def __init__(self, config): self.config = config self.PROCESS_NUMBER = 4 self.num_agents = self.config.num_agents self.size_map = [self.config.map_w, self.config.map_w] self.label_density = str(self.config.map_density).split('.')[-1] self.AgentState = AgentState(self.num_agents) self.communicationRadius = 5 # communicationRadius self.zeroTolerance = 1e-9 self.delta = [[-1, 0], # go up [0, -1], # go left [1, 0], # go down [0, 1], # go right [0, 0]] # stop self.num_actions = 5 self.list_seqtrain_file = [] self.list_train_file = [] self.list_seqvalid_file = [] self.list_validStep_file = [] self.list_valid_file = [] self.list_test_file = [] self.hashids = Hashids(alphabet='01234567789abcdef', min_length=5) self.pathtransformer = self.pathtransformer_RelativeCoordinate self.label_setup = '{}{:02d}x{:02d}_density_p{}/{}_Agent'.format(self.config.loadmap_TYPE, self.size_map[0],self.size_map[1], self.label_density, self.num_agents) self.dirName_parent = os.path.join(self.config.solCases_dir, self.label_setup) self.dirName_Store = os.path.join(self.config.dir_SaveData, self.label_setup) self.dirName_input = os.path.join(self.dirName_parent, 'input') self.dirName_output = os.path.join(self.dirName_parent, 'output_{}'.format(config.chosen_solver)) self.set_up()
def __init__(self, config): self.config = config self.AgentState = AgentState(self.config.num_agents) self.delta_list =[[-1, 0], # go up [0, -1], # go left [1, 0], # go down [0, 1], # go right [0, 0]] # stop self.delta = torch.FloatTensor(self.delta_list).to(self.config.device) self.List_MultiAgent_ActionVec_target = None self.store_MultiAgent = None self.channel_map = None self.size_map = None self.maxstep = None self.posObstacle = None self.numObstacle = None self.posStart = None self.posGoal = None self.currentState_predict = None self.makespanTarget = None self.flowtimeTarget = None self.makespanPredict = None self.flowtimePredict = None self.count_reachgoal = None self.count_reachgoalTarget = None self.fun_Softmax = None self.zeroTolerance = 1e-9 print("run on multirobotsim with collision shielding")
def __init__(self, config, mode): self.config = config self.datapath_exp = '{}{:02d}x{:02d}_density_p{}/{}_Agent/'.format( self.config.map_type, self.config.map_w, self.config.map_h, self.config.map_density, self.config.num_agents) self.dirName = os.path.join(self.config.data_root, self.datapath_exp) self.AgentState = AgentState(self.config.num_agents) if mode == "train": self.dir_data = os.path.join(self.dirName, 'train') self.search_files = self.search_target_files_withStep self.data_paths, self.id_stepdata = self.update_data_path_trainingset( self.dir_data) self.load_data = self.load_train_data elif mode == "test_trainingSet": self.dir_data = os.path.join(self.dirName, 'train') # self.search_files = self.search_target_files # data_paths, id_stepdata = self.update_data_path_trainingset(self.dir_data) data_paths, id_stepdata = self.search_target_files(self.dir_data) paths_total = list(zip(data_paths, id_stepdata)) random.shuffle(paths_total) data_paths, id_stepdata = zip(*paths_total) self.data_paths = data_paths[:self.config.num_test_trainingSet] self.id_stepdata = id_stepdata[:self.config.num_test_trainingSet] self.load_data = self.load_data_during_training elif mode == "valid": self.dir_data = os.path.join(self.dirName, 'valid') self.data_paths, self.id_stepdata = self.obtain_data_path_validset( self.dir_data, self.config.num_validset) self.load_data = self.load_data_during_training elif mode == "validStep": self.dir_data = os.path.join(self.dirName, 'valid') self.data_paths, self.id_stepdata = self.search_valid_files_withStep( self.dir_data, self.config.num_validset) self.load_data = self.load_train_data elif mode == "test": self.dir_data = os.path.join(self.dirName, mode) self.data_paths, self.id_stepdata = self.obtain_data_path_validset( self.dir_data, self.config.num_testset) self.load_data = self.load_test_data self.data_size = len(self.data_paths)
def __init__(self, config): self.config = config self.PROCESS_NUMBER = 4 self.num_agents = self.config.num_agents self.size_map = [self.config.map_w, self.config.map_h] self.AgentState = AgentState(self.num_agents) self.communicationRadius = 5 # communicationRadius self.zeroTolerance = 1e-9 self.delta = [ [-1, 0], # go up [0, -1], # go left [1, 0], # go down [0, 1], # go right [0, 0] ] # stop self.num_actions = 5 self.root_path_save = self.config.failCases_dir self.list_seqtrain_file = [] self.list_train_file = [] self.pathtransformer = self.pathtransformer_RelativeCoordinate
class multiRobotSim: def __init__(self, config): self.config = config self.AgentState = AgentState(self.config.num_agents) self.delta_list =[[-1, 0], # go up [0, -1], # go left [1, 0], # go down [0, 1], # go right [0, 0]] # stop self.delta = torch.FloatTensor(self.delta_list).to(self.config.device) # self.onlineExpert = ComputeCBSSolution(self.config) self.List_MultiAgent_ActionVec_target = None self.store_MultiAgent = None self.channel_map = None self.size_map = None self.maxstep = None self.posObstacle = None self.numObstacle = None self.posStart = None self.posGoal = None self.currentState_predict = None self.makespanTarget = None self.flowtimeTarget = None self.makespanPredict = None self.flowtimePredict = None self.count_reachgoal = None self.count_reachgoalTarget = None self.fun_Softmax = None self.zeroTolerance = 1e-9 print("run on multirobotsim with collision shielding") def setup(self, loadInput, loadTarget, makespanTarget, tensor_map, ID_dataset): # self.fun_Softmax = nn.Softmax(dim=-1) self.fun_Softmax = nn.LogSoftmax(dim=-1) self.ID_dataset = ID_dataset self.store_GSO = [] self.store_communication_radius = [] self.status_MultiAgent = {} # setupState = loadInput.permute(3, 4, 2, 1, 0) target = loadTarget.permute(1, 2, 3, 0) self.List_MultiAgent_ActionVec_target = target[:, :, :,0] # self.List_MultiAgent_ActionVec_target = target[:, :, 0] self.channel_map = tensor_map[0] # setupState[:, :, 0, 0, 0] self.AgentState.setmap(self.channel_map) self.posObstacle = self.findpos(self.channel_map).to(self.config.device) self.numObstacle = self.posObstacle.shape[0] self.size_map = self.channel_map.shape # self.communicationRadius = 5 #self.size_map[0] * 0.5 # self.maxstep = self.size_map[0] * self.size_map[1] if self.config.num_agents >=20: self.rate_maxstep = 3 else: self.rate_maxstep = self.config.rate_maxstep self.maxstep = int(makespanTarget.type(torch.int32) * self.rate_maxstep) self.check_predictCollsion = False self.check_moveCollision = True self.check_predictEdgeCollsion = [False] * self.config.num_agents self.count_reachgoal = [False] * self.config.num_agents self.count_reachgoalTarget = [False] * self.config.num_agents self.allReachGoal_Target = False self.makespanTarget = 0 self.flowtimeTarget = 0 self.makespanPredict = self.maxstep self.flowtimePredict = self.maxstep * self.config.num_agents #0 # used for determining flowtimes of non rouge agents self.nonRogueFlowtimePredict = None self.stopKeyValue = torch.tensor(4).to(self.config.device) self.reset_disabled_action = torch.tensor([1.0, 1.0, 1.0, 1.0, 1.0]).float().to(self.config.device) self.store_goalAgents = loadInput[0, 0, :,:] self.store_stateAgents = loadInput[0, 1, :, :] for id_agent in range(self.config.num_agents): status_CurrentAgent = {} posGoal = loadInput[:, 0,id_agent,:] #self.findpos(goal_CurrentAgent) posStart = loadInput[:, 1,id_agent,:] #self.findpos(start_CurrentAgent) path_predict = {0:posStart} path_target = {0:posStart} len_action_predict = 0 list_actionKey_predict = [] actionVec_target_CurrentAgents = self.List_MultiAgent_ActionVec_target[id_agent, :, :] actionKeyList_target_CurrentAgents = torch.max(actionVec_target_CurrentAgents, 1)[1] disabled_action_predict_currentAgent = self.reset_disabled_action startStep_action_currentAgent = None endStep_action_currentAgent = None len_action_target = actionKeyList_target_CurrentAgents.shape[0] status_CurrentAgents = {"goal": posGoal, "start": posStart,#torch.FloatTensor([[0,0]]).to(self.config.device), "currentState": posStart, "path_target": path_target, "action_target": actionKeyList_target_CurrentAgents, "len_action_target": len_action_target, "startStep_action_target": startStep_action_currentAgent, "endStep_action_target": endStep_action_currentAgent, "path_predict": path_predict, "nextState_predict": posStart, "action_predict": list_actionKey_predict, "disabled_action_predict": disabled_action_predict_currentAgent, "len_action_predict": len_action_predict, "startStep_action_predict": startStep_action_currentAgent, "endStep_action_predict": endStep_action_currentAgent } # print("Agent{} - goal:{} - start:{} - currentState:{}".format(id_agent, posGoal,posStart,posStart)) name_agent = "agent{}".format(id_agent) self.status_MultiAgent.update({name_agent: status_CurrentAgents}) self.getPathTarget() pass def findpos(self, channel): pos_object = channel.nonzero() num_object = pos_object.shape[0] pos = torch.zeros(num_object, 2) # pos_list = [] for i in range(num_object): pos[i][0] = pos_object[i][0] pos[i][1] = pos_object[i][1] # pos_list.append([pos_object[i][0], pos_object[i][1]]) # pos = torch.FloatTensor(pos_list) return pos def getPathTarget(self): #todo check the length for ground truth, out of index list_len_action_target = [] for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) len_actionTarget_currentAgent = self.status_MultiAgent[name_agent]["len_action_target"] list_len_action_target.append(len_actionTarget_currentAgent) maxStep = max(list_len_action_target) for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) pathTarget_currentAgent = self.status_MultiAgent[name_agent]["path_target"] currentState_target = self.status_MultiAgent[name_agent]['start'] goal_currentAgent = self.status_MultiAgent[name_agent]['goal'] nextState_target = currentState_target goalIndexX = int(goal_currentAgent[0][0]) goalIndexY = int(goal_currentAgent[0][1]) for step in range(maxStep): actionKey_target = self.status_MultiAgent[name_agent]['action_target'][step] check_move = (actionKey_target != self.stopKeyValue) check_startStep_action = self.status_MultiAgent[name_agent]["startStep_action_target"] if check_move == 1 and check_startStep_action is None: self.status_MultiAgent[name_agent]["startStep_action_target"] = step else: currentState_target = nextState_target action_target = self.delta[actionKey_target] nextState_target = torch.add(currentState_target, action_target) pathTarget_currentAgent.update({step+1: nextState_target}) self.status_MultiAgent[name_agent]["path_target"] = pathTarget_currentAgent if nextState_target[0][0] == goalIndexX and nextState_target[0][1] == goalIndexY and not self.count_reachgoalTarget[id_agent]: self.count_reachgoalTarget[id_agent] = True self.status_MultiAgent[name_agent]["endStep_action_target"] = step + 1 self.allReachGoal_Target = all(self.count_reachgoalTarget) if self.allReachGoal_Target: List_endStep_target = [] List_startStep_target = [] self.flowtimeTarget = 0 for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) List_endStep_target.append(self.status_MultiAgent[name_agent]["endStep_action_target"]) List_startStep_target.append(self.status_MultiAgent[name_agent]["startStep_action_target"]) self.flowtimeTarget += self.status_MultiAgent[name_agent]["endStep_action_target"] - \ self.status_MultiAgent[name_agent]["startStep_action_target"] len_action_predict = self.status_MultiAgent[name_agent]["endStep_action_target"] - \ self.status_MultiAgent[name_agent]["startStep_action_target"] self.status_MultiAgent[name_agent]["len_action_target"] = len_action_predict self.makespanTarget = max(List_endStep_target) - min(List_startStep_target) # print("Makespane(target):{} \n Flowtime(target): {} \n ").format(self.makespanTarget, self.flowtimeTarget) break def getOptimalityMetrics(self): return [self.makespanPredict, self.makespanTarget], [self.flowtimePredict, self.flowtimeTarget] def getMaxstep(self): return self.maxstep def getMapsize(self): return self.size_map def initCommunicationRadius(self): self.communicationRadius = self.config.commR # self.communicationRadius = 5 # self.communicationRadius = 6 # self.communicationRadius = 7 # self.communicationRadius = 8 # self.communicationRadius = 9 # self.communicationRadius = 10 def reachObstacle(self, state): reach_obstacle = False # name_agent = "agent{}".format(id_agent) currentState_predict = state #self.status_MultiAgent[name_agent]["currentState"] currentStateIndexX = currentState_predict[0][0] currentStateIndexY = currentState_predict[0][1] # print(self.channel_map.shape) # print(self.channel_map) # time.sleep(10) if self.channel_map[int(currentStateIndexX)][int(currentStateIndexY)] == 1: # print('Reach obstacle.') reach_obstacle = True else: reach_obstacle = False # if reach_obstacle: # break return reach_obstacle def reachEdge(self, state): reach_edge = False # name_agent = "agent{}".format(id_agent) currentState_predict = state #self.status_MultiAgent[name_agent]["currentState"] currentStateIndexX = currentState_predict[0][0] currentStateIndexY = currentState_predict[0][1] if currentStateIndexX >= self.size_map[0] or currentStateIndexX < 0 or currentStateIndexY >= self.size_map[1] or currentStateIndexY < 0: # print('Reach edge.') reach_edge = True # break else: reach_edge = False return reach_edge def computeAdjacencyMatrix_fixedCommRadius(self, step, agentPos, CommunicationRadius, graphConnected=False): len_TimeSteps = agentPos.shape[0] # length of timesteps nNodes = agentPos.shape[1] # Number of nodes # Create the space to hold the adjacency matrices W = np.zeros([len_TimeSteps, nNodes, nNodes]) # Initial matrix distances = squareform(pdist(agentPos[0])) # nNodes x nNodes # I will increase the communication radius by 10% each time, # but I have to do it consistently within the while loop, # so in order to not affect the first value set of communication radius, I will account for that initial 10% outside distances = squareform(pdist(agentPos[0])) # nNodes x nNodes W[0] = (distances < self.communicationRadius).astype(agentPos.dtype) W[0] = W[0] - np.diag(np.diag(W[0])) graphConnected = graph.isConnected(W[0]) deg = np.sum(W[0], axis=1) # nNodes (degree vector) zeroDeg = np.nonzero(np.abs(deg) < self.zeroTolerance)[0] deg[zeroDeg] = 1. invSqrtDeg = np.sqrt(1. / deg) invSqrtDeg[zeroDeg] = 0. Deg = np.diag(invSqrtDeg) W[0] = Deg @ W[0] @ Deg return W, self.communicationRadius, graphConnected def computeAdjacencyMatrix(self, step, agentPos, CommunicationRadius, graphConnected=False): len_TimeSteps = agentPos.shape[0] # length of timesteps nNodes = agentPos.shape[1] # Number of nodes # Create the space to hold the adjacency matrices W = np.zeros([len_TimeSteps, nNodes, nNodes]) # Initial matrix distances = squareform(pdist(agentPos[0])) # nNodes x nNodes # I will increase the communication radius by 10% each time, # but I have to do it consistently within the while loop, # so in order to not affect the first value set of communication radius, I will account for that initial 10% outside if step == 0: self.communicationRadius = self.communicationRadius / 1.1 while graphConnected is False: self.communicationRadius = self.communicationRadius * 1.1 W[0] = (distances < self.communicationRadius).astype(agentPos.dtype) W[0] = W[0] - np.diag(np.diag(W[0])) graphConnected = graph.isConnected(W[0]) # And once we have found a connected initial position, we normalize it deg = np.sum(W[0], axis=1) # nNodes (degree vector) zeroDeg = np.nonzero(np.abs(deg) < self.zeroTolerance)[0] deg[zeroDeg] = 1. invSqrtDeg = np.sqrt(1. / deg) invSqrtDeg[zeroDeg] = 0. Deg = np.diag(invSqrtDeg) W[0] = Deg @ W[0] @ Deg # And once we have found a communication radius that makes the initial graph connected, # just follow through with the rest of the times, with that communication radius else: distances = squareform(pdist(agentPos[0])) # nNodes x nNodes W[0] = (distances < self.communicationRadius).astype(agentPos.dtype) W[0] = W[0] - np.diag(np.diag(W[0])) graphConnected = graph.isConnected(W[0]) deg = np.sum(W[0], axis=1) # nNodes (degree vector) zeroDeg = np.nonzero(np.abs(deg) < self.zeroTolerance)[0] deg[zeroDeg] = 1. invSqrtDeg = np.sqrt(1. / deg) invSqrtDeg[zeroDeg] = 0. Deg = np.diag(invSqrtDeg) W[0] = Deg @ W[0] @ Deg return W, self.communicationRadius, graphConnected def get_PosAgents(self): list_PosAgents = [] action_CurrentAgents=[] for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) currentState_predict = self.status_MultiAgent[name_agent]["currentState"] currentPredictIndexX = int(currentState_predict[0][0]) currentPredictIndexY = int(currentState_predict[0][1]) action_CurrentAgents.append([currentPredictIndexX, currentPredictIndexY]) list_PosAgents.append(action_CurrentAgents) return np.asarray(list_PosAgents) def getGSO(self, step): store_PosAgents = self.get_PosAgents() if step == 0: self.initCommunicationRadius() # print("{} - Step-{} - initCommunication Radius:{}".format(self.ID_dataset, step, self.communicationRadius)) # comm radius fixed # GSO, communicationRadius, graphConnected = self.computeAdjacencyMatrix_fixedCommRadius(step, store_PosAgents, self.communicationRadius) # comm radius that ensure initial graph connected GSO, communicationRadius, graphConnected = self.computeAdjacencyMatrix(step, store_PosAgents, self.communicationRadius) GSO_tensor = torch.from_numpy(GSO) self.store_GSO.append(GSO) self.store_communication_radius.append(communicationRadius) # print("################## currentstep {} - size of GSO: {}".format(step, len(self.store_GSO))) # print("{} - Step-{} - Communication Radius:{} - graphConnected:{}".format(self.ID_dataset, step, communicationRadius, graphConnected)) return GSO_tensor def getCurrentState__(self): tensor_currentState = torch.zeros([1, self.config.num_agents, 3, self.size_map[0], self.size_map[1]]) # tensor_currentState_all = torch.zeros([1, self.size_map[0], self.size_map[1]]) for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) goal_CurrentAgent = self.status_MultiAgent[name_agent]["goal"] goalIndexX = int(goal_CurrentAgent[0][0]) goalIndexY = int(goal_CurrentAgent[0][1]) channel_goal = torch.zeros([self.size_map[0], self.size_map[1]]) currentState_predict = self.status_MultiAgent[name_agent]["currentState"] currentPredictIndexX = int(currentState_predict[0][0]) currentPredictIndexY = int(currentState_predict[0][1]) channel_state = torch.zeros([self.size_map[0], self.size_map[1]]) channel_goal[goalIndexX][goalIndexY] = 1 channel_state[currentPredictIndexX][currentPredictIndexY] = 1 tensor_currentState[0, id_agent, 0, :, :] = self.channel_map tensor_currentState[0, id_agent, 1, :, :] = channel_goal tensor_currentState[0, id_agent, 2, :, :] = channel_state # tensor_currentState_allagents = torch.add(tensor_currentState_all, channel_state) # print(tensor_currentState_allagents) return tensor_currentState def getCurrentState(self, return_GPos=False): store_goalAgents = torch.zeros([self.config.num_agents, 2]) store_stateAgents = torch.zeros([self.config.num_agents, 2]) for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) goal_CurrentAgent = self.status_MultiAgent[name_agent]["goal"] goalIndexX = int(goal_CurrentAgent[0][0]) goalIndexY = int(goal_CurrentAgent[0][1]) store_goalAgents[id_agent,:] = torch.FloatTensor([goalIndexX,goalIndexY]) currentState_predict = self.status_MultiAgent[name_agent]["currentState"] currentPredictIndexX = int(currentState_predict[0][0]) currentPredictIndexY = int(currentState_predict[0][1]) store_stateAgents[id_agent, :] = torch.FloatTensor([currentPredictIndexX, currentPredictIndexY]) tensor_currentState = self.AgentState.toInputTensor(store_goalAgents, store_stateAgents) tensor_currentState = tensor_currentState.unsqueeze(0) # print(tensor_currentState_allagents) if return_GPos: return tensor_currentState, store_stateAgents.unsqueeze(0) else: return tensor_currentState def getCurrentState_(self): tensor_currentState = self.AgentState.toInputTensor(self.store_goalAgents, self.store_stateAgents) tensor_currentState = tensor_currentState.unsqueeze(0) return tensor_currentState def interRobotCollision(self): # collision = 0 collision = False allagents_pos = {} list_pos = [] for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) nextstate_currrentAgent = self.status_MultiAgent[name_agent]["nextState_predict"].tolist() list_pos.append(nextstate_currrentAgent) allagents_pos.update({id_agent: nextstate_currrentAgent}) for i in range(self.config.num_agents): pos = list_pos[i] count_collision = list_pos.count(pos) if count_collision > 1: collision = True collided_agents = [] for id_agent, pos_agent in allagents_pos.items(): if pos_agent == pos: name_agent = "agent{}".format(id_agent) collided_agents.append(name_agent) # id_agent2move = max(heuristic_agents.items(), key=operator.itemgetter(1))[0] id_agent2move = random.choice(collided_agents) # print("In {}, {} need to move".format(collided_agents, id_agent2move)) for name_agent in collided_agents: # name_agent = "agent{}".format(id_agent) # print("The action list of {}:\n{}".format(name_agent,self.status_MultiAgent[name_agent]["action_predict"])) list_actionKey_predict = self.status_MultiAgent[name_agent]["action_predict"] if list_actionKey_predict[-1] == self.stopKeyValue: # print('##### one of the agent has stoppted.#####') for name_agent in collided_agents: list_actionKey_predict = self.status_MultiAgent[name_agent]["action_predict"] list_actionKey_predict[-1] = self.stopKeyValue self.status_MultiAgent[name_agent]["action_predict"] = list_actionKey_predict self.status_MultiAgent[name_agent]["nextState_predict"] = self.status_MultiAgent[name_agent]["currentState"] # print("All agents {} stops:\n{}\n".format(name_agent,self.status_MultiAgent[name_agent]["action_predict"])) id_agent = int(name_agent.replace("agent","")) list_pos[id_agent] = self.status_MultiAgent[name_agent]["nextState_predict"].tolist() else: if name_agent != id_agent2move: list_actionKey_predict = self.status_MultiAgent[name_agent]["action_predict"] list_actionKey_predict[-1] = self.stopKeyValue self.status_MultiAgent[name_agent]["action_predict"] = list_actionKey_predict self.status_MultiAgent[name_agent]["nextState_predict"] = self.status_MultiAgent[name_agent]["currentState"] id_agent = int(name_agent.replace("agent","")) list_pos[id_agent] = self.status_MultiAgent[name_agent]["nextState_predict"].tolist() # print('{} stop'.format(name_agent)) # print("The action list of {} after changed:\n{}\n".format(name_agent, self.status_MultiAgent[name_agent]["action_predict"])) ## position swap list_nextpos = [] for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) nextstate_currrentAgent = self.status_MultiAgent[name_agent]["nextState_predict"].tolist() list_nextpos.append(nextstate_currrentAgent) for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) currentstate_currrentAgent = self.status_MultiAgent[name_agent]["currentState"].tolist() if currentstate_currrentAgent in list_nextpos: id_agent_swap = list_nextpos.index(currentstate_currrentAgent) name_agent_swap = "agent{}".format(id_agent_swap) if name_agent_swap != name_agent: if self.status_MultiAgent[name_agent_swap]["currentState"].tolist() == self.status_MultiAgent[name_agent]["nextState_predict"].tolist(): # print("In #{}case(test), {} and {} swap position happens.".format(self.ID_dataset,name_agent,name_agent_swap)) self.status_MultiAgent[name_agent]["nextState_predict"] = self.status_MultiAgent[name_agent]["currentState"] self.status_MultiAgent[name_agent_swap]["nextState_predict"] = self.status_MultiAgent[name_agent_swap]["currentState"] id_agent = int(name_agent.replace("agent","")) list_pos[id_agent] = self.status_MultiAgent[name_agent]["nextState_predict"].tolist() id_agent_swap = int(name_agent_swap.replace("agent","")) list_pos[id_agent_swap] = self.status_MultiAgent[name_agent_swap]["nextState_predict"].tolist() self.status_MultiAgent[name_agent]["action_predict"][-1] = self.stopKeyValue self.status_MultiAgent[name_agent_swap]["action_predict"][-1] = self.stopKeyValue collision = True return collision def heuristic(self, current_pos, goal): value = abs(goal[0] - current_pos[0]) + abs(goal[1] - current_pos[1]) return value def move(self, actionVec, currentstep, rouge_agent_count=None): allReachGoal = all(self.count_reachgoal) allReachGoal_withoutcollision = False self.check_predictCollsion = False self.check_moveCollision = False if (not allReachGoal) or (currentstep < self.maxstep): # if not allReachGoal and currentstep < self.maxstep: for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) # disabled_actionPredict_currentAgent = self.status_MultiAgent[name_agent]["disabled_action_predict"] # if self.config.num_agents == 1: # actionVec_predict_CurrentAgents = torch.mul(self.fun_Softmax(actionVec), # disabled_actionPredict_currentAgent) # else: # # actionVec_tmp = actionVec[id_agent] # actionVec_current = self.fun_Softmax(actionVec[id_agent]) # actionVec_predict_CurrentAgents = torch.mul(actionVec_current, disabled_actionPredict_currentAgent) actionVec_current = self.fun_Softmax(actionVec[id_agent]) if rouge_agent_count and id_agent < rouge_agent_count and currentstep % 2 == 1: # if this is a rouge agent, randomly select an action every other step actionKey_predict = torch.tensor([np.random.randint(5)]) else: actionKey_predict = torch.max(actionVec_current, 1)[1] # set flag of the timestep that agent start to move check_move = (actionKey_predict != self.stopKeyValue) startStep_action = self.status_MultiAgent[name_agent]["startStep_action_predict"] if check_move == 1 and startStep_action is None: self.status_MultiAgent[name_agent]["startStep_action_predict"] = currentstep - 1 list_actionKey_predict = self.status_MultiAgent[name_agent]["action_predict"] currentState_predict = self.status_MultiAgent[name_agent]["currentState"] nextState_predict = torch.add(currentState_predict, self.delta[actionKey_predict]) # ----- check edge and obstacle checkEdge = self.reachEdge(nextState_predict) checkObstacle = False if not checkEdge: checkObstacle = self.reachObstacle(nextState_predict) if checkEdge or checkObstacle: # print('Reach obstacle or edge.') # break # todo : remove the collision motion disabled # disabled_actionPredict_currentAgent[actionKey_predict] = 0.0 # self.status_MultiAgent[name_agent]["disabled_action_predict"] = disabled_actionPredict_currentAgent # self.move(actionVec, currentstep) self.check_predictCollsion = True list_actionKey_predict.append(self.stopKeyValue) self.status_MultiAgent[name_agent]["action_predict"] = list_actionKey_predict self.status_MultiAgent[name_agent]["nextState_predict"] = currentState_predict # self.check_predictEdgeCollsion else: # self.status_MultiAgent[name_agent]["currentState"] = nextState_predict self.status_MultiAgent[name_agent]["nextState_predict"] = nextState_predict # self.status_MultiAgent[name_agent]["disabled_action_predict"] = self.reset_disabled_action list_actionKey_predict.append(actionKey_predict[0]) self.status_MultiAgent[name_agent]["action_predict"] = list_actionKey_predict # if not self.check_predictCollsion: detect_interRobotCollision = self.interRobotCollision() # while detect_interRobotCollision: for _ in range(self.config.num_agents): # print('Collision happens.') if detect_interRobotCollision: detect_interRobotCollision = self.interRobotCollision() self.check_predictCollsion = True # print("Collision happens") else: # print("Collision avoided by collision shielding") break self.check_moveCollision = self.interRobotCollision() for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) nextState_predict = self.status_MultiAgent[name_agent]["nextState_predict"] self.status_MultiAgent[name_agent]["currentState"] = nextState_predict # self.store_stateAgents[id_agent,:] = nextState_predict path_predict = self.status_MultiAgent[name_agent]["path_predict"] path_predict.update({currentstep: nextState_predict}) # print("################## Current Step:{} - size of path_predict: {}".format(currentstep, len(path_predict))) self.status_MultiAgent[name_agent]["path_predict"] = path_predict goal_CurrentAgent = self.status_MultiAgent[name_agent]["goal"] goalIndexX = int(goal_CurrentAgent[0][0]) goalIndexY = int(goal_CurrentAgent[0][1]) if nextState_predict[0][0] == goalIndexX and nextState_predict[0][1] == goalIndexY and not \ self.count_reachgoal[id_agent]: self.count_reachgoal[id_agent] = True self.status_MultiAgent[name_agent]["endStep_action_predict"] = currentstep if currentstep >= (self.maxstep) and not self.count_reachgoal[id_agent]: # self.count_reachgoal[id_agent] = False self.status_MultiAgent[name_agent]["endStep_action_predict"] = currentstep # print("\t \t {} - status(Reach Goal) - {}".format(name_agent, self.count_reachgoal[id_agent])) if self.status_MultiAgent[name_agent]["startStep_action_predict"] is None: self.status_MultiAgent[name_agent]["startStep_action_predict"] = 0 #currentstep # if allReachGoal or (currentstep >= self.maxstep): List_endStep = [] List_startStep = [] self.flowtimePredict = 0 if rouge_agent_count: self.nonRogueFlowtimePredict = 0 for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) List_endStep.append(self.status_MultiAgent[name_agent]["endStep_action_predict"]) List_startStep.append(self.status_MultiAgent[name_agent]["startStep_action_predict"]) if self.config.distribution == "default": self.flowtimePredict += self.status_MultiAgent[name_agent]["endStep_action_predict"] - \ self.status_MultiAgent[name_agent]["startStep_action_predict"] # keep track of flowtime of non rogue agents if rouge_agent_count and id_agent >= rouge_agent_count: self.nonRogueFlowtimePredict += self.status_MultiAgent[name_agent]["endStep_action_predict"] - \ self.status_MultiAgent[name_agent]["startStep_action_predict"] len_action_predict = self.status_MultiAgent[name_agent]["endStep_action_predict"] - \ self.status_MultiAgent[name_agent]["startStep_action_predict"] self.status_MultiAgent[name_agent]["len_action_predict"] = len_action_predict if self.config.distribution == "default": self.makespanPredict = max(List_endStep) - min(List_startStep) if rouge_agent_count: self.nonRogueFlowtimePredict /= (self.config.num_agents - rouge_agent_count) # if not self.check_predictCollsion: # if not self.check_moveCollision: # print("##################### Find collisionfreeSol #############################") # # allReachGoal_withoutcollision = True # print("Makespane(predict):{} \n Flowtime(predict): {} \n ".format(self.makespanPredict , self.flowtimePredict)) return allReachGoal, self.check_moveCollision, self.check_predictCollsion def count_numAgents_ReachGoal(self): return self.count_reachgoal.count(True) def count_GSO_communcationRadius(self, step): _ = self.getGSO(step) return self.store_GSO, self.store_communication_radius def save_failure_cases(self): inputfile_name = os.path.join(self.failureCases_input,'failureCases_ID{:05d}.yaml'.format(self.ID_dataset)) print('############## failureCases in training set ID{} ###############'.format(self.ID_dataset)) f = open(inputfile_name, 'w') f.write("map:\n") f.write(" dimensions: {}\n".format([self.size_map[0], self.size_map[1]])) f.write(" obstacles:\n") for ID_obs in range(self.numObstacle): obstacleIndexX = int(self.posObstacle[ID_obs][0].cpu().detach().numpy()) obstacleIndexY = int(self.posObstacle[ID_obs][1].cpu().detach().numpy()) list_obs = [obstacleIndexX,obstacleIndexY] f.write(" - {}\n".format(list_obs)) f.write("agents:\n") for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) log_goal_currentAgent = self.status_MultiAgent[name_agent]["goal"].cpu().detach().numpy() log_currentState_currentAgent = self.status_MultiAgent[name_agent]["nextState_predict"].cpu().detach().numpy() goalX = int(log_goal_currentAgent[0][0]) goalY = int(log_goal_currentAgent[0][1]) currentStateX = int(log_currentState_currentAgent[0][0]) currentStateY = int(log_currentState_currentAgent[0][1]) goal_currentAgent = [goalX, goalY] currentState_currentAgent = [currentStateX, currentStateY] f.write(" - name: agent{}\n start: {}\n goal: {}\n".format(id_agent, currentState_currentAgent, goal_currentAgent)) f.close() def save_success_cases(self, mode): inputfile_name = os.path.join(self.config.result_AnimeDemo_dir_input, '{}Cases_ID{:05d}.yaml'.format(mode, self.ID_dataset)) if mode == 'success': outputfile_name = os.path.join(self.config.result_AnimeDemo_dir_predict_success, '{}Cases_ID{:05d}.yaml'.format(mode,self.ID_dataset)) checkSuccess = 1 else: outputfile_name = os.path.join(self.config.result_AnimeDemo_dir_predict_failure, '{}Cases_ID{:05d}.yaml'.format(mode, self.ID_dataset)) checkSuccess = 0 targetfile_name = os.path.join(self.config.result_AnimeDemo_dir_target, '{}Cases_ID{:05d}.yaml'.format(mode, self.ID_dataset)) gsofile_name = os.path.join(self.config.result_AnimeDemo_dir_GSO, '{}Cases_ID{:05d}.mat'.format(mode, self.ID_dataset)) save_statistics_GSO = {'gso':self.store_GSO, 'commRadius': self.store_communication_radius} sio.savemat(gsofile_name, save_statistics_GSO) # print('############## successCases in training set ID{} ###############'.format(self.ID_dataset)) f = open(inputfile_name, 'w') f.write("map:\n") f.write(" dimensions: {}\n".format([self.size_map[0], self.size_map[1]])) f.write(" obstacles:\n") for ID_obs in range(self.numObstacle): obstacleIndexX = int(self.posObstacle[ID_obs][0].cpu().detach().numpy()) obstacleIndexY = int(self.posObstacle[ID_obs][1].cpu().detach().numpy()) list_obs = [obstacleIndexX, obstacleIndexY] f.write(" - {}\n".format(list_obs)) f.write("agents:\n") for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) log_goal_currentAgent = self.status_MultiAgent[name_agent]["goal"].cpu().detach().numpy() log_currentState_currentAgent = self.status_MultiAgent[name_agent]["start"].cpu().detach().numpy() goalX = int(log_goal_currentAgent[0][0]) goalY = int(log_goal_currentAgent[0][1]) startX = int(log_currentState_currentAgent[0][0]) startY = int(log_currentState_currentAgent[0][1]) goal_currentAgent = [goalX, goalY] currentState_currentAgent = [startX, startY] f.write(" - name: agent{}\n start: {}\n goal: {}\n".format(id_agent, currentState_currentAgent, goal_currentAgent)) f.close() f_sol = open(outputfile_name, 'w') f_sol.write("statistics:\n") f_sol.write(" cost: {}\n".format(self.flowtimePredict)) f_sol.write(" makespan: {}\n".format(self.makespanPredict)) f_sol.write(" succeed: {}\n".format(int(checkSuccess))) f_sol.write("schedule:\n") for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) # print(self.status_MultiAgent[name_agent]["path_predict"]) path = self.status_MultiAgent[name_agent]["path_predict"] len_path = len(path) f_sol.write(" agent{}:\n".format(id_agent)) for step in range(len_path): pathIndexX = int(path[step][0][0].cpu().detach().numpy()) pathIndexY = int(path[step][0][1].cpu().detach().numpy()) f_sol.write(" - x: {}\n y: {}\n t: {}\n".format(pathIndexX,pathIndexY, step)) f_sol.close() f_target = open(targetfile_name, 'w') f_target.write("statistics:\n") f_target.write(" cost: {}\n".format(self.flowtimeTarget)) f_target.write(" makespan: {}\n".format(self.makespanTarget)) f_target.write("schedule:\n") for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) # print(self.status_MultiAgent[name_agent]["path_predict"]) path = self.status_MultiAgent[name_agent]["path_target"] len_path = len(path) f_target.write(" agent{}:\n".format(id_agent)) for step in range(len_path): pathIndexX = int(path[step][0][0].cpu().detach().numpy()) pathIndexY = int(path[step][0][1].cpu().detach().numpy()) f_target.write(" - x: {}\n y: {}\n t: {}\n".format(pathIndexX, pathIndexY, step)) f_target.close() def createfolder_failure_cases(self): self.failureCases_input = self.config.failCases_dir + 'input/' self.dir_sol = os.path.join(self.config.failCases_dir, "output_ECBS/") if os.path.exists(self.failureCases_input) and os.path.isdir(self.failureCases_input): shutil.rmtree(self.failureCases_input) if os.path.exists(self.dir_sol) and os.path.isdir(self.dir_sol): shutil.rmtree(self.dir_sol) try: # Create target Directory os.makedirs(self.failureCases_input) except FileExistsError: # print("Directory ", dirName, " already exists") pass def checkOptimality(self, collisionFreeSol): if self.makespanPredict <= self.makespanTarget and self.flowtimePredict <= self.flowtimeTarget and collisionFreeSol: findOptimalSolution = True else: findOptimalSolution = False return findOptimalSolution, [self.makespanPredict, self.makespanTarget], [self.flowtimePredict, self.flowtimeTarget] def draw(self, ID_dataset): status_MultiAgent = {} status_MultiAgent_Target = {} status_MultiAgent_Predict = {} for id_agent in range(self.config.num_agents): name_agent = "agent{}".format(id_agent) status_CurrentAgents_Target = {"goal": self.status_MultiAgent[name_agent]["goal"], "start": self.status_MultiAgent[name_agent]["start"], "path": self.status_MultiAgent[name_agent]["path_target"], "action": self.status_MultiAgent[name_agent]["action_target"], "len_action": self.status_MultiAgent[name_agent]["len_action_target"] } status_CurrentAgents_Predict = {"goal": self.status_MultiAgent[name_agent]["goal"], "start": self.status_MultiAgent[name_agent]["start"], "path": self.status_MultiAgent[name_agent]["path_predict"], "action": self.status_MultiAgent[name_agent]["action_predict"], "len_action": self.status_MultiAgent[name_agent]["len_action_predict"] } status_MultiAgent_Target.update({name_agent: status_CurrentAgents_Target}) status_MultiAgent_Predict.update({name_agent: status_CurrentAgents_Predict}) status_MultiAgent_Target.update({"makespan": self.makespanTarget, "flowtime": self.flowtimeTarget}) status_MultiAgent_Predict.update({"makespan": self.makespanPredict, "flowtime": self.flowtimePredict}) status_MultiAgent.update({"target":status_MultiAgent_Target, "predict":status_MultiAgent_Predict}) draw = DrawpathCombine(self.config, self.channel_map, self.posObstacle, status_MultiAgent) draw.draw(ID_dataset) draw.save()
class DataTransformer: def __init__(self, config): self.config = config self.PROCESS_NUMBER = 4 self.num_agents = self.config.num_agents self.size_map = [self.config.map_w, self.config.map_w] self.label_density = str(self.config.map_density).split('.')[-1] self.AgentState = AgentState(self.num_agents) self.communicationRadius = 5 # communicationRadius self.zeroTolerance = 1e-9 self.delta = [[-1, 0], # go up [0, -1], # go left [1, 0], # go down [0, 1], # go right [0, 0]] # stop self.num_actions = 5 self.list_seqtrain_file = [] self.list_train_file = [] self.list_seqvalid_file = [] self.list_validStep_file = [] self.list_valid_file = [] self.list_test_file = [] self.hashids = Hashids(alphabet='01234567789abcdef', min_length=5) self.pathtransformer = self.pathtransformer_RelativeCoordinate self.label_setup = '{}{:02d}x{:02d}_density_p{}/{}_Agent'.format(self.config.loadmap_TYPE, self.size_map[0],self.size_map[1], self.label_density, self.num_agents) self.dirName_parent = os.path.join(self.config.solCases_dir, self.label_setup) self.dirName_Store = os.path.join(self.config.dir_SaveData, self.label_setup) self.dirName_input = os.path.join(self.dirName_parent, 'input') self.dirName_output = os.path.join(self.dirName_parent, 'output_{}'.format(config.chosen_solver)) self.set_up() def set_up(self): self.list_failureCases_solution = self.search_failureCases(self.dirName_output) self.list_failureCases_input = self.search_failureCases(self.dirName_input) self.nameprefix_input = self.list_failureCases_input[0].split('input/')[-1].split('ID')[0] self.list_failureCases_solution = sorted(self.list_failureCases_solution) self.len_failureCases_solution = len(self.list_failureCases_solution) def reset(self): self.task_queue = Queue() dirpath = self.dirName_Store if os.path.exists(dirpath) and os.path.isdir(dirpath): shutil.rmtree(dirpath) self.path_save_solDATA = self.dirName_Store try: # Create target Directory os.makedirs(self.path_save_solDATA) os.makedirs(os.path.join(self.path_save_solDATA, 'train')) os.makedirs(os.path.join(self.path_save_solDATA, 'valid')) os.makedirs(os.path.join(self.path_save_solDATA, 'test')) except FileExistsError: # print("Directory ", dirName, " already exists") pass def solutionTransformer(self): # div_train = 21000 # div_valid = 61 # div_test = 4500 # div_train = 0 # div_valid = 0 # div_test = 1500 div_train = self.config.div_train div_valid = self.config.div_valid div_test = self.config.div_test # div_train = 5 # div_valid = 2 # div_test = 2 num_used_data = div_train + div_valid + div_test num_data_loop = min(num_used_data, self.len_failureCases_solution) # for id_sol in range(num_data_loop): for id_sol in range(self.config.id_start, num_data_loop): if id_sol < div_train: mode = "train" case_config = (mode, id_sol) self.task_queue.put(case_config) elif id_sol < (div_train+div_valid): mode = "valid" case_config = (mode, id_sol) self.task_queue.put(case_config) elif id_sol <= num_used_data: mode = "test" case_config = (mode, id_sol) self.task_queue.put(case_config) time.sleep(0.3) processes = [] for i in range(self.PROCESS_NUMBER): # Run Multiprocesses p = Process(target=self.compute_thread, args=(str(i))) processes.append(p) [x.start() for x in processes] def compute_thread(self,thread_id): while True: try: case_config = self.task_queue.get(block=False) (mode, id_sol) = case_config print('thread {} get task:{} - {}'.format(thread_id, mode, id_sol)) self.pipeline(case_config) except: # print('thread {} no task, exit'.format(thread_id)) return def pipeline(self, case_config): (mode, id_sol) = case_config agents_schedule, agents_goal, makespan, map_data, id_case = self.load_ExpertSolution(id_sol) # agents_schedule, agents_goal, makespan, map_data, id_case = self.load_ExpertSolution_(id_sol) log_str = 'Transform_failureCases_ID_#{} in MAP_ID{}'.format(id_case[1],id_case[0]) print('############## {} ###############'.format(log_str)) # print(agents_schedule) if mode == "train" or mode == "valid": self.pathtransformer(map_data, agents_schedule, agents_goal, makespan + 1, id_case, mode) else: self.pathtransformer_test(map_data, agents_schedule, agents_goal, makespan + 1, id_case, mode) def load_ExpertSolution(self, ID_case): name_solution_file = self.list_failureCases_solution[ID_case] # id_solved_case = name_solution_file.split('_ID')[-1].split('.yaml')[0] map_setup = name_solution_file.split('output_')[-1].split('_IDMap')[0] id_sol_map = name_solution_file.split('_IDMap')[-1].split('_IDCase')[0] id_sol_case = name_solution_file.split('_IDCase')[-1].split('_')[0] name_inputfile = os.path.join(self.dirName_input, 'input_{}_IDMap{}_IDCase{}.yaml'.format(map_setup, id_sol_map, id_sol_case)) # print(name_inputfile) # print(name_solution_file) with open(name_inputfile, 'r') as stream: try: # print(yaml.safe_load(stream)) data_config = yaml.safe_load(stream) except yaml.YAMLError as exc: print(exc) with open(name_solution_file, 'r') as stream: try: # print(yaml.safe_load(stream)) data_output = yaml.safe_load(stream) except yaml.YAMLError as exc: print(exc) agentsConfig = data_config['agents'] num_agent = len(agentsConfig) list_posObstacle = data_config['map']['obstacles'] if list_posObstacle == None: map_data = np.zeros(self.size_map, dtype=np.int64) else: map_data = self.setup_map(list_posObstacle) schedule = data_output['schedule'] makespan = data_output['statistics']['makespan'] goal_allagents = np.zeros([num_agent, 2]) schedule_agentsState = np.zeros([makespan + 1, num_agent, 2]) schedule_agentsActions = np.zeros([makespan + 1, num_agent, self.num_actions]) schedule_agents = [schedule_agentsState, schedule_agentsActions] hash_ids = np.zeros(self.num_agents) for id_agent in range(num_agent): goalX = agentsConfig[id_agent]['goal'][0] goalY = agentsConfig[id_agent]['goal'][1] goal_allagents[id_agent][:] = [goalX, goalY] schedule_agents = self.obtainSchedule(id_agent, schedule, schedule_agents, goal_allagents, makespan + 1) str_id = '{}_{}_{}'.format(id_sol_map,id_sol_case,id_agent) int_id = int(hashlib.sha256(str_id.encode('utf-8')).hexdigest(), 16) % (10 ** 5) # hash_ids[id_agent]=np.divide(int_id,10**5) hash_ids[id_agent] = int_id # print(id_sol_map, id_sol_case, hash_ids) return schedule_agents, goal_allagents, makespan, map_data, (id_sol_map, id_sol_case, hash_ids) def load_ExpertSolution_(self, ID_case): name_solution_file = self.list_failureCases_solution[ID_case] id_sol_case = name_solution_file.split('_ID')[-1].split('.yaml')[0] map_setup = 'demo' id_sol_map = '0' name_inputfile = os.path.join(self.dirName_input, 'failureCases_ID{}.yaml'.format(id_sol_case)) # print(name_inputfile) # print(name_solution_file) with open(name_inputfile, 'r') as stream: try: # print(yaml.safe_load(stream)) data_config = yaml.safe_load(stream) except yaml.YAMLError as exc: print(exc) with open(name_solution_file, 'r') as stream: try: # print(yaml.safe_load(stream)) data_output = yaml.safe_load(stream) except yaml.YAMLError as exc: print(exc) agentsConfig = data_config['agents'] num_agent = len(agentsConfig) list_posObstacle = data_config['map']['obstacles'] if list_posObstacle == None: map_data = np.zeros(self.size_map, dtype=np.int64) else: map_data = self.setup_map(list_posObstacle) schedule = data_output['schedule'] makespan = data_output['statistics']['makespan'] # print(schedule) goal_allagents = np.zeros([num_agent, 2]) schedule_agentsState = np.zeros([makespan + 1, num_agent, 2]) schedule_agentsActions = np.zeros([makespan + 1, num_agent, self.num_actions]) schedule_agents = [schedule_agentsState, schedule_agentsActions] hash_ids = np.zeros(self.num_agents) for id_agent in range(num_agent): goalX = agentsConfig[id_agent]['goal'][0] goalY = agentsConfig[id_agent]['goal'][1] goal_allagents[id_agent][:] = [goalX, goalY] schedule_agents = self.obtainSchedule(id_agent, schedule, schedule_agents, goal_allagents, makespan + 1) str_id = '{}_{}_{}'.format(id_sol_map, id_sol_case, id_agent) int_id = int(hashlib.sha256(str_id.encode('utf-8')).hexdigest(), 16) % (10 ** 5) # hash_ids[id_agent]=np.divide(int_id,10**5) hash_ids[id_agent] = int_id print(schedule_agents) # print(id_sol_map, id_sol_case, hash_ids) return schedule_agents, goal_allagents, makespan, map_data, (id_sol_map, id_sol_case, hash_ids) def obtainSchedule(self, id_agent, agentplan, schedule_agents, goal_allagents, teamMakeSpan): name_agent = "agent{}".format(id_agent) [schedule_agentsState, schedule_agentsActions] = schedule_agents planCurrentAgent = agentplan[name_agent] pathLengthCurrentAgent = len(planCurrentAgent) actionKeyListAgent = [] for step in range(teamMakeSpan): if step < pathLengthCurrentAgent: currentX = planCurrentAgent[step]['x'] currentY = planCurrentAgent[step]['y'] else: currentX = goal_allagents[id_agent][0] currentY = goal_allagents[id_agent][1] schedule_agentsState[step][id_agent][:] = [currentX, currentY] # up left down right stop actionVectorTarget = [0, 0, 0, 0, 0] # map action with respect to the change of position of agent if step < (pathLengthCurrentAgent - 1): nextX = planCurrentAgent[step + 1]['x'] nextY = planCurrentAgent[step + 1]['y'] # actionCurrent = [nextX - currentX, nextY - currentY] elif step >= (pathLengthCurrentAgent - 1): nextX = goal_allagents[id_agent][0] nextY = goal_allagents[id_agent][1] actionCurrent = [nextX - currentX, nextY - currentY] actionKeyIndex = self.delta.index(actionCurrent) actionKeyListAgent.append(actionKeyIndex) actionVectorTarget[actionKeyIndex] = 1 schedule_agentsActions[step][id_agent][:] = actionVectorTarget return [schedule_agentsState,schedule_agentsActions] def setup_map(self, list_posObstacle): num_obstacle = len(list_posObstacle) map_data = np.zeros(self.size_map) for ID_obs in range(num_obstacle): obstacleIndexX = list_posObstacle[ID_obs][0] obstacleIndexY = list_posObstacle[ID_obs][1] map_data[obstacleIndexX][obstacleIndexY] = 1 return map_data def pathtransformer_RelativeCoordinate(self, map_data, agents_schedule, agents_goal, makespan, ID_case, mode): # input: start and goal position, # output: a set of file, # each file consist of state (map. goal, state) and target (action for current state) [schedule_agentsState, schedule_agentsActions] = agents_schedule save_PairredData = {} # print(ID_case) # compute AdjacencyMatrix GSO, communicationRadius = self.computeAdjacencyMatrix(schedule_agentsState, self.communicationRadius) # transform into relative Coordinate, loop "makespan" times self.AgentState.setmap(map_data) input_seq_tensor = self.AgentState.toSeqInputTensor(agents_goal, schedule_agentsState, makespan) list_input = input_seq_tensor.cpu().detach().numpy() save_PairredData.update({'map': map_data, 'goal': agents_goal, 'inputState': schedule_agentsState, 'inputTensor': list_input, 'target': schedule_agentsActions, 'GSO': GSO,'makespan':makespan, 'HashIDs':ID_case[2]}) # print(save_PairredData) self.save(mode, save_PairredData, ID_case, makespan) print("Save as {}set_#{} from MAP ID_{}.".format(mode, ID_case[1], ID_case[0])) def pathtransformer_test(self, map_data, agents_schedule, agents_goal, makespan, ID_case, mode): # input: start and goal position, # output: a set of file, # each file consist of state (map. goal, state) and target (action for current state) [schedule_agentsState, schedule_agentsActions] = agents_schedule save_PairredData = {} save_PairredData.update({'map': map_data, 'goal': agents_goal, 'inputState': schedule_agentsState[0], 'target': schedule_agentsActions, 'makespan': makespan, 'HashIDs':ID_case[2]}) # print(save_PairredData) self.save(mode, save_PairredData, ID_case, makespan) print("Save as {}set_#{} from MAP ID_{}.".format(mode, ID_case[1], ID_case[0])) def save(self, mode, save_PairredData, ID_case, makespan): (id_sol_map, id_sol_case,_) = ID_case file_name = os.path.join(self.path_save_solDATA, mode,'{}_IDMap{}_IDCase{}_MP{}.mat'.format(mode, id_sol_map, id_sol_case, makespan)) # print(file_name) sio.savemat(file_name, save_PairredData) def record_pathdata(self, mode, ID_case, makespan): (id_sol_map, id_sol_case) = ID_case data_name_mat = '{}_IDMap{}_IDCase{}_MP{}.mat'.format(mode, id_sol_map, id_sol_case, makespan) if mode == "train": self.list_seqtrain_file.append([data_name_mat, makespan, 0]) # print("\n train --", self.list_seqtrain_file) for step in range(makespan): self.list_train_file.append([data_name_mat, step, 0]) elif mode =='validStep': self.list_seqvalid_file.append([data_name_mat, makespan, 0]) for step in range(makespan): self.list_validStep_file.append([data_name_mat, step, 0]) elif mode == "valid": self.list_valid_file.append([data_name_mat, makespan, 0]) # 0 elif mode == "test": self.list_test_file.append([data_name_mat, makespan, 0]) # 0 def save_filepath(self): dirName = self.path_save_solDATA file_seqtrain_name = os.path.join(dirName,'{}seq_filename.csv'.format('train')) with open(file_seqtrain_name, "w", newline="") as f: writer = csv.writer(f) print("\n train hello --", self.list_seqtrain_file) writer.writerows(self.list_seqtrain_file) file_train_name = os.path.join(dirName,'{}_filename.csv'.format('train')) with open(file_train_name, "w", newline="") as f: writer = csv.writer(f) writer.writerows(self.list_train_file) file_seqvalid_name = os.path.join(dirName,'{}seq_filename.csv'.format('valid')) with open(file_seqvalid_name, "w", newline="") as f: writer = csv.writer(f) writer.writerows(self.list_seqvalid_file) file_validStep_name = os.path.join(dirName,'{}_filename.csv'.format('validStep')) with open(file_validStep_name, "w", newline="") as f: writer = csv.writer(f) writer.writerows(self.list_validStep_file) file_valid_name = os.path.join(dirName,'{}_filename.csv'.format('valid')) with open(file_valid_name, "w", newline="") as f: writer = csv.writer(f) writer.writerows(self.list_valid_file) file_test_name = os.path.join(dirName,'{}_filename.csv'.format('test')) with open(file_test_name, "w", newline="") as f: writer = csv.writer(f) writer.writerows(self.list_test_file) def search_failureCases(self, dir): # make a list of file name of input yaml list_path = [] assert os.path.isdir(dir), '%s is not a valid directory' % dir for root, _, fnames in sorted(os.walk(dir)): for fname in fnames: if self.is_target_file(fname): path = os.path.join(root, fname) list_path.append(path) return list_path def is_target_file(self, filename): DATA_EXTENSIONS = ['.yaml'] return any(filename.endswith(extension) for extension in DATA_EXTENSIONS) def computeAdjacencyMatrix(self, pos, CommunicationRadius, connected=True): # First, transpose the axis of pos so that the rest of the code follows # through as legible as possible (i.e. convert the last two dimensions # from 2 x nNodes to nNodes x 2) # pos: TimeSteps x nAgents x 2 (X, Y) # Get the appropriate dimensions nSamples = pos.shape[0] len_TimeSteps = pos.shape[0] # length of timesteps nNodes = pos.shape[1] # Number of nodes # Create the space to hold the adjacency matrices W = np.zeros([len_TimeSteps, nNodes, nNodes]) threshold = CommunicationRadius # We compute a different # threshold for each sample, because otherwise one bad trajectory # will ruin all the adjacency matrices for t in range(len_TimeSteps): # Compute the distances distances = squareform(pdist(pos[t])) # nNodes x nNodes # Threshold them W[t] = (distances < threshold).astype(pos.dtype) # And get rid of the self-loops W[t] = W[t] - np.diag(np.diag(W[t])) # Now, check if it is connected, if not, let's make the # threshold bigger while (not graph.isConnected(W[t])) and (connected): # while (not graph.isConnected(W[t])) and (connected): # Increase threshold threshold = threshold * 1.1 # Increase 10% # Compute adjacency matrix W[t] = (distances < threshold).astype(pos.dtype) W[t] = W[t] - np.diag(np.diag(W[t])) # And since the threshold has probably changed, and we want the same # threshold for all nodes, we repeat: W = np.zeros([len_TimeSteps, nNodes, nNodes]) for t in range(len_TimeSteps): distances = squareform(pdist(pos[t])) W[t] = (distances < threshold).astype(pos.dtype) W[t] = W[t] - np.diag(np.diag(W[t])) # And, when we compute the adjacency matrix, we normalize it by # the degree deg = np.sum(W[t], axis=1) # nNodes (degree vector) # Build the degree matrix powered to the -1/2 Deg = np.diag(np.sqrt(1. / deg)) # And finally get the correct adjacency W[t] = Deg @ W[t] @ Deg return W, threshold
class CreateDataset(data.Dataset): def __init__(self, config, mode): self.config = config self.datapath_exp = '{}{:02d}x{:02d}_density_p{}/{}_Agent/'.format( self.config.map_type, self.config.map_w, self.config.map_h, self.config.map_density, self.config.num_agents) self.dirName = os.path.join(self.config.data_root, self.datapath_exp) self.AgentState = AgentState(self.config.num_agents) if mode == "train": self.dir_data = os.path.join(self.dirName, 'train') self.search_files = self.search_target_files_withStep self.data_paths, self.id_stepdata = self.update_data_path_trainingset( self.dir_data) self.load_data = self.load_train_data elif mode == "test_trainingSet": self.dir_data = os.path.join(self.dirName, 'train') # self.search_files = self.search_target_files # data_paths, id_stepdata = self.update_data_path_trainingset(self.dir_data) data_paths, id_stepdata = self.search_target_files(self.dir_data) paths_total = list(zip(data_paths, id_stepdata)) random.shuffle(paths_total) data_paths, id_stepdata = zip(*paths_total) self.data_paths = data_paths[:self.config.num_test_trainingSet] self.id_stepdata = id_stepdata[:self.config.num_test_trainingSet] self.load_data = self.load_data_during_training elif mode == "valid": self.dir_data = os.path.join(self.dirName, 'valid') self.data_paths, self.id_stepdata = self.obtain_data_path_validset( self.dir_data, self.config.num_validset) self.load_data = self.load_data_during_training elif mode == "validStep": self.dir_data = os.path.join(self.dirName, 'valid') self.data_paths, self.id_stepdata = self.search_valid_files_withStep( self.dir_data, self.config.num_validset) self.load_data = self.load_train_data elif mode == "test": self.dir_data = os.path.join(self.dirName, mode) self.data_paths, self.id_stepdata = self.obtain_data_path_validset( self.dir_data, self.config.num_testset) self.load_data = self.load_test_data self.data_size = len(self.data_paths) def __getitem__(self, index): path = self.data_paths[index % self.data_size] id_step = int(self.id_stepdata[index % self.data_size]) input, target, GSO, map_tensor = self.load_data(path, id_step) return input, target, id_step, GSO, map_tensor def update_data_path_trainingset(self, dir_data): # only used for training set and online expert - training purpose data_paths_total = [] step_paths_total = [] # load common training set (21000) data_paths, step_paths = self.search_files(dir_data) data_paths_total.extend(data_paths) step_paths_total.extend(step_paths) # load training set from online expert based on failCases data_paths_failcases, step_paths_failcases = self.search_files( self.config.failCases_dir) data_paths_total.extend(data_paths_failcases) step_paths_total.extend(step_paths_failcases) paths_total = list(zip(data_paths_total, step_paths_total)) random.shuffle(paths_total) data_paths_total, step_paths_total = zip(*paths_total) return data_paths_total, step_paths_total def obtain_data_path_validset(self, dir_data, case_limit): # obtain validation data to valid the decision making at given state data_paths, id_stepdata = self.search_target_files(dir_data) paths_bundle = list(zip(data_paths, id_stepdata)) paths_bundle = sorted(paths_bundle) data_paths, id_stepdata = zip(*paths_bundle) data_paths = data_paths[:case_limit] id_stepdata = id_stepdata[:case_limit] return data_paths, id_stepdata def load_train_data(self, path, id_step): data_contents = sio.loadmat(path) map_channel = data_contents['map'] # W x H input_tensor = data_contents[ 'inputTensor'] # step x num_agent x 3 x 11 x 11 target_sequence = data_contents['target'] # step x num_agent x 5 input_GSO_sequence = data_contents[ 'GSO'] # Step x num_agent x num_agent tensor_map = torch.from_numpy(map_channel).float() step_input_tensor = torch.from_numpy(input_tensor[id_step][:]).float() step_input_GSO = torch.from_numpy( input_GSO_sequence[id_step, :, :]).float() step_target = torch.from_numpy(target_sequence[id_step, :, :]).long() return step_input_tensor, step_target, step_input_GSO, tensor_map def load_data_during_training(self, path, _): # load dataset into validation mode during training - only initial position, predict action towards goal # test on training set and test on validation set data_contents = sio.loadmat(path) map_channel = data_contents['map'] # W x H goal_allagents = data_contents['goal'] # num_agent x 2 input_sequence = data_contents['inputState'][ 0] # from step x num_agent x 2 to # initial pos x num_agent x 2 target_sequence = data_contents['target'] # step x num_agent x 5 self.AgentState.setmap(map_channel) step_input_tensor = self.AgentState.stackinfo(goal_allagents, input_sequence) step_target = torch.from_numpy(target_sequence).long() # from step x num_agent x action (5) to id_agent x step x action(5) step_target = step_target.permute(1, 0, 2) step_input_rs = step_input_tensor.squeeze(0) step_target_rs = step_target.squeeze(0) tensor_map = torch.from_numpy(map_channel).float() GSO_none = torch.zeros(1) return step_input_rs, step_target_rs, GSO_none, tensor_map def load_test_data(self, path, _): # load dataset into test mode - only initial position, predict action towards goal data_contents = sio.loadmat(path) map_channel = data_contents['map'] # W x H goal_allagents = data_contents['goal'] # num_agent x 2 input_sequence = data_contents['inputState'] # num_agent x 2 target_sequence = data_contents['target'] # step x num_agent x 5 bots_to_change = 0 if int(self.config.distribution_num) != 0: bots_to_change = int(self.config.distribution_num) idxs = [] for i in range(bots_to_change): num = 0 idx = np.random.choice(range(len(input_sequence))) while (idx in idxs) and num < 10: idx = np.random.choice(range(len(input_sequence))) num += 1 if self.config.distribution == "LLQ": print("RUNNING LLQ") choice = [ np.random.choice(range(int(map_channel.shape[0] / 2))), np.random.choice(range(int(map_channel.shape[1] / 2))) ] num = 0 while (map_channel[choice[0]][choice[1]] == 1 or choice in input_sequence) and num < 10: choice = [ np.random.choice(range(int(map_channel.shape[0] / 2))), np.random.choice(range(int(map_channel.shape[1] / 2))) ] input_sequence[idx] = np.array(choice) if self.config.distribution == "edge": print("RUNNING EDGE") edge = np.random.choice([0, 1]) if edge == 0 and map_channel[0][int( input_sequence[idx][1])] != 1 and [ 0, int(input_sequence[idx][1]) ] not in input_sequence: input_sequence[idx][0] = 0 if edge == 1 and map_channel[int( input_sequence[idx][0])][0] != 1 and [ input_sequence[idx][0], 0 ] not in input_sequence: input_sequence[idx][1] = 0 self.AgentState.setmap(map_channel) step_input_tensor = self.AgentState.stackinfo(goal_allagents, input_sequence) step_target = torch.from_numpy(target_sequence).long() # from step x num_agent x action (5) to id_agent x step x action(5) step_target = step_target.permute(1, 0, 2) step_input_rs = step_input_tensor.squeeze(0) step_target_rs = step_target.squeeze(0) tensor_map = torch.from_numpy(map_channel).float() GSO_none = torch.zeros(1) return step_input_rs, step_target_rs, GSO_none, tensor_map def search_target_files(self, dir): # make a list of file name of input yaml list_path = [] list_path_stepdata = [] assert os.path.isdir(dir), '%s is not a valid directory' % dir for root, _, fnames in sorted(os.walk(dir)): for fname in fnames: if self.is_target_file(fname): makespan = int(fname.split('_MP')[-1].split('.mat')[0]) path = os.path.join(root, fname) list_path.append(path) list_path_stepdata.append(makespan) return list_path, list_path_stepdata def search_target_files_withStep(self, dir): # make a list of file name of input yaml list_path = [] list_path_stepdata = [] assert os.path.isdir(dir), '%s is not a valid directory' % dir for root, _, fnames in sorted(os.walk(dir)): for fname in fnames: if self.is_target_file(fname): makespan = int(fname.split('_MP')[-1].split('.mat')[0]) path = os.path.join(root, fname) for step in range(makespan): # path = os.path.join(root, fname, str(step)) list_path.append(path) list_path_stepdata.append(step) return list_path, list_path_stepdata def search_valid_files_withStep(self, dir, case_limit): # make a list of file name of input yaml list_path = [] list_path_stepdata = [] count_num_cases = 0 assert os.path.isdir(dir), '%s is not a valid directory' % dir for root, _, fnames in sorted(os.walk(dir)): for fname in fnames: if self.is_target_file(fname): makespan = int(fname.split('_MP')[-1].split('.mat')[0]) path = os.path.join(root, fname) if count_num_cases <= case_limit: for step in range(makespan): # path = os.path.join(root, fname, str(step)) list_path.append(path) list_path_stepdata.append(step) count_num_cases += 1 else: break return list_path, list_path_stepdata def is_target_file(self, filename): DATA_EXTENSIONS = ['.mat'] return any( filename.endswith(extension) for extension in DATA_EXTENSIONS) def __len__(self): return self.data_size