["target"]) obstacle_pos.append(o["obstacle_pos"]) obstacle_ori.append(o["obstacle_ori"]) config, goal, obstacle_pos, obstacle_ori = map( np.array, [config, goal, obstacle_pos, obstacle_ori]) args = config, goal, obstacle_pos, obstacle_ori, seg["ac"], atarg fvpargs = [arr[::5] for arr in args] assign_old_eq_new( ) # set old parameter values to new parameter values with timed("computegrad"): *lossbefore, g = compute_lossandgrad(*args) lossbefore = allmean(np.array(lossbefore)) g = allmean(g) if np.allclose(g, 0): logger.log("Got zero gradient. not updating") else: with timed("cg"): '''stepdir0 = cg(fisher_vector_product, g, cg_iters=15, verbose=rank == 0) print('iter:10, norm of g: {:.4f}, error of cg: {:.4f}'.format(np.linalg.norm(g), np.linalg.norm( g - compute_fvp(stepdir0, *fvpargs))))''' stepdir = cg(fisher_vector_product, g, cg_iters=cg_iters, verbose=rank == 0) print('iter:{:d}, norm of g: {:.4f}, error of cg: {:.4f}'. format( cg_iters, np.linalg.norm(g), np.linalg.norm(g - compute_fvp(stepdir, *fvpargs)))) '''stepdir2 = cg(fisher_vector_product, g, cg_iters=200, verbose=rank == 0) print('iter:200, norm of g: {:.4f}, error of cg: {:.4f}'.format(np.linalg.norm(g), np.linalg.norm( g - compute_fvp(stepdir2, *fvpargs))))''' assert np.isfinite(stepdir).all() shs = .5 * stepdir.dot(fisher_vector_product(stepdir)) lm = np.sqrt(shs / max_kl) # logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g)) fullstep = stepdir / lm expectedimprove = g.dot(fullstep) surrbefore = lossbefore[0] stepsize = 1.0 thbefore = get_flat() for _ in range(10): thnew = thbefore + fullstep * stepsize set_from_flat(thnew) meanlosses = surr, kl, *_ = allmean( np.array(compute_losses(*args))) improve = surr - surrbefore logger.log("Expected: %.3f Actual: %.3f" % (expectedimprove, improve)) if not np.isfinite(meanlosses).all(): logger.log("Got non-finite value of losses -- bad!") elif kl > max_kl * 1.5: logger.log("violated KL constraint. shrinking step.") elif improve < 0: logger.log("surrogate didn't improve. shrinking step.") else: logger.log("Stepsize OK!") break stepsize *= .5 else: logger.log("couldn't compute a good step") set_from_flat(thbefore) if nworkers > 1 and iters_so_far % 20 == 0: paramsums = MPI.COMM_WORLD.allgather( (thnew.sum(), vfadam.getflat().sum())) # list of tuples assert all( np.allclose(ps, paramsums[0]) for ps in paramsums[1:]) with timed("vf"): for _ in range(vf_iters): for (mbob, mbg, mbop, mboo, mbret) in dataset.iterbatches( (config, goal, obstacle_pos, obstacle_ori, seg["tdlamret"]), include_final_partial_batch=False, batch_size=128): if hasattr(pi, "ob_rms"): pi.ob_rms.update( mbob) # update running mean/std for policy g = allmean( compute_vflossandgrad(mbob, mbg, mbop, mboo, mbret)) vfadam.update(g, vf_stepsize) g_losses = meanlosses for (lossname, lossval) in zip(loss_names, meanlosses): logger.record_tabular(lossname, lossval) logger.record_tabular("ev_tdlam_before", explained_variance(vpredbefore, tdlamret)) #mean = pi.pd.mean.eval() #print(mean) # ------------------ Update D ------------------ logger.log("Optimizing Discriminator...") logger.log(fmt_row(13, reward_giver.loss_name)) ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob)) batch_size = len(ob) // d_step d_losses = [ ] # list of tuples, each of which gives the loss for a minibatch dof = env.env.env.dof for ob_batch, goal_batch, obs_pos_batch, obs_ori_batch, ac_batch in dataset.iterbatches( (config, goal, obstacle_pos, obstacle_ori, ac), include_final_partial_batch=False, batch_size=batch_size): ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob_batch)) # update running mean/std for reward_giver if hasattr(reward_giver, "obs_rms"): reward_giver.obs_rms.update( np.concatenate((ob_batch, ob_expert), 0)) *newlosses, g = reward_giver.lossandgrad( ob_batch, goal_batch, obs_pos_batch, obs_ori_batch, ac_batch, ob_expert[:, :dof], ob_expert[:, dof:2 * dof], ob_expert[:, -6:-3], ob_expert[:, -3:], ac_expert) d_adam.update(allmean(g), d_stepsize) d_losses.append(newlosses) logger.log(fmt_row(13, np.mean(d_losses, axis=0))) lrlocal = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"] ) # local values listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples lens, rews, true_rets = map(flatten_lists, zip(*listoflrpairs)) true_rewbuffer.extend(true_rets) lenbuffer.extend(lens) rewbuffer.extend(rews) logger.record_tabular("EpLenMean", np.mean(lenbuffer)) logger.record_tabular("EpRewMean", np.mean(rewbuffer)) logger.record_tabular("EpTrueRewMean", np.mean(true_rewbuffer)) logger.record_tabular("EpThisIter", len(lens)) episodes_so_far += len(lens) timesteps_so_far += sum(lens) iters_so_far += 1 logger.record_tabular("EpisodesSoFar", episodes_so_far) logger.record_tabular("TimestepsSoFar", timesteps_so_far) logger.record_tabular("TimeElapsed", time.time() - tstart) if rank == 0: logger.dump_tabular()
def learn( *, env_type, env, eval_env, plotter_env, total_timesteps, num_cpu, allow_run_as_root, bind_to_core, seed=None, save_interval=5, clip_return=True, override_params=None, load_path=None, save_path=None, policy_pkl=None, ): rank = MPI.COMM_WORLD.Get_rank() logger.info('before mpi_fork: rank', rank, 'num_cpu', MPI.COMM_WORLD.Get_size()) if num_cpu > 1: if allow_run_as_root: whoami = mpi_fork_run_as_root(num_cpu, bind_to_core=bind_to_core) else: whoami = mpi_fork(num_cpu, bind_to_core=bind_to_core) if whoami == 'parent': logger.info('parent exiting with code 0...') sys.exit(0) U.single_threaded_session().__enter__() rank = MPI.COMM_WORLD.Get_rank() num_cpu = MPI.COMM_WORLD.Get_size() logger.info('after mpi_fork: rank', rank, 'num_cpu', num_cpu) override_params = override_params or {} # Seed everything. rank_seed = seed + 1000000 * rank if seed is not None else None set_global_seeds(rank_seed) # Prepare params. params = config.DEFAULT_PARAMS env_name = env.spec.id params['env_name'] = env_name if env_name in config.DEFAULT_ENV_PARAMS: params.update(config.DEFAULT_ENV_PARAMS[env_name] ) # merge env-specific parameters in params.update( **override_params) # makes it possible to override any parameter params['rollout_batch_size'] = env.num_envs params['num_cpu'] = num_cpu params['env_type'] = env_type with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f: json.dump(params, f) params = config.prepare_ve_params(params) dims = config.configure_dims(params) policy, value_ensemble, sample_disagreement_goals_fun, sample_uniform_goals_fun = \ config.configure_ve_ddpg(dims=dims, params=params, clip_return=clip_return, policy_pkl=policy_pkl) if policy_pkl is not None: env.set_sample_goals_fun(sample_dummy_goals_fun) else: env.envs_op("update_goal_sampler", goal_sampler=sample_disagreement_goals_fun) eval_env.envs_op("update_goal_sampler", goal_sampler=sample_uniform_goals_fun) if plotter_env is not None: plotter_env.envs_op("update_goal_sampler", goal_sampler=sample_uniform_goals_fun) if load_path is not None: tf_util.load_variables( os.path.join(load_path, 'final_policy_params.joblib')) return play(env=env, policy=policy) rollout_params, eval_params, plotter_params = config.configure_rollout_worker_params( params) rollout_worker = RolloutWorker(env, policy, dims, logger, monitor=True, **rollout_params) n_cycles = params['n_cycles'] n_epochs = total_timesteps // n_cycles // rollout_worker.T // rollout_worker.rollout_batch_size params['n_epochs'] = n_epochs params[ 'total_timesteps'] = n_epochs * n_cycles * rollout_worker.T * rollout_worker.rollout_batch_size config.log_params(params, logger=logger) if policy_pkl is not None: train_fun = train_ve evaluator = None else: train_fun = train # construct evaluator # assert eval_env.sample_goals_fun is None # eval_env.set_sample_goals_fun(sample_dummy_goals_fun) evaluator = RolloutWorker(eval_env, policy, dims, logger, **eval_params) if plotter_env is not None: raise NotImplementedError # from baselines.misc.html_report import HTMLReport # plotter_worker = RolloutWorker(plotter_env, policy, dims, logger, **plotter_params) # rank = MPI.COMM_WORLD.Get_rank() # report = HTMLReport(os.path.join(save_path, f'report-{rank}.html'), images_per_row=8) # # # report.add_header("{}".format(EXPERIMENT_TYPE)) # # report.add_text(format_dict(v)) # plotter = config.configure_plotter(policy, value_ensemble, plotter_worker, params, report) else: plotter = None return train_fun(save_path=save_path, policy=policy, value_ensemble=value_ensemble, rollout_worker=rollout_worker, evaluator=evaluator, n_epochs=n_epochs, n_test_rollouts=params['n_test_rollouts'], n_cycles=params['n_cycles'], n_batches=params['n_batches'], ve_n_batches=params['ve_n_batches'], save_interval=save_interval, plotter=plotter)
def learn(env, network, seed=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, # 每 train_freq step 更新一次模型 train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, # 开始训练前需要收集多少transition的信息 learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, **network_kwargs ): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. batch_size: int size of a batch sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) print("deepq.py network parameter", network) print("deepq.py network_kwargs parameter", network_kwargs) # q_func 得到对每个动作的评分 q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( # 输入的 observation_space 表示为 batch * obs.shape * one-hot_dim make_obs_ph=make_obs_ph, q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() reset = True with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, _ = env.step(env_action) # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0.0) reset = True if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len(episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) load_variables(model_file) return act
def learn(*, network, env, total_timesteps, seed=None, eval_env=None, replay_strategy='future', policy_save_interval=5, clip_return=True, demo_file=None, override_params=None, load_path=None, save_path=None, **kwargs ): override_params = override_params or {} if MPI is not None: rank = MPI.COMM_WORLD.Get_rank() num_cpu = MPI.COMM_WORLD.Get_size() # Seed everything. rank_seed = seed + 1000000 * rank if seed is not None else None set_global_seeds(rank_seed) # Prepare params. params = config.DEFAULT_PARAMS env_name = env.spec.id params['env_name'] = env_name params['replay_strategy'] = replay_strategy if env_name in config.DEFAULT_ENV_PARAMS: params.update(config.DEFAULT_ENV_PARAMS[env_name]) # merge env-specific parameters in params.update(**override_params) # makes it possible to override any parameter with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f: json.dump(params, f) params = config.prepare_params(params) params['rollout_batch_size'] = env.num_envs if demo_file is not None: params['bc_loss'] = 1 params.update(kwargs) config.log_params(params, logger=logger) if num_cpu == 1: logger.warn() logger.warn('*** Warning ***') logger.warn( 'You are running HER with just a single MPI worker. This will work, but the ' + 'experiments that we report in Plappert et al. (2018, https://arxiv.org/abs/1802.09464) ' + 'were obtained with --num_cpu 19. This makes a significant difference and if you ' + 'are looking to reproduce those results, be aware of this. Please also refer to ' + 'https://github.com/openai/baselines/issues/314 for further details.') logger.warn('****************') logger.warn() dims = config.configure_dims(params) policy = config.configure_ddpg(dims=dims, params=params, clip_return=clip_return) if load_path is not None: tf_util.load_variables(load_path) rollout_params = { 'exploit': False, 'use_target_net': False, 'use_demo_states': True, 'compute_Q': False, 'T': params['T'], } eval_params = { 'exploit': True, 'use_target_net': params['test_with_polyak'], 'use_demo_states': False, 'compute_Q': True, 'T': params['T'], } for name in ['T', 'rollout_batch_size', 'gamma', 'noise_eps', 'random_eps']: rollout_params[name] = params[name] eval_params[name] = params[name] eval_env = eval_env or env rollout_worker = RolloutWorker(env, policy, dims, logger, monitor=True, **rollout_params) evaluator = RolloutWorker(eval_env, policy, dims, logger, **eval_params) n_cycles = params['n_cycles'] n_epochs = total_timesteps // n_cycles // rollout_worker.T // rollout_worker.rollout_batch_size return train( save_path=save_path, policy=policy, rollout_worker=rollout_worker, evaluator=evaluator, n_epochs=n_epochs, n_test_rollouts=params['n_test_rollouts'], n_cycles=params['n_cycles'], n_batches=params['n_batches'], policy_save_interval=policy_save_interval, demo_file=demo_file)
def load(self, path): load_variables(path)
def load(self, load_path, extra_vars=None): tf_util.load_variables(load_path, extra_vars=extra_vars) #, sess=self.sess
def evalModel(self, toteps=1, err_thresh=.1): # Timestep Data actions = [] observations = [] rewards = [] errors = [] desireds = [] actuals = [] # Episode Data episodes = [] ep_num = [] # Set up parameters for quick initialization tp = tg.trainParams() tp.num_timesteps = 1 tp.timesteps_per_actorbatch = 1000 tp.optim_epochs = 1 tp.optim_batchsize = 1 tp.seed = 17 pp = tg.policyParams() # Initialize and load model if tp.model_path: tp.model_path = None # prevent override of model with U.tf.Graph().as_default(): # Allow Re-running of tf pi = tg.train(tp, pp, self.env_id) # Load Model tp.modelName(self.model_name) # Set up name self.model_dir = tp.model_dir # Save extracted model dir self.model_path = tp.model_path # Save extracted model path U.load_variables(tp.model_path) # Make Training Log # self.train_log = TrainLog(tp.model_dir) # Setup gym env = gym.make(self.env_id) # Seed Set rank = MPI.COMM_WORLD.Get_rank() workerseed = tp.seed + 1000000 * rank env.seed(workerseed) env.reset() ob = env.reset() # reset object for pi print('----------=================--------------') print('rank: ', rank, 'workerseed: ', workerseed) print('----------=================--------------') #env.render() #input('Press enter to continue') for eps in range (toteps): print(eps) #action = rand_action(env) action = pi.act(stochastic=False, ob=ob)[0] ob, r, done, info = env.step(action) # Initialize records ittr = 0 ep_data={} headers = ['stats', 'actions', 'observations', 'rewards', 'roll_err', 'pitch_err', 'yaw_err', 'droll_v', 'dpitch_v', 'dyaw_v', 'aroll_v', 'apitch_v', 'ayaw_v'] for header in headers: ep_data[header] = [] # Run Environment while done != True: #action = rand_action(env) # action = env.action_space.sample() # Random action for ctrl action = pi.act(stochastic=False, ob=ob)[0] # choose action ob, r, done, info = env.step(action) # perform action des = env.omega_target # desired angular velocities actual = env.omega_actual # current angular velocities # Record Data ep_data['actions'].append(action) ep_data['observations'].append(ob) ep_data['rewards'].append(r) # Errors ep_data['roll_err'].append(abs(ob[0])) ep_data['pitch_err'].append(abs(ob[1])) ep_data['yaw_err'].append(abs(ob[2])) # Step functions ep_data['droll_v'].append(env.omega_target[0]) ep_data['dpitch_v'].append(env.omega_target[1]) ep_data['dyaw_v'].append(env.omega_target[2]) ep_data['aroll_v'].append(env.omega_actual[0]) ep_data['apitch_v'].append(env.omega_actual[1]) ep_data['ayaw_v'].append(env.omega_actual[2]) ittr += 1 episodes.append(ep_data) env.reset() ep_num.append(eps) env.close() self.eps = episodes self.procEval()
import random if not ITS_THE_REAL_DEAL: benchmark.benchmark("imports") with open("model_params.pkl", "rb") as f: learn_params = pkl.load(f) env, policy, nenvs, ob_space, ac_space, nstack, model = create_model( **learn_params) params = {} with open("params.json") as f: params = json.load(f) if not ITS_THE_REAL_DEAL: benchmark.benchmark("create model") load_variables("actor.ckpt") if not ITS_THE_REAL_DEAL: benchmark.benchmark("load weights") devnull.close() sys.stdout = oldstdout from replay_parser import gen_obs """ <<<Game Begin>>> """ # This game object contains the initial game state. game = hlt.Game() # At this point "game" variable is populated with initial map data. # This is a good place to do computationally expensive start-up pre-processing. # As soon as you call "ready" function below, the 2 second per turn timer will start. game.ready("MyPythonBot")
def learn(env, policy_func, reward_giver, expert_dataset, rank, pretrained, pretrained_weight, *, g_step, d_step, entcoeff, save_per_iter, ckpt_dir, timesteps_per_batch, task_name, gamma, lam, max_kl, cg_iters, cg_damping=1e-2, vf_stepsize=3e-4, d_stepsize=3e-4, vf_iters=3, max_timesteps=0, max_episodes=0, max_iters=0, callback=None): nworkers = MPI.COMM_WORLD.Get_size() rank = MPI.COMM_WORLD.Get_rank() np.set_printoptions(precision=3) # Setup losses and stuff # ---------------------------------------- ob_space = env.observation_space ac_space = env.action_space pi = policy_func("pi", ob_space, ac_space, reuse=(pretrained_weight != None)) oldpi = policy_func("oldpi", ob_space, ac_space) atarg = tf.placeholder( dtype=tf.float32, shape=[None]) # Target advantage function (if applicable) ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return ob = U.get_placeholder_cached(name="ob") ac = pi.pdtype.sample_placeholder([None]) kloldnew = oldpi.pd.kl(pi.pd) ent = pi.pd.entropy() meankl = tf.reduce_mean(kloldnew) meanent = tf.reduce_mean(ent) entbonus = entcoeff * meanent vferr = tf.reduce_mean(tf.square(pi.vpred - ret)) ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # advantage * pnew / pold surrgain = tf.reduce_mean(ratio * atarg) optimgain = surrgain + entbonus losses = [optimgain, meankl, entbonus, surrgain, meanent] loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"] dist = meankl all_var_list = pi.get_trainable_variables() var_list = [ v for v in all_var_list if v.name.startswith("pi/pol") or v.name.startswith("pi/logstd") ] vf_var_list = [v for v in all_var_list if v.name.startswith("pi/vff")] assert len(var_list) == len(vf_var_list) + 1 d_adam = MpiAdam(reward_giver.get_trainable_variables()) vfadam = MpiAdam(vf_var_list) get_flat = U.GetFlat(var_list) set_from_flat = U.SetFromFlat(var_list) klgrads = tf.gradients(dist, var_list) flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan") shapes = [var.get_shape().as_list() for var in var_list] start = 0 tangents = [] for shape in shapes: sz = U.intprod(shape) tangents.append(tf.reshape(flat_tangent[start:start + sz], shape)) start += sz gvp = tf.add_n([ tf.reduce_sum(g * tangent) for (g, tangent) in zipsame(klgrads, tangents) ]) # pylint: disable=E1111 fvp = U.flatgrad(gvp, var_list) assign_old_eq_new = U.function( [], [], updates=[ tf.assign(oldv, newv) for (oldv, newv) in zipsame(oldpi.get_variables(), pi.get_variables()) ]) compute_losses = U.function([ob, ac, atarg], losses) compute_lossandgrad = U.function([ob, ac, atarg], losses + [U.flatgrad(optimgain, var_list)]) compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp) compute_vflossandgrad = U.function([ob, ret], U.flatgrad(vferr, vf_var_list)) def allmean(x): assert isinstance(x, np.ndarray) out = np.empty_like(x) MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM) out /= nworkers return out U.initialize() th_init = get_flat() MPI.COMM_WORLD.Bcast(th_init, root=0) set_from_flat(th_init) d_adam.sync() vfadam.sync() if rank == 0: print("Init param sum", th_init.sum(), flush=True) # Prepare for rollouts # ---------------------------------------- seg_gen = traj_segment_generator(pi, env, reward_giver, timesteps_per_batch, stochastic=True) episodes_so_far = 0 timesteps_so_far = 0 iters_so_far = 0 tstart = time.time() lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards true_rewbuffer = deque(maxlen=40) test_true_rewbuffer = deque(maxlen=40) test_lenbuffer = deque(maxlen=40) assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1 # g_loss_stats = stats(loss_names) # d_loss_stats = stats(reward_giver.loss_name) ep_stats = stats(["True_rewards", "Rewards", "Episode_length"]) # if provide pretrained weight if pretrained_weight is not None: U.load_variables(pretrained_weight, variables=pi.get_variables()) best = -2000 save_ind = 0 max_save = 3 while True: if callback: callback(locals(), globals()) # if max_timesteps and timesteps_so_far >= max_timesteps: # break # elif max_episodes and episodes_so_far >= max_episodes: # break # elif max_iters and iters_so_far >= max_iters: # break # if max_iters and iters_so_far >= max_iters: break logger.log("********** Iteration %i ************" % iters_so_far) def fisher_vector_product(p): return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p # ------------------ Update G ------------------ logger.log("Optimizing Policy...") for _ in range(g_step): seg = seg_gen.__next__() if (iters_so_far == 0) or ((iters_so_far + 1) % 60 == 0): obs_list = [] acs_list = [] len_list = [] ret_list = [] for _ in tqdm(range(50)): from run_expert import traj_1_generator traj = traj_1_generator(pi, env, 1000, stochastic=False) obs, acs, ep_len, ep_ret = traj['ob'], traj['ac'], traj[ 'ep_len'], traj['ep_ret'] obs_list.append(obs) acs_list.append(acs) len_list.append(ep_len) ret_list.append(ep_ret) avg_len = np.mean(len_list) avg_ret = np.mean(ret_list) std_ret = np.std(ret_list) logger.record_tabular("TestEpTrueRewMean", avg_ret) logger.record_tabular("TestEpTrueRewStd", std_ret) else: logger.record_tabular("TestEpTrueRewMean", -1) logger.record_tabular("TestEpTrueRewStd", -1) #report stats and save policy if any good lrlocal = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"] ) # local values listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples lens, rews, true_rets = map(flatten_lists, zip(*listoflrpairs)) true_rewbuffer.extend(true_rets) lenbuffer.extend(lens) rewbuffer.extend(rews) true_rew_avg = np.mean(true_rewbuffer) true_rew_std = np.std(true_rewbuffer) logger.record_tabular("EpLenMean", np.mean(lenbuffer)) logger.record_tabular("EpRewMean", np.mean(rewbuffer)) logger.record_tabular("EpRewStd", np.std(rewbuffer)) logger.record_tabular("EpTrueRewMean", true_rew_avg) logger.record_tabular("EpTrueStd", true_rew_std) logger.record_tabular("EpThisIter", len(lens)) episodes_so_far += len(lens) timesteps_so_far += sum(lens) iters_so_far += 1 logger.record_tabular("EpisodesSoFar", episodes_so_far) logger.record_tabular("TimestepsSoFar", timesteps_so_far) logger.record_tabular("TimeElapsed", time.time() - tstart) logger.record_tabular("Best so far", best) # # Save model # if ckpt_dir is not None and true_rew_avg >= best and len(true_rewbuffer) > 30: # best = true_rew_avg # fname = os.path.join(ckpt_dir, task_name) # os.makedirs(os.path.dirname(fname), exist_ok=True) # pi.save_policy(fname) # Save model if ckpt_dir is not None: fname = os.path.join(ckpt_dir, task_name) os.makedirs(os.path.dirname(fname), exist_ok=True) if true_rew_avg >= best: best = true_rew_avg pi.save_policy(fname + "_" + str(save_ind)) pi.save_policy(fname + "_best") save_ind = (save_ind + 1) % max_save if (iters_so_far + 1) % 1000 == 0: pi.save_policy(fname + "_iter_" + str(iters_so_far + 1)) #compute gradient towards next policy add_vtarg_and_adv(seg, gamma, lam) # ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets)) ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[ "tdlamret"] vpredbefore = seg[ "vpred"] # predicted value function before udpate atarg = (atarg - atarg.mean()) / atarg.std( ) # standardized advantage function estimate if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy args = seg["ob"], seg["ac"], atarg fvpargs = [arr[::5] for arr in args] assign_old_eq_new( ) # set old parameter values to new parameter values *lossbefore, g = compute_lossandgrad(*args) lossbefore = allmean(np.array(lossbefore)) g = allmean(g) if np.allclose(g, 0): logger.log("Got zero gradient. not updating") else: stepdir = cg(fisher_vector_product, g, cg_iters=cg_iters, verbose=False) assert np.isfinite(stepdir).all() shs = .5 * stepdir.dot(fisher_vector_product(stepdir)) lm = np.sqrt(shs / max_kl) fullstep = stepdir / lm expectedimprove = g.dot(fullstep) surrbefore = lossbefore[0] stepsize = 1.0 thbefore = get_flat() for _ in range(10): thnew = thbefore + fullstep * stepsize set_from_flat(thnew) meanlosses = surr, kl, *_ = allmean( np.array(compute_losses(*args))) improve = surr - surrbefore logger.log("Expected: %.3f Actual: %.3f" % (expectedimprove, improve)) if not np.isfinite(meanlosses).all(): logger.log("Got non-finite value of losses -- bad!") elif kl > max_kl * 1.5: logger.log("violated KL constraint. shrinking step.") elif improve < 0: logger.log("surrogate didn't improve. shrinking step.") else: logger.log("Stepsize OK!") break stepsize *= .5 else: logger.log("couldn't compute a good step") set_from_flat(thbefore) if nworkers > 1 and iters_so_far % 20 == 0: paramsums = MPI.COMM_WORLD.allgather( (thnew.sum(), vfadam.getflat().sum())) # list of tuples assert all( np.allclose(ps, paramsums[0]) for ps in paramsums[1:]) for _ in range(vf_iters): for (mbob, mbret) in dataset.iterbatches( (seg["ob"], seg["tdlamret"]), include_final_partial_batch=False, batch_size=128): if hasattr(pi, "ob_rms"): pi.ob_rms.update( mbob) # update running mean/std for policy g = allmean(compute_vflossandgrad(mbob, mbret)) vfadam.update(g, vf_stepsize) if rank == 0: logger.dump_tabular() g_losses = meanlosses for (lossname, lossval) in zip(loss_names, meanlosses): logger.record_tabular(lossname, lossval) logger.record_tabular("ev_tdlam_before", explained_variance(vpredbefore, tdlamret)) pi.save_policy(fname + "_converged") # ------------------ Update D ------------------ batch_size = len(ob) // d_step d_losses = [ ] # list of tuples, each of which gives the loss for a minibatch for ob_batch, ac_batch in dataset.iterbatches( (ob, ac), include_final_partial_batch=False, batch_size=batch_size): ob_expert, ac_expert = expert_dataset.next_batch(len(ob_batch)) # update running mean/std for reward_giver if hasattr(reward_giver, "obs_rms"): reward_giver.obs_rms.update( np.concatenate((ob_batch, ob_expert), 0)) *newlosses, g = reward_giver.lossandgrad(ob_batch, ac_batch, ob_expert, ac_expert) d_adam.update(allmean(g), d_stepsize) d_losses.append(newlosses) logger.log(fmt_row(13, reward_giver.loss_name)) logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
def learn(*, network, env, total_timesteps, seed=None, eval_env=None, replay_strategy='future', policy_save_interval=5, clip_return=True, demo_file=None, override_params=None, load_path=None, save_path=None, params=None, **kwargs): override_params = override_params or {} if MPI is not None: rank = MPI.COMM_WORLD.Get_rank() num_cpu = MPI.COMM_WORLD.Get_size() # Seed everything. rank_seed = seed + 1000000 * rank if seed is not None else None set_global_seeds(rank_seed) # Prepare params. params = { # env 'max_u': 1., # max absolute value of actions on different coordinates # ddpg 'layers': 3, # number of layers in the critic/actor networks 'hidden': 256, # number of neurons in each hidden layers 'network_class': 'baselines.her.actor_critic:ActorCritic', 'Q_lr': 0.001, # critic learning rate 'pi_lr': 0.001, # actor learning rate 'buffer_size': int(1E6), # for experience replay 'polyak': 0.95, # polyak averaging coefficient 'action_l2': 1.0, # quadratic penalty on actions (before rescaling by max_u) 'clip_obs': 200., 'scope': 'ddpg', # can be tweaked for testing 'relative_goals': False, # training 'n_cycles': 50, # per epoch 'rollout_batch_size': 2, # per mpi thread 'n_batches': 40, # training batches per cycle 'batch_size': 256, # per mpi thread, measured in transitions and reduced to even multiple of chunk_length. 'n_test_rollouts': 10, # number of test rollouts per epoch, each consists of rollout_batch_size rollouts 'test_with_polyak': False, # run test episodes with the target network # exploration 'random_eps': 0.2, # percentage of time a random action is taken 'noise_eps': 0.3, # std of gaussian noise added to not-completely-random actions as a percentage of max_u # HER 'replay_strategy': 'future', # supported modes: future, none 'replay_k': 4, # number of additional goals used for replay, only used if off_policy_data=future # normalization 'norm_eps': 0.01, # epsilon used for observation normalization 'norm_clip': 5, # normalized observations are cropped to this values 'bc_loss': 0, # whether or not to use the behavior cloning loss as an auxilliary loss 'q_filter': 0, # whether or not a Q value filter should be used on the Actor outputs 'num_demo': 25, # number of expert demo episodes 'demo_batch_size': 128, #number of samples to be used from the demonstrations buffer, per mpi thread 128/1024 or 32/256 'prm_loss_weight': 0.001, #Weight corresponding to the primary loss 'aux_loss_weight': 0.0078, #Weight corresponding to the auxilliary loss also called the cloning loss 'perturb': kwargs['pert_type'], 'n_actions': kwargs['n_actions'], } params['replay_strategy'] = replay_strategy if env is not None: env_name = env.spec.id params['env_name'] = env_name if env_name in config.DEFAULT_ENV_PARAMS: params.update(config.DEFAULT_ENV_PARAMS[env_name] ) # merge env-specific parameters in else: params['env_name'] = 'NuFingers_Experiment' params.update( **override_params) # makes it possible to override any parameter with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f: json.dump(params, f) if demo_file is not None: params['bc_loss'] = 1 params['q_filter'] = 1 params['n_cycles'] = 20 params['random_eps'] = 0.1 # chip: ON params['noise_eps'] = 0.1 # chip: ON # params['batch_size']: 1024 params = config.prepare_params(params) params['rollout_batch_size'] = 1 params.update(kwargs) config.log_params(params, logger=logger) if num_cpu == 1: logger.warn() logger.warn('*** Warning ***') logger.warn( 'You are running HER with just a single MPI worker. This will work, but the ' + 'experiments that we report in Plappert et al. (2018, https://arxiv.org/abs/1802.09464) ' + 'were obtained with --num_cpu 19. This makes a significant difference and if you ' + 'are looking to reproduce those results, be aware of this. Please also refer to ' + 'https://github.com/openai/baselines/issues/314 for further details.' ) logger.warn('****************') logger.warn() if env is not None: dims = config.configure_dims(params) else: dims = dict(o=15, u=4, g=7, info_is_success=1) policy = config.configure_ddpg(dims=dims, params=params, clip_return=clip_return) if load_path is not None: tf_util.load_variables(load_path) rollout_params = { 'exploit': False, 'use_target_net': False, 'use_demo_states': True, 'compute_Q': False, 'T': params['T'], } eval_params = { 'exploit': True, 'use_target_net': params['test_with_polyak'], 'use_demo_states': False, 'compute_Q': True, 'T': params['T'], } for name in [ 'T', 'rollout_batch_size', 'gamma', 'noise_eps', 'random_eps' ]: rollout_params[name] = params[name] eval_params[name] = params[name] eval_env = eval_env or env print("NAME={}".format(params['env_name'])) print(rollout_params) if params['env_name'].find('NuFingers_Experiment') == -1: rollout_worker = RolloutWorker(env, policy, dims, logger, monitor=True, **rollout_params) evaluator = RolloutWorker(eval_env, policy, dims, logger, **eval_params) else: rollout_worker = RolloutNuFingers(policy, dims, logger, monitor=True, **rollout_params) evaluator = RolloutNuFingers(policy, dims, logger, **eval_params) n_cycles = params['n_cycles'] n_epochs = total_timesteps // n_cycles // rollout_worker.T // rollout_worker.rollout_batch_size return train(save_path=save_path, policy=policy, rollout_worker=rollout_worker, evaluator=evaluator, n_epochs=n_epochs, n_test_rollouts=params['n_test_rollouts'], n_cycles=params['n_cycles'], n_batches=params['n_batches'], policy_save_interval=policy_save_interval, demo_file=demo_file)
def load(self, load_path): tf_util.load_variables(load_path, sess=self.sess)
def learn(env, network, seed=None, lr=5e-4, expert_lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.5, initial_exploration_p=1.0, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1, gamma=1.0, target_network_update_freq=100, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, double_q=True, obs_dim=None, **network_kwargs ): """Train a bootstrap-dqn model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) nenvs = env.num_envs print("Bootstrap DQN with {} envs".format(nenvs)) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph # import IPython; IPython.embed() #assert isinstance(env.envs[0].env.env.env, ExplicitBayesEnv) #belief_space = env.envs[0].env.env.env.belief_space #observation_space = env.envs[0].env.env.env.internal_observation_space obs_space = env.observation_space assert obs_dim is not None observation_space = Box(obs_space.low[:obs_dim], obs_space.high[:obs_dim], dtype=np.float32) belief_space = Box(obs_space.low[obs_dim:], obs_space.high[obs_dim:], dtype=np.float32) num_experts = belief_space.high.size print("Num experts", num_experts) def make_obs_ph(name): return ObservationInput(observation_space, name=name) def make_bel_ph(name): return ObservationInput(belief_space, name=name) q_func = build_q_func(network, num_experts, **network_kwargs) print('=============== got qfunc ============== ') act, train, update_target, debug = residual_bqn_separate_expert.build_train( make_obs_ph=make_obs_ph, make_bel_ph=make_bel_ph, q_func=q_func, num_experts=num_experts, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), expert_optimizer=tf.train.AdamOptimizer(learning_rate=expert_lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise, double_q=double_q ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size, num_experts) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=initial_exploration_p, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() episode_reward = np.zeros(nenvs, dtype = np.float32) saved_mean_reward = None reset = True epoch_episode_rewards = [] epoch_episode_steps = [] epoch_actions = [] epoch_episodes = 0 episode_rewards_history = deque(maxlen=100) episode_step = np.zeros(nenvs, dtype = int) episodes = 0 #scalar with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") print("Model will be saved at " , model_file) model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) print('Loaded model from {}'.format(load_path)) t = 0 while t < total_timesteps: if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} update_eps = exploration.value(t) update_param_noise_threshold = 0. obs = env.reset() obs, bel = obs[:, :-belief_space.shape[0]], obs[:, -belief_space.shape[0]:] for m in range(100): action, q_values, expert_q_values = act(np.array(obs)[None], np.array(bel)[None], update_eps=update_eps, **kwargs) env_action = action new_obs, rew, done, info = env.step(env_action) new_obs, new_bel = new_obs[:, :-belief_space.shape[0]], new_obs[:, -belief_space.shape[0]:] expert = np.array([_info['expert'] for _info in info]) # Store transition in the replay buffer. replay_buffer.add(obs, bel, action, rew, new_obs, new_bel, done, expert) if np.random.rand() < 0.01: # write to file with open('tiger_rbqn_sep_exp.csv', 'a') as f: out = str(expert[0]) + ',' + ','.join(str(np.around(x,1)) for x in [bel[0], obs[0], q_values[0], expert_q_values[:, 0].ravel()]) f.write(out + "\n") print(out) obs = new_obs bel = new_bel episode_reward += rew episode_step += 1 for d in range(len(done)): if done[d]: epoch_episode_rewards.append(episode_reward[d]) episode_rewards_history.append(episode_reward[d]) epoch_episode_steps.append(episode_step[d]) episode_reward[d] = 0. episode_step[d] = 0 epoch_episodes += 1 episodes += 1 episode_step[d] = 0 epoch_episodes += 1 episodes += 1 t += 100 * nenvs # import IPython; IPython.embed(); import sys; sys.exit(0) if t > learning_starts and t % train_freq == 0: # for _ in range(5): # expert_i = np.random.choice(num_experts) for expert_i in range(num_experts): # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t), expert=expert_i) if experience is None: continue obses_t, bels_t, actions, rewards, obses_tp1, bels_tp1, dones, weights, batch_idxes = experience else: experience = replay_buffer.sample(batch_size, expert=expert_i) if experience is None: continue obses_t, bels_t, actions, rewards, obses_tp1, bels_tp1, dones, exps = experience weights, batch_idxes = np.ones_like(rewards), None assert np.all(exps == expert_i) td_errors, expert_td_errors = train(obses_t, bels_t, actions, rewards, obses_tp1, bels_tp1, dones, weights, expert_i) if np.random.rand() < 0.01: print("TD error", td_errors, expert_td_errors) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) """ obses_t, bels_t, actions, rewards, obses_tp1, bels_tp1, dones, exps = replay_buffer.sample(batch_size, expert=None) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, bels_t, actions, rewards, obses_tp1, bels_tp1, dones, weights, np.array([0])) """ if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards_history), 2) num_episodes = episodes if print_freq is not None: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() print("episodes ", num_episodes, "steps {}/{}".format(t, total_timesteps)) print("mean reward", mean_100ep_reward) print("% time spent exploring", int(100 * exploration.value(t))) if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: # if print_freq is not None: print("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) print("saving model") save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) load_variables(model_file) return act
def learn(env, network, seed=None, lr=1e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, multiplayer=False, callback=None, load_path=None, load_path_1=None, load_path_2=None, **network_kwargs): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # This was all handled in not the most elegant way # Variables have a _1 or _2 appended to them to separate them # and a bunch of if statementss to have the _2 variables not do anything in single-player # when in multiplayer Space Invaders, need to not reward players for other player dying isSpaceInvaders = False if "SpaceInvaders" in str(env): isSpaceInvaders = True # put a limit on the amount of memory used, otherwise TensorFlow will consume nearly everything # this leaves 1 GB free on my computer, others may need to change it # Create all the functions necessary to train the model # Create two separate TensorFlow sessions graph_1 = tf.Graph() sess_1 = tf.Session(graph=graph_1) if multiplayer: graph_2 = tf.Graph() sess_2 = tf.Session(graph=graph_2) else: # set session 2 to None if it's not being used sess_2 = None set_global_seeds(seed) # specify the q functions as separate objects q_func_1 = build_q_func(network, **network_kwargs) if multiplayer: q_func_2 = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) # build everything for the first model # pass in the session and the "_1" suffix act_1, train_1, update_target_1, debug_1 = deepq.build_train( sess=sess_1, make_obs_ph=make_obs_ph, q_func=q_func_1, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise, scope="deepq") # a lot of if multiplayer statements duplicating these actions for a second network # pass in session 2 and "_2" instead if multiplayer: act_2, train_2, update_target_2, debug_2 = deepq.build_train( sess=sess_2, make_obs_ph=make_obs_ph, q_func=q_func_2, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise, scope="deepq") # separate act_params for each wrapper act_params_1 = { 'make_obs_ph': make_obs_ph, 'q_func': q_func_1, 'num_actions': env.action_space.n, } if multiplayer: act_params_2 = { 'make_obs_ph': make_obs_ph, 'q_func': q_func_2, 'num_actions': env.action_space.n, } # make the act wrappers act_1 = ActWrapper(act_1, act_params_1) if multiplayer: act_2 = ActWrapper(act_2, act_params_2) # I need to return something if it's single-player else: act_2 = None # Create the replay buffer # separate replay buffers are required for each network # this is required for competitive because the replay buffers hold rewards # and player 2 has different rewards than player 1 if prioritized_replay: replay_buffer_1 = PrioritizedReplayBuffer( buffer_size, alpha=prioritized_replay_alpha) if multiplayer: replay_buffer_2 = PrioritizedReplayBuffer( buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer_1 = ReplayBuffer(buffer_size) if multiplayer: replay_buffer_2 = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. # initialize both sessions U.initialize(sess_1) if multiplayer: U.initialize(sess_2) # the session was passed into these functions when they were created # the separate update functions work within the different sessions update_target_1() if multiplayer: update_target_2() # keep track of rewards for both models separately episode_rewards_1 = [0.0] saved_mean_reward_1 = None if multiplayer: episode_rewards_2 = [0.0] saved_mean_reward_2 = None obs = env.reset() reset = True # storing stuff in a temporary directory while it's working with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file_1 = os.path.join(td, "model_1") temp_file_1 = os.path.join(td, "temp_1") model_saved_1 = False if multiplayer: model_file_2 = os.path.join(td, "model_2") temp_file_2 = os.path.join(td, "temp_2") model_saved_2 = False if tf.train.latest_checkpoint(td) is not None: if multiplayer: # load both models if multiplayer is on load_variables(model_file_1, sess_1) logger.log('Loaded model 1 from {}'.format(model_file_1)) model_saved_1 = True load_variables(model_file_2, sess_2) logger.log('Loaded model 2 from {}'.format(model_file_2)) model_saved_2 = True # otherwise just load the first one else: load_variables(model_file_1, sess_1) logger.log('Loaded model from {}'.format(model_file_1)) model_saved_1 = True # I have separate load variables for single-player and multiplayer # this should be None if multiplayer is on elif load_path is not None: load_variables(load_path, sess_1) logger.log('Loaded model from {}'.format(load_path)) # load the separate models in for multiplayer # should load the variables into the appropriate sessions # my format may restrict things to working properly only when a Player 1 model is loaded into session 1, and same for Player 2 # however, in practice, the models won't work properly otherwise elif multiplayer: if load_path_1 is not None: load_variables(load_path_1, sess_1) logger.log('Loaded model 1 from {}'.format(load_path_1)) if load_path_2 is not None: load_variables(load_path_2, sess_2) logger.log('Loaded model 2 from {}'.format(load_path_2)) # actual training starts here for t in range(total_timesteps): # use this for updating purposes actual_t = t + 1 if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value( t) + exploration.value(t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True # receive model 1's action based on the model and observation action_1 = act_1(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action_1 = action_1 # do the same for model 2 if in multiplayer if multiplayer: action_2 = act_2(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action_2 = action_2 reset = False # apply actions to the environment if multiplayer: new_obs, rew_1, rew_2, done, _ = env.step( env_action_1, env_action_2) # apply single action if there isn't a second model else: new_obs, rew_1, rew_2, done, _ = env.step(env_action_1) # manual clipping for Space Invaders multiplayer if isSpaceInvaders and multiplayer: # don't reward a player when the other player dies # change the reward to 0 # the only time either player will get rewarded 200 is when the other player dies if rew_1 >= 200: rew_1 = rew_1 - 200.0 if rew_2 >= 200: rew_2 = rew_2 - 200.0 # manually clip the rewards using the sign function rew_1 = np.sign(rew_1) rew_2 = np.sign(rew_2) combo_factor = 0.25 rew_1_combo = rew_1 + combo_factor * rew_2 rew_2_combo = rew_2 + combo_factor * rew_1 rew_1 = rew_1_combo rew_2 = rew_2_combo # Store transition in the replay buffers replay_buffer_1.add(obs, action_1, rew_1, new_obs, float(done)) if multiplayer: # pass reward_2 to the second player # this reward will vary based on the game replay_buffer_2.add(obs, action_2, rew_2, new_obs, float(done)) obs = new_obs # separate rewards for each model episode_rewards_1[-1] += rew_1 if multiplayer: episode_rewards_2[-1] += rew_2 if done: obs = env.reset() episode_rewards_1.append(0.0) if multiplayer: episode_rewards_2.append(0.0) reset = True if actual_t > learning_starts and actual_t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. # sample from the two replay buffers if prioritized_replay: experience_1 = replay_buffer_1.sample( batch_size, beta=beta_schedule.value(t)) (obses_t_1, actions_1, rewards_1, obses_tp1_1, dones_1, weights_1, batch_idxes_1) = experience_1 # keep all the variables with separate names if multiplayer: experience_2 = replay_buffer_2.sample( batch_size, beta=beta_schedule.value(t)) (obses_t_2, actions_2, rewards_2, obses_tp1_2, dones_2, weights_2, batch_idxes_2) = experience_2 # do the same if there's no prioritization else: obses_t_1, actions_1, rewards_1, obses_tp1_1, dones_1 = replay_buffer_1.sample( batch_size) weights_1, batch_idxes_1 = np.ones_like(rewards_1), None if multiplayer: obses_t_2, actions_2, rewards_2, obses_tp1_2, dones_2 = replay_buffer_2.sample( batch_size) weights_2, batch_idxes_2 = np.ones_like( rewards_2), None # actually train the model based on the samples td_errors_1 = train_1(obses_t_1, actions_1, rewards_1, obses_tp1_1, dones_1, weights_1) if multiplayer: td_errors_2 = train_2(obses_t_2, actions_2, rewards_2, obses_tp1_2, dones_2, weights_2) # give new priority weights to the observations if prioritized_replay: new_priorities_1 = np.abs( td_errors_1) + prioritized_replay_eps replay_buffer_1.update_priorities(batch_idxes_1, new_priorities_1) if multiplayer: new_priorities_2 = np.abs( td_errors_2) + prioritized_replay_eps replay_buffer_2.update_priorities( batch_idxes_2, new_priorities_2) if actual_t > learning_starts and actual_t % target_network_update_freq == 0: # Update target networks periodically. update_target_1() if multiplayer: update_target_2() # this section is for the purposes of logging stuff # calculate the average reward over the last 100 episodes mean_100ep_reward_1 = round(np.mean(episode_rewards_1[-101:-1]), 1) if multiplayer: mean_100ep_reward_2 = round( np.mean(episode_rewards_2[-101:-1]), 1) num_episodes = len(episode_rewards_1) # every given number of episodes log and print out the appropriate stuff if done and print_freq is not None and len( episode_rewards_1) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) # print out both rewards if multiplayer if multiplayer: logger.record_tabular("mean 100 episode reward 1", mean_100ep_reward_1) logger.record_tabular("mean 100 episode reward 2", mean_100ep_reward_2) else: logger.record_tabular("mean 100 episode reward", mean_100ep_reward_1) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() # save best-performing version of each model # I've opted out of this for competitive multiplayer because it's difficult to determine what's "best" if (checkpoint_freq is not None and actual_t > learning_starts and num_episodes > 100 and actual_t % checkpoint_freq == 0): # if there's a best reward, save it as the new best model if saved_mean_reward_1 is None or mean_100ep_reward_1 > saved_mean_reward_1: if print_freq is not None: if multiplayer: logger.log( "Saving model 1 due to mean reward increase: {} -> {}" .format(saved_mean_reward_1, mean_100ep_reward_1)) else: logger.log( "Saving model due to mean reward increase: {} -> {}" .format(saved_mean_reward_1, mean_100ep_reward_1)) save_variables(model_file_1, sess_1) model_saved_1 = True saved_mean_reward_1 = mean_100ep_reward_1 if multiplayer and (saved_mean_reward_2 is None or mean_100ep_reward_2 > saved_mean_reward_2): if print_freq is not None: logger.log( "Saving model 2 due to mean reward increase: {} -> {}" .format(saved_mean_reward_2, mean_100ep_reward_2)) save_variables(model_file_2, sess_2) model_saved_2 = True saved_mean_reward_2 = mean_100ep_reward_2 # restore models at the end to the best performers if model_saved_1: if print_freq is not None: logger.log("Restored model 1 with mean reward: {}".format( saved_mean_reward_1)) load_variables(model_file_1, sess_1) if multiplayer and model_saved_2: if print_freq is not None: logger.log("Restored model 2 with mean reward: {}".format( saved_mean_reward_2)) load_variables(model_file_2, sess_2) return act_1, act_2, sess_1, sess_2
def learn( # env flags env, raw_env, use_2D_env=True, use_multiple_starts=False, use_rich_reward=False, total_timesteps=100000, # dqn network=identity_fn, exploration_fraction=0.1, exploration_final_eps=0.02, # hr use_feedback=False, use_real_feedback=False, only_use_hr_until=int(1e3), trans_to_rl_in=int(2e4), good_feedback_acc=0.7, bad_feedback_acc=0.7, # dqn training lr=5e-4, batch_size=32, dqn_epochs=3, train_freq=1, target_network_update_freq=500, learning_starts=1000, param_noise=True, gamma=1.0, # hr training feedback_lr=1e-3, feedback_epochs=4, feedback_batch_size=16, feedback_minibatch_size=8, min_feedback_buffer_size=32, feedback_training_prop=0.7, feedback_training_new_prop=0.4, # replay buffer buffer_size=50000, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, # rslts saving and others checkpoint_freq=10000, checkpoint_path=None, print_freq=100, load_path=None, callback=None, seed=0, **network_kwargs): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. batch_size: int size of a batch sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model # sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) hr_func = build_hr_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space observation_space.dtype = np.float32 def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train_rl, train_hr, evaluate_hr, update_target, debug = build_train( make_obs_ph=make_obs_ph, q_func=q_func, hr_func=hr_func, num_actions=env.action_space.n, rl_optimizer=tf.train.AdamOptimizer(learning_rate=lr), hr_optimizer=tf.train.AdamOptimizer(learning_rate=feedback_lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'hr_func': hr_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() obs, cor = obs['obs'], obs['nonviz_sensor'] reset = True if use_feedback and use_real_feedback: import pylsl print("looking for an EEG_Pred stream...", end="", flush=True) feedback_LSL_stream = pylsl.StreamInlet( pylsl.resolve_stream('type', 'EEG_Pred')[0]) print(" done") target_position = raw_env.robot.get_target_position() if use_2D_env: judge_action, *_ = run_dijkstra(raw_env, target_position) else: judge_action = judge_action_1D(raw_env, target_position) state_action_buffer = deque(maxlen=100) action_idx_buffer = deque(maxlen=100) feedback_buffer_train = [] feedback_buffer_valid = [] performance = {"feedback": [], "sparse_reward": [], "rich_reward": []} epi_feedback_test_num = 0 with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if use_feedback: update_rl_importance = (t - only_use_hr_until) / trans_to_rl_in update_rl_importance = np.clip(update_rl_importance, 0, 1) kwargs['update_rl_importance'] = update_rl_importance if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value( t) + exploration.value(t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs[ 'update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False raw_env.action_idx = t new_obs, rewards_dict, done, _ = env.step(env_action) new_obs, new_cor = new_obs['obs'], new_obs['nonviz_sensor'] sparse_reward = rewards_dict["sparse"] rich_reward = rewards_dict["rich"] rew = rich_reward if use_rich_reward else sparse_reward # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) state_action_buffer.append([obs, action]) action_idx_buffer.append(t) action_idxs, feedbacks, correct_feedbacks = \ get_simulated_feedback([cor] if use_2D_env else [obs], [action], [t], judge_action, good_feedback_acc, bad_feedback_acc) performance["feedback"].extend(correct_feedbacks) performance["sparse_reward"].append(sparse_reward) performance["rich_reward"].append(rich_reward) obs, cor = new_obs, new_cor if use_feedback: if use_real_feedback: feedbacks, action_idxs = get_feedback_from_LSL( feedback_LSL_stream) feedback_epi_buffer = [ state_action_buffer[action_idx_buffer.index(a_idx)] + [feedback] for a_idx, feedback in zip(action_idxs, feedbacks) ] # add feedbacks into feedback replay buffer if feedback_epi_buffer: epi_feedback_test_num += len(feedback_epi_buffer) * ( 1 - feedback_training_prop) epi_test_int = int(epi_feedback_test_num) epi_feedback_test_num -= epi_test_int epi_test_inds = np.random.choice(len(feedback_epi_buffer), epi_test_int, replace=False) epi_train_inds = [ ind for ind in range(len(feedback_epi_buffer)) if ind not in epi_test_inds ] feedback_buffer_train.extend( [feedback_epi_buffer[ind] for ind in epi_train_inds]) feedback_buffer_valid.extend( [feedback_epi_buffer[ind] for ind in epi_test_inds]) episode_rewards[-1] += rew if done: obs = env.reset() obs, cor = obs['obs'], obs['nonviz_sensor'] episode_rewards.append(0.0) reset = True if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. for _ in range(dqn_epochs): if prioritized_replay: experience = replay_buffer.sample( batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train_rl(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs( td_errors) + prioritized_replay_eps replay_buffer.update_priorities( batch_idxes, new_priorities) # train feedback regressor if use_feedback and len( feedback_buffer_train ) >= min_feedback_buffer_size and t <= only_use_hr_until: for i in range(feedback_epochs): if i < feedback_epochs * feedback_training_new_prop: inds = np.arange( len(feedback_buffer_train) - feedback_batch_size, len(feedback_buffer_train)) else: inds = np.random.choice(len(feedback_buffer_train), feedback_batch_size, replace=False) np.random.shuffle(inds) for start in range(0, feedback_batch_size, feedback_minibatch_size): end = start + feedback_minibatch_size obses = np.asarray([ feedback_buffer_train[idx][0] for idx in inds[start:end] ]) actions = np.asarray([ feedback_buffer_train[idx][1] for idx in inds[start:end] ]) feedbacks = np.asarray([ feedback_buffer_train[idx][2] for idx in inds[start:end] ]) pred, loss = train_hr(obses, actions, feedbacks) obs_train = np.asarray( [feedback[0] for feedback in feedback_buffer_train]) actions_train = np.asarray( [feedback[1] for feedback in feedback_buffer_train]) feedbacks_train = np.asarray( [feedback[2] for feedback in feedback_buffer_train]) obs_valid = np.asarray( [feedback[0] for feedback in feedback_buffer_valid]) actions_valid = np.asarray( [feedback[1] for feedback in feedback_buffer_valid]) feedbacks_valid = np.asarray( [feedback[2] for feedback in feedback_buffer_valid]) train_acc, train_loss = evaluate_hr(obs_train, actions_train, feedbacks_train) valid_acc, valid_loss = evaluate_hr(obs_valid, actions_valid, feedbacks_valid) print( "HR: train acc {:>4.2f}, loss {:>5.2f}; valid acc {:>4.2f}, loss {:>5.2f}" .format(train_acc, train_loss, valid_acc, valid_loss)) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len( episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log( "Saving model due to mean reward increase: {} -> {}" .format(saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format( saved_mean_reward)) load_variables(model_file) return act, performance
def learn(*, network, env, total_timesteps, seed=None, eval_env=None, replay_strategy='future', policy_save_interval=5, clip_return=True, demo_file=None, override_params=None, load_path=None, save_path=None, **kwargs): override_params = override_params or {} if MPI is not None: rank = MPI.COMM_WORLD.Get_rank() num_cpu = MPI.COMM_WORLD.Get_size() # Seed everything. rank_seed = seed + 1000000 * rank if seed is not None else None set_global_seeds(rank_seed) # Prepare params. params = config.DEFAULT_PARAMS env_name = env.specs[0].id params['env_name'] = env_name params['replay_strategy'] = replay_strategy if env_name in config.DEFAULT_ENV_PARAMS: params.update(config.DEFAULT_ENV_PARAMS[env_name] ) # merge env-specific parameters in params.update( **override_params) # makes it possible to override any parameter with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f: json.dump(params, f) params = config.prepare_params(params) params['rollout_batch_size'] = env.num_envs if demo_file is not None: params['bc_loss'] = 1 params.update(kwargs) config.log_params(params, logger=logger) if num_cpu == 1: logger.warn() logger.warn('*** Warning ***') logger.warn( 'You are running HER with just a single MPI worker. This will work, but the ' + 'experiments that we report in Plappert et al. (2018, https://arxiv.org/abs/1802.09464) ' + 'were obtained with --num_cpu 19. This makes a significant difference and if you ' + 'are looking to reproduce those results, be aware of this. Please also refer to ' + 'https://github.com/openai/baselines/issues/314 for further details.' ) logger.warn('****************') logger.warn() dims = config.configure_dims(params) policy = config.configure_ddpg(dims=dims, params=params, clip_return=clip_return) if load_path is not None: tf_util.load_variables(load_path) rollout_params = { 'exploit': False, 'use_target_net': False, 'use_demo_states': True, 'compute_Q': False, 'T': params['T'], } eval_params = { 'exploit': True, 'use_target_net': params['test_with_polyak'], 'use_demo_states': False, 'compute_Q': True, 'T': params['T'], } for name in [ 'T', 'rollout_batch_size', 'gamma', 'noise_eps', 'random_eps' ]: rollout_params[name] = params[name] eval_params[name] = params[name] eval_env = eval_env or env rollout_worker = RolloutWorker(env, policy, dims, logger, monitor=True, **rollout_params) evaluator = RolloutWorker(eval_env, policy, dims, logger, **eval_params) n_cycles = params['n_cycles'] n_epochs = total_timesteps // n_cycles // rollout_worker.T // rollout_worker.rollout_batch_size return train(save_path=save_path, policy=policy, rollout_worker=rollout_worker, evaluator=evaluator, n_epochs=n_epochs, n_test_rollouts=params['n_test_rollouts'], n_cycles=params['n_cycles'], n_batches=params['n_batches'], policy_save_interval=policy_save_interval, demo_file=demo_file)
def learn(env, use_ddpg=False, gamma=0.9, controller_kargs={}, option_kargs={}, seed=None, total_timesteps=100000, print_freq=100, callback=None, checkpoint_path=None, checkpoint_freq=10000, load_path=None, **others): """Train a deepq model. Parameters ------- env: gym.Env environment to train on use_ddpg: bool whether to use DDPG or DQN to learn the option's policies gamma: float discount factor controller_kargs arguments for learning the controller policy. option_kargs arguments for learning the option policies. seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. total_timesteps: int number of env steps to optimizer for print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. load_path: str path to load the model from. (default: None) Returns ------- act: ActWrapper (meta-controller) Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. act: ActWrapper (option policies) Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) controller = ControllerDQN(env, **controller_kargs) if use_ddpg: options = OptionDDPG(env, gamma, total_timesteps, **option_kargs) else: options = OptionDQN(env, gamma, total_timesteps, **option_kargs) option_s = None # State where the option initiated option_id = None # Id of the current option being executed option_rews = [] # Rewards obtained by the current option episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() options.reset() reset = True with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): if callback is not None: if callback(locals(), globals()): break # Selecting an option if needed if option_id is None: valid_options = env.get_valid_options() option_s = obs option_id = controller.get_action(option_s, valid_options) option_rews = [] # Take action and update exploration to the newest value action = options.get_action(env.get_option_observation(option_id), t, reset) reset = False new_obs, rew, done, info = env.step(action) # Saving the real reward that the option is getting option_rews.append(rew) # Store transition for the option policies for _s, _a, _r, _sn, _done in env.get_experience(): options.add_experience(_s, _a, _r, _sn, _done) # Learn and update the target networks if needed for the option policies options.learn(t) options.update_target_network(t) # Update the meta-controller if needed # Note that this condition always hold if done is True if env.did_option_terminate(option_id): option_sn = new_obs option_reward = sum( [_r * gamma**_i for _i, _r in enumerate(option_rews)]) valid_options = [] if done else env.get_valid_options() controller.add_experience(option_s, option_id, option_reward, option_sn, done, valid_options, gamma**(len(option_rews))) controller.learn() controller.update_target_network() controller.increase_step() option_id = None obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() options.reset() episode_rewards.append(0.0) reset = True # General stats mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len( episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.dump_tabular() if (checkpoint_freq is not None and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log( "Saving model due to mean reward increase: {} -> {}" .format(saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format( saved_mean_reward)) load_variables(model_file) return controller.act, options.act
def prepare_agent(_env, eval_env, active, exploration='eps_greedy', action_l2=None, scope=None, ss=False, load_path=None): # Prepare params. _params = copy.deepcopy(config.DEFAULT_PARAMS) _kwargs = copy.deepcopy(kwargs) _override_params = copy.deepcopy(override_params) env_name = _env.spec.id _params['env_name'] = env_name _params['replay_strategy'] = replay_strategy _params['ss'] = ss if action_l2 is not None: _params['action_l2'] = action_l2 if not active: _params["buffer_size"] = 1 if env_name in config.DEFAULT_ENV_PARAMS: _params.update(config.DEFAULT_ENV_PARAMS[env_name] ) # merge env-specific parameters in _params.update( **_override_params) # makes it possible to override any parameter with open(os.path.join(logger.get_dir(), 'params.json'), 'w') as f: json.dump(_params, f) _params = config.prepare_params(_params) _params['rollout_batch_size'] = _env.num_envs if demo_file is not None: _params['bc_loss'] = 1 _params.update(_kwargs) config.log_params(_params, logger=logger) if num_cpu == 1: logger.warn() logger.warn('*** Warning ***') logger.warn( 'You are running HER with just a single MPI worker. This will work, but the ' + 'experiments that we report in Plappert et al. (2018, https://arxiv.org/abs/1802.09464) ' + 'were obtained with --num_cpu 19. This makes a significant difference and if you ' + 'are looking to reproduce those results, be aware of this. Please also refer to ' + 'https://github.com/openai/baselines/issues/314 for further details.' ) logger.warn('****************') logger.warn() dims, coord_dict = config.configure_dims(_params) _params['ddpg_params']['scope'] = scope policy, reward_fun = config.configure_ddpg(dims=dims, params=_params, active=active, clip_return=clip_return) if load_path is not None: tf_util.load_variables(load_path) print(f"Loaded model: {load_path}") rollout_params = { 'exploit': False, 'use_target_net': False, 'use_demo_states': True, 'compute_Q': False, 'exploration': exploration } eval_params = { 'exploit': True, 'use_target_net': _params['test_with_polyak'], 'use_demo_states': False, 'compute_Q': True, } for name in [ 'T', 'rollout_batch_size', 'gamma', 'noise_eps', 'random_eps' ]: rollout_params[name] = _params[name] eval_params[name] = _params[name] eval_env = eval_env or _env rollout_worker = RolloutWorker(_env, policy, dims, logger, active, monitor=True, **rollout_params) evaluator = RolloutWorker(eval_env, policy, dims, logger, active, **eval_params) return policy, rollout_worker, evaluator, _params, coord_dict, reward_fun
def learn(env, network, seed=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, **network_kwargs ): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. batch_size: int size of a batch sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=make_obs_ph, q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) ############################## RL-S Prepare ############################################# # model saved name saved_name = "0817" ##### # Setup Training Record ##### save_new_data = False create_new_file = False create_new_file_rule = create_new_file save_new_data_rule = save_new_data create_new_file_RL = False save_new_data_RL = save_new_data create_new_file_replay_buffer = False save_new_data_replay_buffer = save_new_data is_training = False trajectory_buffer = deque(maxlen=20) if create_new_file_replay_buffer: if osp.exists("recorded_replay_buffer.txt"): os.remove("recorded_replay_buffer.txt") else: replay_buffer_dataset = np.loadtxt("recorded_replay_buffer.txt") for data in replay_buffer_dataset: obs, action, rew, new_obs, done = _extract_data(data) replay_buffer.add(obs, action, rew, new_obs, done) recorded_replay_buffer_outfile = open("recorded_replay_buffer.txt","a") recorded_replay_buffer_format = " ".join(("%f",)*31)+"\n" ##### # Setup Rule-based Record ##### create_new_file_rule = True # create state database if create_new_file_rule: if osp.exists("state_index_rule.dat"): os.remove("state_index_rule.dat") os.remove("state_index_rule.idx") if osp.exists("visited_state_rule.txt"): os.remove("visited_state_rule.txt") if osp.exists("visited_value_rule.txt"): os.remove("visited_value_rule.txt") visited_state_rule_value = [] visited_state_rule_counter = 0 else: visited_state_rule_value = np.loadtxt("visited_value_rule.txt") visited_state_rule_value = visited_state_rule_value.tolist() visited_state_rule_counter = len(visited_state_rule_value) visited_state_rule_outfile = open("visited_state_rule.txt", "a") visited_state_format = " ".join(("%f",)*14)+"\n" visited_value_rule_outfile = open("visited_value_rule.txt", "a") visited_value_format = " ".join(("%f",)*2)+"\n" visited_state_tree_prop = rindex.Property() visited_state_tree_prop.dimension = 14 visited_state_dist = np.array([[0.2, 2, 10, 0.2, 2, 10, 0.2, 2, 10, 0.2, 2, 10, 0.2, 2]]) visited_state_rule_tree = rindex.Index('state_index_rule',properties=visited_state_tree_prop) ##### # Setup RL-based Record ##### if create_new_file_RL: if osp.exists("state_index_RL.dat"): os.remove("state_index_RL.dat") os.remove("state_index_RL.idx") if osp.exists("visited_state_RL.txt"): os.remove("visited_state_RL.txt") if osp.exists("visited_value_RL.txt"): os.remove("visited_value_RL.txt") if create_new_file_RL: visited_state_RL_value = [] visited_state_RL_counter = 0 else: visited_state_RL_value = np.loadtxt("visited_value_RL.txt") visited_state_RL_value = visited_state_RL_value.tolist() visited_state_RL_counter = len(visited_state_RL_value) visited_state_RL_outfile = open("visited_state_RL.txt", "a") visited_state_format = " ".join(("%f",)*14)+"\n" visited_value_RL_outfile = open("visited_value_RL.txt", "a") visited_value_format = " ".join(("%f",)*2)+"\n" visited_state_tree_prop = rindex.Property() visited_state_tree_prop.dimension = 14 visited_state_dist = np.array([[0.2, 2, 10, 0.2, 2, 10, 0.2, 2, 10, 0.2, 2, 10, 0.2, 2]]) visited_state_RL_tree = rindex.Index('state_index_RL',properties=visited_state_tree_prop) ############################## RL-S Prepare End ############################################# # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() reset = True with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True action, q_function_cz = act(np.array(obs)[None], update_eps=update_eps, **kwargs) # RLS_action = generate_RLS_action(obs,q_function_cz,action,visited_state_rule_value, # visited_state_rule_tree,visited_state_RL_value, # visited_state_RL_tree,is_training) RLS_action = 0 env_action = RLS_action reset = False new_obs, rew, done, _ = env.step(env_action) ########### Record data in trajectory buffer and local file, but not in replay buffer ########### trajectory_buffer.append((obs, action, float(rew), new_obs, float(done))) # Store transition in the replay buffer. # replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew # safe driving is 1, collision is 0 while len(trajectory_buffer)>10: # if safe driving for 10(can be changed) steps, the state is regarded as safe obs_left, action_left, rew_left, new_obs_left, done_left = trajectory_buffer.popleft() # save this state in local replay buffer file if save_new_data_replay_buffer: recorded_data = _wrap_data(obs_left, action_left, rew_left, new_obs_left, done_left) recorded_replay_buffer_outfile.write(recorded_replay_buffer_format % tuple(recorded_data)) # put this state in replay buffer replay_buffer.add(obs_left[0], action_left, float(rew_left), new_obs_left[0], float(done_left)) action_to_record = action_left r_to_record = rew_left obs_to_record = obs_left # save this state in rule-based or RL-based visited state if action_left == 0: if save_new_data_rule: visited_state_rule_value.append([action_to_record,r_to_record]) visited_state_rule_tree.insert(visited_state_rule_counter, tuple((obs_to_record-visited_state_dist).tolist()[0]+(obs_to_record+visited_state_dist).tolist()[0])) visited_state_rule_outfile.write(visited_state_format % tuple(obs_to_record[0])) visited_value_rule_outfile.write(visited_value_format % tuple([action_to_record,r_to_record])) visited_state_rule_counter += 1 else: if save_new_data_RL: visited_state_RL_value.append([action_to_record,r_to_record]) visited_state_RL_tree.insert(visited_state_RL_counter, tuple((obs_to_record-visited_state_dist).tolist()[0]+(obs_to_record+visited_state_dist).tolist()[0])) visited_state_RL_outfile.write(visited_state_format % tuple(obs_to_record[0])) visited_value_RL_outfile.write(visited_value_format % tuple([action_to_record,r_to_record])) visited_state_RL_counter += 1 ################# Record data end ######################## if done: """ Get collision or out of multilane map """ ####### Record the trajectory data and add data in replay buffer ######### _, _, rew_right, _, _ = trajectory_buffer[-1] while len(trajectory_buffer)>0: obs_left, action_left, rew_left, new_obs_left, done_left = trajectory_buffer.popleft() action_to_record = action_left r_to_record = (rew_right-rew_left)*gamma**len(trajectory_buffer) + rew_left # record in local replay buffer file if save_new_data_replay_buffer: obs_to_record = obs_left recorded_data = _wrap_data(obs_left, action_left, r_to_record, new_obs_left, done_left) recorded_replay_buffer_outfile.write(recorded_replay_buffer_format % tuple(recorded_data)) # record in replay buffer for trainning replay_buffer.add(obs_left[0], action_left, float(r_to_record), new_obs_left[0], float(done_left)) # save visited rule/RL state data in local file if action_left == 0: if save_new_data_rule: visited_state_rule_value.append([action_to_record,r_to_record]) visited_state_rule_tree.insert(visited_state_rule_counter, tuple((obs_to_record-visited_state_dist).tolist()[0]+(obs_to_record+visited_state_dist).tolist()[0])) visited_state_rule_outfile.write(visited_state_format % tuple(obs_to_record[0])) visited_value_rule_outfile.write(visited_value_format % tuple([action_to_record,r_to_record])) visited_state_rule_counter += 1 else: if save_new_data_RL: visited_state_RL_value.append([action_to_record,r_to_record]) visited_state_RL_tree.insert(visited_state_RL_counter, tuple((obs_to_record-visited_state_dist).tolist()[0]+(obs_to_record+visited_state_dist).tolist()[0])) visited_state_RL_outfile.write(visited_state_format % tuple(obs_to_record[0])) visited_value_RL_outfile.write(visited_value_format % tuple([action_to_record,r_to_record])) visited_state_RL_counter += 1 ####### Recorded ##### obs = env.reset() episode_rewards.append(0.0) reset = True ############### Trainning Part Start ##################### if not is_training: # don't need to train the model continue if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len(episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward rew_str = str(mean_100ep_reward) path = osp.expanduser("~/models/carlaok_checkpoint/"+saved_name+"_"+rew_str) act.save(path) #### close the file #### visited_state_rule_outfile.close() visited_value_rule_outfile.close() recorded_replay_buffer_outfile.close() if not is_training: testing_record_outfile.close() #### close the file ### if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) load_variables(model_file) return act
def learn(env, network, seed=None, lr=5e-4, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, callback=None, load_path=None, **network_kwargs ): """Train a deepq model. Parameters ------- env: gym.Env environment to train on network: string or a function neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models (mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that) seed: int or None prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used. lr: float learning rate for adam optimizer total_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to total_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. param_noise: bool whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905) callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. load_path: str path to load the model from. (default: None) **network_kwargs additional keyword arguments to pass to the network builder. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = get_session() set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) # capture the shape outside the closure so that the env object is not serialized # by cloudpickle when serializing make_obs_ph observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=make_obs_ph, q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, param_noise=param_noise ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() reset = True with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(env.action_space.n)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0] env_action = action reset = False new_obs, rew, done, _ = env.step(env_action) # Store transition in the replay buffer. replay_buffer.add(obs, action, rew, new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0.0) reset = True if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len(episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) load_variables(model_file) return act
def learn(env, network, seed=None, lr=1e-3, total_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=100, checkpoint_freq=10000, checkpoint_path=None, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, param_noise=False, num_cpu=5, callback=None, scope='co_deepq', pilot_tol=0, pilot_is_human=False, reuse=False, load_path=None, **network_kwargs): # Create all the functions necessary to train the model sess = get_session() #tf.Session(graph=tf.Graph()) set_global_seeds(seed) q_func = build_q_func(network, **network_kwargs) observation_space = env.observation_space def make_obs_ph(name): return ObservationInput(observation_space, name=name) using_control_sharing = True #pilot_tol > 0 if pilot_is_human: utils.human_agent_action = init_human_action() utils.human_agent_active = False act, train, update_target, debug = co_build_train( scope=scope, make_obs_ph=make_obs_ph, q_func=q_func, num_actions=env.action_space.n, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10, reuse=tf.AUTO_REUSE if reuse else False, using_control_sharing=using_control_sharing) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': env.action_space.n, } act = ActWrapper(act, act_params) # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = total_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] episode_outcomes = [] saved_mean_reward = None obs = env.reset() reset = True prev_t = 0 rollouts = [] if not using_control_sharing: exploration = LinearSchedule(schedule_timesteps=int( exploration_fraction * total_timesteps), initial_p=1.0, final_p=exploration_final_eps) with tempfile.TemporaryDirectory() as td: td = checkpoint_path or td model_file = os.path.join(td, "model") model_saved = False if tf.train.latest_checkpoint(td) is not None: load_variables(model_file) logger.log('Loaded model from {}'.format(model_file)) model_saved = True elif load_path is not None: load_variables(load_path) logger.log('Loaded model from {}'.format(load_path)) for t in range(total_timesteps): masked_obs = mask_helipad(obs) act_kwargs = {} if using_control_sharing: if pilot_is_human: act_kwargs['pilot_action'] = env.unwrapped.pilot_policy( obs[None, :9]) else: act_kwargs[ 'pilot_action'] = env.unwrapped.pilot_policy.step( obs[None, :9]) act_kwargs['pilot_tol'] = pilot_tol if not pilot_is_human or ( pilot_is_human and utils.human_agent_active) else 0 else: act_kwargs['update_eps'] = exploration.value(t) #action = act(masked_obs[None, :], **act_kwargs)[0][0] action = act(np.array(masked_obs)[None], **act_kwargs)[0][0] env_action = action reset = False new_obs, rew, done, info = env.step(env_action) if pilot_is_human: env.render() # Store transition in the replay buffer. masked_new_obs = mask_helipad(new_obs) replay_buffer.add(masked_obs, action, rew, masked_new_obs, float(done)) obs = new_obs episode_rewards[-1] += rew if done: obs = env.reset() episode_rewards.append(0.0) reset = True if pilot_is_human: utils.human_agent_action = init_human_action() utils.human_agent_active = False time.sleep(2) if t > learning_starts and t % train_freq == 0: if prioritized_replay: experience = replay_buffer.sample( batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample( batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() episode_outcomes.append(rew) episode_rewards.append(0.0) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) mean_100ep_succ = round( np.mean( [1 if x == 100 else 0 for x in episode_outcomes[-101:-1]]), 2) mean_100ep_crash = round( np.mean([ 1 if x == -100 else 0 for x in episode_outcomes[-101:-1] ]), 2) num_episodes = len(episode_rewards) if done and print_freq is not None and len( episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("mean 100 episode succ", mean_100ep_succ) logger.record_tabular("mean 100 episode crash", mean_100ep_crash) logger.dump_tabular() if checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0 and ( saved_mean_reward is None or mean_100ep_reward > saved_mean_reward): if print_freq is not None: logger.log( "Saving model due to mean reward increase: {} -> {}". format(saved_mean_reward, mean_100ep_reward)) save_variables(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format( saved_mean_reward)) load_variables(model_file) reward_data = {'rewards': episode_rewards, 'outcomes': episode_outcomes} return act, reward_data