def is_valid(self, state): if has_nonterminals(state): return False robot = build_normalized_robot(state) _, has_self_collision = presimulate(robot) return not has_self_collision
def sample_design(args, task_id, seed, env, V, eps, results_queue, time_queue, done_event): tt0 = time.time() random.seed(seed) valid = False samples = [] while not valid: state = make_initial_graph() rule_seq = [] no_action_flag = False for _ in range(args.depth): action, step_type = select_action(env, V, state, eps) if action is None: no_action_flag = True break rule_seq.append(action) next_state = env.transite(state, action) state = next_state if not has_nonterminals(state): break valid = env.is_valid(state) if not valid: # update the invalid sample's count if no_action_flag: info = 'no_action' elif has_nonterminals(state): info = 'step_exceeded' else: info = 'self_collision' samples.append(Sample(task_id, rule_seq, -2.0, info)) else: samples.append( Sample(task_id, rule_seq, predict(V, state), info='valid')) tt = time.time() - tt0 time_queue.put(tt) results_queue.put(samples) done_event.wait()
def compute_Vhat(robot_graph, env, V): if has_nonterminals(robot_graph): available_actions = env.get_available_actions(robot_graph) if len(available_actions) == 0: return -5.0 next_states = [] for action in available_actions: next_states.append(env.transite(robot_graph, action)) values = predict_batch(V, next_states) return np.max(values) else: return env.get_reward(robot_graph)[1]
def step(self, action): next_state = self.transite(self.state, action) self.rule_seq.append(action) if has_nonterminals(next_state): reward = 0. done = False else: input_sequence, reward = self.get_reward(next_state) done = True self.state = next_state return self.state, reward, done
def search_algo(args): # iniailize random seed random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) torch.set_num_threads(1) # initialize/load task_class = getattr(tasks, args.task) if args.no_noise: task = task_class(force_std=0.0, torque_std=0.0) else: task = task_class() graphs = rd.load_graphs(args.grammar_file) rules = [rd.create_rule_from_graph(g) for g in graphs] # initialize preprocessor # Find all possible link labels, so they can be one-hot encoded all_labels = set() for rule in rules: for node in rule.lhs.nodes: all_labels.add(node.attrs.require_label) all_labels = sorted(list(all_labels)) # TODO: use 80 to fit the input of trained MPC GNN, use args.depth * 3 later for real mpc max_nodes = args.depth * 3 global preprocessor # preprocessor = Preprocessor(max_nodes = max_nodes, all_labels = all_labels) preprocessor = Preprocessor(all_labels=all_labels) # initialize the env env = RobotGrammarEnv(task, rules, seed=args.seed, mpc_num_processes=args.mpc_num_processes) # initialize Value function device = 'cpu' state = env.reset() sample_adj_matrix, sample_features, sample_masks = preprocessor.preprocess( state) num_features = sample_features.shape[1] V = Net(max_nodes=max_nodes, num_channels=num_features, num_outputs=1).to(device) # load pretrained V function if args.load_V_path is not None: V.load_state_dict(torch.load(args.load_V_path)) print_info('Loaded pretrained V function from {}'.format( args.load_V_path)) # initialize target V_hat look up table V_hat = dict() # load pretrained V_hat if args.load_Vhat_path is not None: V_hat_fp = open(args.load_Vhat_path, 'rb') V_hat = pickle.load(V_hat_fp) V_hat_fp.close() print_info('Loaded pretrained Vhat from {}'.format( args.load_Vhat_path)) # initialize invalid_his invalid_his = dict() num_invalid_samples, num_valid_samples = 0, 0 repeated_cnt = 0 # initialize the seen states pool states_pool = StatesPool(capacity=args.states_pool_capacity) states_set = set() # explored designs designs = [] design_rewards = [] design_opt_seeds = [] # record prediction error prediction_error_sum = 0.0 if not args.test: # initialize save folders and files fp_log = open(os.path.join(args.save_dir, 'log.txt'), 'w') fp_log.close() fp_eval = open(os.path.join(args.save_dir, 'eval.txt'), 'w') fp_eval.close() design_csv_path = os.path.join(args.save_dir, 'designs.csv') fp_csv = open(design_csv_path, 'w') fieldnames = ['rule_seq', 'reward', 'opt_seed'] writer = csv.DictWriter(fp_csv, fieldnames=fieldnames) writer.writeheader() fp_csv.close() # initialize the optimizer global optimizer optimizer = torch.optim.Adam(V.parameters(), lr=args.lr) # initialize best design rule sequence best_design, best_reward = None, -np.inf # reward history epoch_rew_his = [] last_checkpoint = -1 # recording time t_sample_sum = 0. # record the count for invalid samples no_action_samples, step_exceeded_samples, self_collision_samples = 0, 0, 0 for epoch in range(args.num_iterations): t_start = time.time() V.eval() # update eps and eps_sample if args.eps_schedule == 'linear-decay': eps = args.eps_start + epoch / args.num_iterations * ( args.eps_end - args.eps_start) elif args.eps_schedule == 'exp-decay': eps = args.eps_end + (args.eps_start - args.eps_end) * np.exp( -1.0 * epoch / args.num_iterations / args.eps_decay) if args.eps_sample_schedule == 'linear-decay': eps_sample = args.eps_sample_start + epoch / args.num_iterations * ( args.eps_sample_end - args.eps_sample_start) elif args.eps_sample_schedule == 'exp-decay': eps_sample = args.eps_sample_end + ( args.eps_sample_start - args.eps_sample_end) * np.exp( -1.0 * epoch / args.num_iterations / args.eps_sample_decay) t_sample, t_update, t_mpc, t_opt = 0, 0, 0, 0 selected_design, selected_reward = None, -np.inf selected_state_seq, selected_rule_seq = None, None p = random.random() if p < eps_sample: num_samples = 1 else: num_samples = args.num_samples # use e-greedy to sample a design within maximum #steps. for _ in range(num_samples): valid = False while not valid: t0 = time.time() state = env.reset() rule_seq = [] state_seq = [state] no_action_flag = False for _ in range(args.depth): action, step_type = select_action(env, V, state, eps) if action is None: no_action_flag = True break rule_seq.append(action) next_state = env.transite(state, action) state_seq.append(next_state) state = next_state if not has_nonterminals(state): break valid = env.is_valid(state) t_sample += time.time() - t0 t0 = time.time() if not valid: # update the invalid sample's count if no_action_flag: no_action_samples += 1 elif has_nonterminals(state): step_exceeded_samples += 1 else: self_collision_samples += 1 # update the Vhat for invalid designs update_Vhat(args, V_hat, state_seq, -2.0, invalid=True, invalid_cnt=invalid_his) # update states pool update_states_pool(states_pool, state_seq, states_set, V_hat) num_invalid_samples += 1 else: num_valid_samples += 1 t_update += time.time() - t0 predicted_value = predict(V, state) if predicted_value > selected_reward: selected_design, selected_reward = state, predicted_value selected_rule_seq, selected_state_seq = rule_seq, state_seq t0 = time.time() repeated = False if (hash(selected_design) in V_hat) and (V_hat[hash(selected_design)] > -2.0 + 1e-3): repeated = True repeated_cnt += 1 reward, best_seed = -np.inf, None for _ in range(args.num_eval): _, rew = env.get_reward(selected_design) if rew > reward: reward, best_seed = rew, env.last_opt_seed t_mpc += time.time() - t0 # save the design and the reward in the list designs.append(selected_rule_seq) design_rewards.append(reward) design_opt_seeds.append(best_seed) # update best design if reward > best_reward: best_design, best_reward = selected_rule_seq, reward print_info( 'new best: reward = {:.4f}, predicted reward = {:.4f}, num_samples = {}' .format(reward, selected_reward, num_samples)) t0 = time.time() # update V_hat for the valid design update_Vhat(args, V_hat, selected_state_seq, reward) # update states pool for the valid design update_states_pool(states_pool, selected_state_seq, states_set, V_hat) t_update += time.time() - t0 t0 = time.time() # optimize V.train() total_loss = 0.0 for _ in range(args.opt_iter): minibatch = states_pool.sample( min(len(states_pool), args.batch_size)) train_adj_matrix, train_features, train_masks, train_reward = [], [], [], [] max_nodes = 0 for robot_graph in minibatch: hash_key = hash(robot_graph) target_reward = V_hat[hash_key] # adj_matrix, features, masks = preprocessor.preprocess(robot_graph) adj_matrix, features, _ = preprocessor.preprocess( robot_graph) max_nodes = max(max_nodes, len(features)) train_adj_matrix.append(adj_matrix) train_features.append(features) # train_masks.append(masks) train_reward.append(target_reward) for i in range(len(minibatch)): train_adj_matrix[i], train_features[i], masks = \ preprocessor.pad_graph(train_adj_matrix[i], train_features[i], max_nodes) train_masks.append(masks) train_adj_matrix_torch = torch.tensor(train_adj_matrix) train_features_torch = torch.tensor(train_features) train_masks_torch = torch.tensor(train_masks) train_reward_torch = torch.tensor(train_reward) optimizer.zero_grad() output, loss_link, loss_entropy = V(train_features_torch, train_adj_matrix_torch, train_masks_torch) loss = F.mse_loss(output[:, 0], train_reward_torch) loss.backward() total_loss += loss.item() optimizer.step() t_opt += time.time() - t0 t_end = time.time() t_sample_sum += t_sample # logging if (epoch + 1 ) % args.log_interval == 0 or epoch + 1 == args.num_iterations: iter_save_dir = os.path.join(args.save_dir, '{}'.format(epoch + 1)) os.makedirs(os.path.join(iter_save_dir), exist_ok=True) # save model save_path = os.path.join(iter_save_dir, 'V_model.pt') torch.save(V.state_dict(), save_path) # save V_hat save_path = os.path.join(iter_save_dir, 'V_hat') fp = open(save_path, 'wb') pickle.dump(V_hat, fp) fp.close() # save explored design and its reward fp_csv = open(design_csv_path, 'a') fieldnames = ['rule_seq', 'reward', 'opt_seed'] writer = csv.DictWriter(fp_csv, fieldnames=fieldnames) for i in range(last_checkpoint + 1, len(designs)): writer.writerow({ 'rule_seq': str(designs[i]), 'reward': design_rewards[i], 'opt_seed': design_opt_seeds[i] }) last_checkpoint = len(designs) - 1 fp_csv.close() epoch_rew_his.append(reward) avg_loss = total_loss / args.opt_iter len_his = min(len(epoch_rew_his), 30) avg_reward = np.sum(epoch_rew_his[-len_his:]) / len_his prediction_error_sum += (selected_reward - reward)**2 avg_prediction_error = prediction_error_sum / (epoch + 1) if repeated: print_white('Epoch {:4}: T_sample = {:5.2f}, T_update = {:5.2f}, T_mpc = {:5.2f}, T_opt = {:5.2f}, eps = {:5.3f}, eps_sample = {:5.3f}, #samples = {:2}, training loss = {:7.4f}, pred_error = {:6.4f}, predicted_reward = {:6.4f}, reward = {:6.4f}, last 30 epoch reward = {:6.4f}, best reward = {:6.4f}'.format(\ epoch, t_sample, t_update, t_mpc, t_opt, eps, eps_sample, num_samples, \ avg_loss, avg_prediction_error, selected_reward, reward, avg_reward, best_reward)) else: print_warning('Epoch {:4}: T_sample = {:5.2f}, T_update = {:5.2f}, T_mpc = {:5.2f}, T_opt = {:5.2f}, eps = {:5.3f}, eps_sample = {:5.3f}, #samples = {:2}, training loss = {:7.4f}, pred_error = {:6.4f}, predicted_reward = {:6.4f}, reward = {:6.4f}, last 30 epoch reward = {:6.4f}, best reward = {:6.4f}'.format(\ epoch, t_sample, t_update, t_mpc, t_opt, eps, eps_sample, num_samples, \ avg_loss, avg_prediction_error, selected_reward, reward, avg_reward, best_reward)) fp_log = open(os.path.join(args.save_dir, 'log.txt'), 'a') fp_log.write('eps = {:.4f}, eps_sample = {:.4f}, num_samples = {}, T_sample = {:4f}, T_update = {:4f}, T_mpc = {:.4f}, T_opt = {:.4f}, loss = {:.4f}, predicted_reward = {:.4f}, reward = {:.4f}, avg_reward = {:.4f}\n'.format(\ eps, eps_sample, num_samples, t_sample, t_update, t_mpc, t_opt, avg_loss, selected_reward, reward, avg_reward)) fp_log.close() if (epoch + 1) % args.log_interval == 0: print_info( 'Avg sampling time for last {} epoch: {:.4f} second'. format(args.log_interval, t_sample_sum / args.log_interval)) t_sample_sum = 0. print_info('size of states_pool = {}'.format(len(states_pool))) print_info( '#valid samples = {}, #invalid samples = {}, #valid / #invalid = {}' .format( num_valid_samples, num_invalid_samples, num_valid_samples / num_invalid_samples if num_invalid_samples > 0 else 10000.0)) print_info( 'Invalid samples: #no_action_samples = {}, #step_exceeded_samples = {}, #self_collision_samples = {}' .format(no_action_samples, step_exceeded_samples, self_collision_samples)) max_trials, cnt = 0, 0 for key in invalid_his.keys(): if invalid_his[key] > max_trials: if key not in V_hat: max_trials = invalid_his[key] elif V_hat[key] < -2.0 + 1e-3: max_trials = invalid_his[key] if invalid_his[key] >= args.max_trials: if V_hat[key] < -2.0 + 1e-3: cnt += 1 print_info( 'max invalid_trials = {}, #failed nodes = {}'.format( max_trials, cnt)) print_info('repeated rate = {}'.format(repeated_cnt / (epoch + 1))) save_path = os.path.join(args.save_dir, 'model_state_dict_final.pt') torch.save(V.state_dict(), save_path) else: import IPython IPython.embed() # test V.eval() print('Start testing') test_epoch = 30 y0 = [] y1 = [] x = [] for ii in range(0, 11): eps = 1.0 - 0.1 * ii print('------------------------------------------') print('eps = ', eps) reward_sum = 0. best_reward = -np.inf for epoch in range(test_epoch): t0 = time.time() # use e-greedy to sample a design within maximum #steps. vaild = False while not valid: state = env.reset() rule_seq = [] state_seq = [state] for _ in range(args.depth): action, step_type = select_action(env, V, state, eps) if action is None: break rule_seq.append(action) next_state = env.transite(state, action) state_seq.append(next_state) if not has_nonterminals(next_state): valid = True break state = next_state _, reward = env.get_reward(state) reward_sum += reward best_reward = max(best_reward, reward) print( f'design {epoch}: reward = {reward}, time = {time.time() - t0}' ) print('test avg reward = ', reward_sum / test_epoch) print('best reward found = ', best_reward) x.append(eps) y0.append(reward_sum / test_epoch) y1.append(best_reward) import matplotlib.pyplot as plt fig, ax = plt.subplots(1, 2, figsize=(10, 5)) ax[0].plot(x, y0) ax[0].set_title('Avg Reward') ax[0].set_xlabel('eps') ax[0].set_ylabel('reward') ax[1].plot(x, y1) ax[0].set_title('Best Reward') ax[0].set_xlabel('eps') ax[0].set_ylabel('reward') plt.show()