class SGDParallel(playdoh.ParallelTask): ''' ''' def initialize(self, nuse, totalNodes ): """ This method is called on each node when playdoh.start_task method is called. Each node initializes its own model. """ self.seed = 1742 self.nuse = nuse self.totalNodes = totalNodes #including master node self.debug = True self.nIter = 1000 self.nFetch = 15 #after how many steps(minibatch processing) should a worker fetch new parameters from master self.rnd = np.random.RandomState(self.seed) def start(self): if self.index == 0: self.startMasterNode() else: self.startWorkerNode() def startWorkerNode(self): '''Receives updated parameters from the master, keeps updating local parameters using its own subset of training data, pushes local parameters to master, all asynchronously ''' self.initMVRNN() #also initializes training and test sets self.func = costFn self.fprime = None self.optimizer = StochasticGradientDescent(niter=self.nIter , learning_rate=0.01, learningrateFactor=1.0, printAt10Iter='.', printAt100Iter='\n+') genObj = None while True: #fetch action and parameters from masters action, master_rnn = self.pop('fromMaster_' + str(self.index)) if(action == "finish"): print "Node: ",self.index," Received finish action from master, exiting..." return if self.debug: print "Node: ", self.index, " Fetched new action & parameters from Master.. " + str(master_rnn.W[0,0]) self.params.resetNumReducedWords() #to unroll all the words # W, WO, Wcat, Wv, Wo = unroll_theta(theta, self.params) self.copy_into_rnn(master_rnn.getParamsList()) #copy to local rnn to global theta = self.rnn.getTheta(self.params, self.all_train_idx) self.params.setNumReducedWords(len(self.all_train_idx)) #set number of reduced words, to be used by costfn for unrolling start_time = time.clock() try: genObj.next() except (AttributeError, StopIteration) : genObj = self.optimizer.minimizeBatchesPll(rnn=self.rnn, rnnData_train=self.rnnData_train, allTrainSentIdx=self.all_train_idx, params=self.params, x0=theta, func=costFn, fprime=None, rnnData_test=self.rnnData_dev, initialSetSize=1, niter=1, seed=self.seed, modelFileName='', printStatistics=False, modelSaveIter=10, nIterInPart=1, nFetch=self.nFetch, rnd=self.rnd, nodeid=self.index) #optimize this theta and save it in self.rnn except: raise end_time = time.clock() # [W, WO, Wcat, Wv, Wo] = self.rnn # theta = np.concatenate((W.flatten(), WO.flatten(), Wcat.flatten(), Wv.flatten(), Wo.flatten())) # current local optimal theta #push local theta to master self.push('toMaster_' + str(self.index), (self.rnn, None)) #push the theta value #pushing a tuple object #playdoh bug if self.debug: print "Node:", self.index, " Execution time for ", self.nFetch, " minibatches: ", (end_time - start_time)/60, 'minutes' print "Node:", self.index, " Pushed local parameters to Master.." def async_eval(self, itr, testSet, result_file=None): if testSet != None: # if len(testSet.allSNum) > 100: # rnnData_test_mini = RNNDataCorpus() # testSet.copy_into_minibatch(rnnData_test_mini, range(100)) # else: rnnData_test_mini = testSet self.rnn.evaluate( self.params, rnnData_test_mini) def async_save(self, itr): timeStr = strftime("%Y-%m-%d_%H-%M-%S", gmtime()) # for generating names modelFileName = config.saved_params_file+"_"+str(itr)+'_'+timeStr #save weights with open(modelFileName, 'wb') as wf: cPickle.dump(self.rnn, wf, protocol=-1) # [W, WO, Wcat, Wv, Wo] = self.rnn # save_dict = {'Wv':Wv, 'Wo':Wo, 'Wcat':Wcat, 'W':W, 'WO':WO} # sio.savemat(modelFileName, mdict=save_dict) print 'Master## saved trained model to ', modelFileName def startMasterNode(self): '''Parameter server, receives gradients from workers, updates parameters of model, sends updated parameters to workers ''' nDocsTest = 150 printStatistics = True action = ["continue", "finish"] evalFreq = 10 saveFreq = 5 self.initMVRNN() #also initializes training and test sets nparts = self.rnnData_dev.ndoc()/nDocsTest if self.rnnData_dev.ndoc() > nDocsTest else 1 testSeq = range(self.rnnData_dev.ndoc()) #start optimization by pushing intial params for iNode in xrange(1, self.totalNodes): if self.debug: print "Master## Pushing parameters to " + str(iNode) + " theta[0]= " + str(self.rnn.W[0,0]) self.push('fromMaster_' + str(iNode),(action[0],self.rnn)) # send updated params to workers #start iterations and evaluate after completion of set of iterations for itr in range(self.nIter): # workerW = np.zeros(self.rnn.W.shape); workerWO = np.zeros(self.rnn.WO.shape); workerWcat = np.zeros(self.rnn.Wcat.shape); # workerWv = np.zeros(self.rnn.Wv.shape); workerWo = np.zeros(self.rnn.Wo.shape) workerParams = [] for p in self.rnn.getParamsList(): workerParams.append(np.zeros(p.shape)) self.rnd.shuffle(testSeq) (testSetPart, _) = self.rnnData_dev.getSubset(idocNext=0, ipart=0, nparts=nparts, sequenceIndex=testSeq) if self.debug: print "Master## Iteration: ",itr ,"Waiting to receive params from workers..." for iNode in xrange(1, self.totalNodes): #get gradients from all workers receivedRNN, _ = self.pop('toMaster_' + str(iNode)) # poll the workers for result # print str(receivedRNN) receivedParams = receivedRNN.getParamsList() for workerParam, receivedParam in zip(workerParams, receivedParams): workerParam += receivedParam # workerW = workerW + rW # workerWO = workerWO + rWO # workerWcat = workerWcat + rWcat # workerWv = workerWv + rWv # workerWo = workerWo + rWo if self.debug: print "Master## Iteration: ",itr ,"Received theta from: " + str(iNode) + " theta[0]="+str(receivedParams[0][0,0]) #average all the received parameters for i in range(len(workerParams)): workerParams[i] /= (self.totalNodes-1) # workerW = workerW /(self.totalNodes - 1) # workerWO = workerWO /(self.totalNodes - 1) # workerWcat = workerWcat /(self.totalNodes - 1) # workerWv = workerWv /(self.totalNodes - 1) # workerWo = workerWo /(self.totalNodes - 1) #one is master rest are workers #copy into master's rnn # workerRnn = [workerW, workerWO, workerWcat, workerWv, workerWo] self.copy_into_rnn(workerParams) if(self.debug): print "Master## Average Params : ", str(workerParams[0][0,0]) for iNode in xrange(1, self.totalNodes): self.push('fromMaster_' + str(iNode),(action[0],self.rnn)) # send updated params to workers if self.debug: print "Master## Sending new parameters to Worker ", iNode, "Param[0]: ", str(self.rnn.W[0,0]) if printStatistics and ((itr % evalFreq) == 0) : p = mp.Process(target=self.async_eval, args=(itr,self.rnnData_dev)) p.start() if((itr % saveFreq == 0)): # save mlp after every 5 iterations of nFetch cycles. p = mp.Process(target=self.async_save, args=(itr,)) p.start() #finish optimization and close workers for iNode in xrange(1, self.totalNodes): self.push('fromMaster_' + str(iNode),(action[1],None)) # send updated params to workers def get_result(self): if self.index == 0: return self.rnn def copy_into_rnn(self, workerRnnParams): for dst, src in zip(self.rnn.getParamsList(), workerRnnParams): dst[:,:] = src def initMVRNN(self): print "Node: ",self.index," Loading training and dev sets.." self.rnnData_train = RNNDataCorpus() self.rnnData_train.load_data_srl(config.train_data_srl, nExamples=self.nuse) self.rnnData_dev = RNNDataCorpus() self.rnnData_dev.load_data_srl(config.dev_data_srl, nExamples=self.nuse) modelfilename = config.saved_params_file+'SGD_SRLiter305' print "Node: ", self.index," Loading model: ", modelfilename # mats = sio.loadmat(config.saved_params_file+'iter120.mat') # Wv = mats.get('Wv') #L, as in paper # W = mats.get('W') #W, as in paper # WO = mats.get('WO') #Wm, as in paper # Wo = mats.get('Wo') # Wcat = mats.get('Wcat') # n = Wv.shape[0] # r = (Wo.shape[0] - n)/(2*n) with open(modelfilename, 'r') as loadfile: self.rnn = cPickle.load(loadfile)#MVRNN(W, WO, Wcat, Wv, Wo) n = self.rnn.Wv.shape[0] r = (self.rnn.Wo.shape[0] - n)/(2*n) print "Node: ",self.index, "initializing params.." self.params = Params(data=self.rnnData_train, wordSize=n, rankWo=r) # #to be removed # Wcat = 0.005*np.random.randn(self.params.categories, self.params.fanIn) # self.rnn.Wcat = Wcat if(self.index == 0): print "Master## Total trees in training set: ", self.rnnData_train.ndoc() print "Master## nFetch: ", self.nFetch if(self.index!=0): self.rnnData_train = self.get_subset(self.rnnData_train, self.index) self.rnnData_dev = None # workers don't need test set and trainTest, so to free memory release them [_, _, self.all_train_idx] = getRelevantWords(self.rnnData_train, self.rnn.Wv,self.rnn.Wo,self.params) #sets nWords_reduced, returns new arrays #set this to none, as unrolling of theta will take all words into account. # self.theta = np.concatenate((W.flatten(), WO.flatten(), Wcat.flatten(), Wv_trainTest.flatten(), Wo_trainTest.flatten())) def get_subset(self, rnnData, workerIndex): ''' data : complete train or test set for which a subset of docs is to be calculated for this worker ''' rnnData_mini = RNNDataCorpus() dataSize = rnnData.ndoc() sizePerNode = dataSize/(self.totalNodes-1) #exclude master node startPos = (workerIndex-1) * sizePerNode if(workerIndex != self.totalNodes-1): #last node on a machine gets all the remaining data endPos = startPos+sizePerNode else: endPos = dataSize rnnData.copy_into_minibatch(rnnData_mini, range(startPos, endPos)) return rnnData_mini