def handler(event, context): start_time = time.time() bucket = event['data_bucket'] worker_index = event['rank'] num_workers = event['num_workers'] key = 'training_{}.pt'.format(worker_index) print('data_bucket = {}\n worker_index:{}\n num_worker:{}\n key:{}'.format(bucket, worker_index, num_workers, key)) # Set thrift connection # Make socket transport = TSocket.TSocket(constants.HOST, constants.PORT) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}" .format(constants.HOST, constants.PORT)) # read file from s3 readS3_start = time.time() s3.Bucket(bucket).download_file(key, os.path.join(local_dir, training_file)) s3.Bucket(bucket).download_file(test_file, os.path.join(local_dir, test_file)) print("read data cost {} s".format(time.time() - readS3_start)) # preprocess dataset preprocess_start = time.time() trainset = torch.load(os.path.join(local_dir, training_file)) testset = torch.load(os.path.join(local_dir, test_file)) trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True) testloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') print("preprocess data cost {} s".format(time.time() - preprocess_start)) device = 'cpu' torch.manual_seed(1234) #Model print('==> Building model..') # net = VGG('VGG19') # net = ResNet18() # net = ResNet50() # net = PreActResNet18() # net = GoogLeNet() # net = DenseNet121() # net = ResNeXt29_2x64d() net = MobileNet() # net = MobileNetV2() # net = DPN92() # net = ShuffleNetG2() # net = SENet18() # net = ShuffleNetV2(1) # net = EfficientNetB0() net = net.to(device) optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9, weight_decay=5e-4) # Loss and Optimizer # Softmax is internally computed. # Set parameters to be updated. # criterion = torch.nn.CrossEntropyLoss() # optimizer = torch.optim.SGD(model.parameters(), lr=LEARNING_RATE) # register model model_name = "dnn" weight = [param.data.numpy() for param in net.parameters()] weight_shape = [layer.shape for layer in weight] weight_size = [layer.size for layer in weight] weight_length = sum(weight_size) ps_client.register_model(t_client, worker_index, model_name, weight_length, num_workers) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format(model_name, weight_length)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(NUM_EPOCHS): epoch_start = time.time() net.train() num_batch = 0 train_acc = Accuracy() train_loss = Average() for batch_idx, (inputs, targets) in enumerate(trainloader): # print("------worker {} epoch {} batch {}------".format(worker_index, epoch+1, batch_idx+1)) batch_start = time.time() # pull latest model pull_start = time.time() ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) latest_model = np.asarray(latest_model,dtype=np.float32) pull_time = time.time() - pull_start # update the model offset = 0 for layer_index, param in enumerate(net.parameters()): layer_value = latest_model[offset : offset + weight_size[layer_index]].reshape(weight_shape[layer_index]) param.data = torch.from_numpy(layer_value) offset += weight_size[layer_index] # Forward + Backward + Optimize inputs, targets = inputs.to(device), targets.to(device) outputs = net(inputs) loss = F.cross_entropy(outputs, targets) optimizer.zero_grad() loss.backward() train_acc.update(outputs, targets) train_loss.update(loss.item(), inputs.size(0)) # flatten and concat gradients of weight and bias for index, param in enumerate(net.parameters()): if index == 0: flattened_grad = param.grad.data.numpy().flatten() else: flattened_grad = np.concatenate((flattened_grad, param.grad.data.numpy().flatten())) flattened_grad = flattened_grad * -1 # push gradient to PS push_start = time.time() ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_grad(t_client, model_name, flattened_grad, learning_rate, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter+1, worker_index) # sync all workers push_time = time.time() - push_start iter_counter += 1 num_batch += 1 step_time = time.time() - batch_start print("Epoch:[{}/{}], Step:[{}/{}];\n Training Loss:{}, Training accuracy:{};\n Step Time:{}, Calculation Time:{}, Communication Time:{}".format( epoch, NUM_EPOCHS, num_batch, len(trainloader), train_loss, train_acc, step_time, step_time - (pull_time + push_time), pull_time + push_time)) # Test the Model net.eval() test_loss = Average() test_acc = Accuracy() with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.to(device), targets.to(device) outputs = net(inputs) loss = F.cross_entropy(outputs, targets) test_loss.update(loss.item(), inputs.size(0)) test_acc.update(outputs, targets) # correct = 0 # total = 0 # test_loss = 0 # for items, labels in validation_loader: # items = Variable(items.view(-1, NUM_FEATURES)) # labels = Variable(labels) # outputs = model(items) # test_loss += criterion(outputs, labels).data # _, predicted = torch.max(outputs.data, 1) # total += labels.size(0) # correct += (predicted == labels).sum() print('Time = %.4f, accuracy of the model on test set: %f, loss = %f' % (time.time() - train_start, test_acc, test_loss)) end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context): start_time = time.time() bucket = event['bucket_name'] worker_index = event['rank'] num_workers = event['num_workers'] key = event['file'] host = event['host'] port = event['port'] print('bucket = {}'.format(bucket)) print('number of workers = {}'.format(num_workers)) print('worker index = {}'.format(worker_index)) print("file = {}".format(key)) print("host = {}".format(host)) print("port = {}".format(port)) # Set thrift connection # Make socket transport = TSocket.TSocket(host, port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( host, port)) # read file from s3 file = get_object(bucket, key).read().decode('utf-8').split("\n") print("read data cost {} s".format(time.time() - start_time)) # parse dataset parse_start = time.time() dataset = DenseDatasetWithLines(file, NUM_FEATURES) print("parse data cost {} s".format(time.time() - parse_start)) # preprocess dataset preprocess_start = time.time() dataset_size = len(dataset) indices = list( range(dataset_size)) # indices for training and validation splits: split = int(np.floor(VALIDATION_RATIO * dataset_size)) if SHUFFLE_DATASET: np.random.seed(RANDOM_SEED) np.random.shuffle(indices) train_indices, val_indices = indices[split:], indices[:split] # Creating PT data samplers and loaders: train_sampler = SubsetRandomSampler(train_indices) valid_sampler = SubsetRandomSampler(val_indices) train_loader = torch.utils.data.DataLoader(dataset, batch_size=BATCH_SIZE, sampler=train_sampler) validation_loader = torch.utils.data.DataLoader(dataset, batch_size=BATCH_SIZE, sampler=valid_sampler) print("preprocess data cost {} s".format(time.time() - preprocess_start)) model = LogisticRegression(NUM_FEATURES, NUM_CLASSES) # Loss and Optimizer # Softmax is internally computed. # Set parameters to be updated. criterion = torch.nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=LEARNING_RATE) # register model model_name = "w.b" weight_shape = model.linear.weight.data.numpy().shape weight_length = weight_shape[0] * weight_shape[1] bias_shape = model.linear.bias.data.numpy().shape bias_length = bias_shape[0] model_length = weight_length + bias_length ps_client.register_model(t_client, worker_index, model_name, model_length, num_workers) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(NUM_EPOCHS): epoch_start = time.time() for batch_index, (items, labels) in enumerate(train_loader): print("------worker {} epoch {} batch {}------".format( worker_index, epoch, batch_index)) batch_start = time.time() # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) model.linear.weight = Parameter( torch.from_numpy( np.asarray(latest_model[:weight_length], dtype=np.double).reshape(weight_shape))) model.linear.bias = Parameter( torch.from_numpy( np.asarray(latest_model[weight_length:], dtype=np.double).reshape(bias_shape[0]))) items = Variable(items.view(-1, NUM_FEATURES)) labels = Variable(labels) # Forward + Backward + Optimize optimizer.zero_grad() outputs = model(items.double()) loss = criterion(outputs, labels) loss.backward() # flatten and concat gradients of weight and bias w_b_grad = np.concatenate( (model.linear.weight.grad.data.numpy().flatten(), model.linear.bias.grad.data.numpy().flatten())) cal_time = time.time() - batch_start # push gradient to PS sync_start = time.time() ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_grad(t_client, model_name, w_b_grad, LEARNING_RATE, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index) # sync all workers sync_time = time.time() - sync_start print( 'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, NUM_EPOCHS, batch_index + 1, len(train_indices) / BATCH_SIZE, time.time() - train_start, loss.data, time.time() - epoch_start, time.time() - batch_start, cal_time, sync_time)) iter_counter += 1 # Test the Model correct = 0 total = 0 test_loss = 0 for items, labels in validation_loader: items = Variable(items.view(-1, NUM_FEATURES)) labels = Variable(labels) outputs = model(items) test_loss += criterion(outputs, labels).data _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum() print( 'Time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f' % (time.time() - train_start, len(val_indices), 100 * correct / total, test_loss)) end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context): # dataset data_bucket = event['data_bucket'] file = event['file'] dataset_type = event["dataset_type"] assert dataset_type == "sparse_libsvm" n_features = event['n_features'] # ps setting host = event['host'] port = event['port'] # hyper-parameter n_clusters = event['n_clusters'] n_epochs = event["n_epochs"] threshold = event["threshold"] sync_mode = event["sync_mode"] n_workers = event["n_workers"] worker_index = event['worker_index'] assert sync_mode.lower() == Synchronization.Reduce print('data bucket = {}'.format(data_bucket)) print("file = {}".format(file)) print('number of workers = {}'.format(n_workers)) print('worker index = {}'.format(worker_index)) print('num clusters = {}'.format(n_clusters)) print('host = {}'.format(host)) print('port = {}'.format(port)) # Set thrift connection # Make socket transport = TSocket.TSocket(host, port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format(host, port)) # Reading data from S3 read_start = time.time() storage = S3Storage() lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n") print("read data cost {} s".format(time.time() - read_start)) parse_start = time.time() dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type) train_set = dataset.ins_list np_dtype = train_set[0].to_dense().numpy().dtype centroid_shape = (n_clusters, n_features) print("parse data cost {} s".format(time.time() - parse_start)) print("dataset type: {}, data type: {}, centroids shape: {}" .format(dataset_type, np_dtype, centroid_shape)) # register model model_name = Prefix.KMeans_Cent model_length = centroid_shape[0] * centroid_shape[1] + 1 ps_client.register_model(t_client, worker_index, model_name, model_length, n_workers) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format(model_name, model_length)) init_centroids_start = time.time() ps_client.can_pull(t_client, model_name, 0, worker_index) ps_model = ps_client.pull_model(t_client, model_name, 0, worker_index) if worker_index == 0: centroids_np = sparse_centroid_to_numpy(train_set[0:n_clusters], n_clusters) ps_client.can_push(t_client, model_name, 0, worker_index) ps_client.push_grad(t_client, model_name, np.append(centroids_np.flatten(), 1000.).astype(np.double) - np.asarray(ps_model).astype(np.double), 1., 0, worker_index) else: centroids_np = np.zeros(centroid_shape) ps_client.can_push(t_client, model_name, 0, worker_index) ps_client.push_grad(t_client, model_name, np.append(centroids_np.flatten(), 0).astype(np.double), 0, 0, worker_index) ps_client.can_pull(t_client, model_name, 1, worker_index) ps_model = ps_client.pull_model(t_client, model_name, 1, worker_index) cur_centroids = np.array(ps_model[0:-1]).astype(np.float32).reshape(centroid_shape) cur_error = float(ps_model[-1]) print("initial centroids cost {} s".format(time.time() - init_centroids_start)) model = cluster_models.get_model(train_set, torch.from_numpy(cur_centroids), dataset_type, n_features, n_clusters) train_start = time.time() for epoch in range(1, n_epochs + 1): epoch_start = time.time() # local computation model.find_nearest_cluster() local_cent = model.get_centroids("numpy").reshape(-1) local_cent_error = np.concatenate((local_cent.astype(np.double).flatten(), np.array([model.error], dtype=np.double))) epoch_cal_time = time.time() - epoch_start # push updates epoch_comm_start = time.time() last_cent_error = np.concatenate((cur_centroids.astype(np.double).flatten(), np.array([cur_error], dtype=np.double))) ps_model_inc = local_cent_error - last_cent_error ps_client.can_push(t_client, model_name, epoch, worker_index) ps_client.push_grad(t_client, model_name, ps_model_inc, 1. / n_workers, epoch, worker_index) # pull new model ps_client.can_pull(t_client, model_name, epoch + 1, worker_index) # sync all workers ps_model = ps_client.pull_model(t_client, model_name, epoch + 1, worker_index) model.centroids = [torch.from_numpy(c).reshape(1, n_features).to_sparse() for c in np.array(ps_model[0:-1]).astype(np.float32).reshape(centroid_shape)] model.error = float(ps_model[-1]) cur_centroids = model.get_centroids("numpy") cur_error = model.error epoch_comm_time = time.time() - epoch_comm_start print("Epoch[{}] Worker[{}], error = {}, cost {} s, cal cost {} s, sync cost {} s" .format(epoch, worker_index, model.error, time.time() - epoch_start, epoch_cal_time, epoch_comm_time)) if model.error < threshold: break print("Worker[{}] finishes training: Error = {}, cost {} s" .format(worker_index, model.error, time.time() - train_start)) return
def handler(event, context): start_time = time.time() worker_index = event['rank'] num_workers = event['num_workers'] host = event['host'] port = event['port'] size = event['size'] print('number of workers = {}'.format(num_workers)) print('worker index = {}'.format(worker_index)) print("host = {}".format(host)) print("port = {}".format(port)) print("size = {}".format(size)) # Set thrift connection # Make socket transport = TSocket.TSocket(host, port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( host, port)) # register model ps_client.register_model(t_client, worker_index, MODEL_NAME, size, num_workers) ps_client.exist_model(t_client, MODEL_NAME) print("register and check model >>> name = {}, length = {}".format( MODEL_NAME, size)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(NUM_EPOCHS): epoch_start = time.time() for batch_index in range(NUM_BATCHES): print("------worker {} epoch {} batch {}------".format( worker_index, epoch, batch_index)) batch_start = time.time() loss = 0.0 # pull latest model ps_client.can_pull(t_client, MODEL_NAME, iter_counter, worker_index) pull_start = time.time() latest_model = ps_client.pull_model(t_client, MODEL_NAME, iter_counter, worker_index) pull_time = time.time() - pull_start w_b_grad = np.random.rand(1, size).astype(np.double).flatten() # push gradient to PS ps_client.can_push(t_client, MODEL_NAME, iter_counter, worker_index) push_start = time.time() ps_client.push_grad(t_client, MODEL_NAME, w_b_grad, LEARNING_RATE, iter_counter, worker_index) push_time = time.time() - push_start ps_client.can_pull(t_client, MODEL_NAME, iter_counter + 1, worker_index) # sync all workers print( 'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: pull model cost %.4f s, push update cost %.4f s' % (epoch + 1, NUM_EPOCHS, batch_index, NUM_BATCHES, time.time() - train_start, loss, time.time() - epoch_start, time.time() - batch_start, pull_time, push_time)) iter_counter += 1 end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def main(): parser = argparse.ArgumentParser() parser.add_argument('--rank', type=int, default=0) parser.add_argument('--host', type=str, default=constants.HOST) parser.add_argument('--port', type=int, default=constants.PORT) args = parser.parse_args() print(args) worker_index = args.rank # Make socket transport = TSocket.TSocket(args.host, args.port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder client = ParameterServer.Client(protocol) # Connect! transport.open() client.ping() print('ping()') # register model if worker_index == 0: client.register_model(MODEL_ID, MODEL_LENGTH, N_WORKER) is_model_exist = False while is_model_exist is not True: is_model_exist = client.exist_model(MODEL_ID) # pull latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 0, worker_index) weight = client.pull_model(MODEL_ID, 0, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # push gradient try: can_push = False while can_push is not True: can_push = client.can_push(MODEL_ID, 0, worker_index) grad = Grad() grad.id = MODEL_ID grad.learning_rate = 0.01 grad.length = 10 grad.data = [i for i in range(MODEL_LENGTH)] grad.n_iter = 0 grad.worker_id = worker_index client.push_grad(grad) except InvalidOperation as e: print('InvalidOperation: %r' % e) # get latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 1, worker_index) weight = client.pull_model(MODEL_ID, 1, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # push update try: can_push = False while can_push is not True: can_push = client.can_push(MODEL_ID, 1, worker_index) update = Update() update.id = MODEL_ID update.length = MODEL_LENGTH update.data = [i for i in range(MODEL_LENGTH)] update.n_iter = 1 update.worker_id = worker_index client.push_update(update) except InvalidOperation as e: print('InvalidOperation: %r' % e) # get latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 2, worker_index) weight = client.pull_model(MODEL_ID, 2, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # Close! transport.close()
def handler(event, context): startTs = time.time() num_features = event['num_features'] learning_rate = event["learning_rate"] batch_size = event["batch_size"] num_epochs = event["num_epochs"] validation_ratio = event["validation_ratio"] # Reading data from S3 bucket_name = event['bucket_name'] key = urllib.parse.unquote_plus(event['key'], encoding='utf-8') print(f"Reading training data from bucket = {bucket_name}, key = {key}") key_splits = key.split("_") worker_index = int(key_splits[0]) num_worker = int(key_splits[1]) # read file from s3 file = get_object(bucket_name, key).read().decode('utf-8').split("\n") print("read data cost {} s".format(time.time() - startTs)) parse_start = time.time() dataset = SparseDatasetWithLines(file, num_features) print("parse data cost {} s".format(time.time() - parse_start)) preprocess_start = time.time() dataset_size = len(dataset) indices = list(range(dataset_size)) split = int(np.floor(validation_ratio * dataset_size)) np.random.seed(42) np.random.shuffle(indices) train_indices, val_indices = indices[split:], indices[:split] train_set = [dataset[i] for i in train_indices] val_set = [dataset[i] for i in val_indices] print("preprocess data cost {} s".format(time.time() - preprocess_start)) # Set thrift connection # Make socket transport = TSocket.TSocket(constants.HOST, constants.PORT) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( constants.HOST, constants.PORT)) svm = SparseSVM(train_set, val_set, num_features, num_epochs, learning_rate, batch_size) # register model model_name = "w.b" model_length = num_features ps_client.register_model(t_client, worker_index, model_name, model_length, num_worker) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 # Training the Model for epoch in range(num_epochs): epoch_start = time.time() num_batches = math.floor(len(train_set) / batch_size) print(f"worker {worker_index} epoch {epoch}") for batch_idx in range(num_batches): batch_start = time.time() # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) svm.weights = torch.from_numpy(latest_model).reshape( num_features, 1) batch_ins, batch_label = svm.next_batch(batch_idx) acc = svm.one_epoch(batch_idx, epoch) compute_end = time.time() sync_start = time.time() w_update = svm.weights - latest_model ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_update(t_client, model_name, w_update, learning_rate, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index) # sync all workers sync_time = time.time() - sync_start print( 'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, train acc: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, NUM_EPOCHS, batch_idx + 1, len(train_indices) / batch_size, time.time() - train_start, acc, time.time() - epoch_start, time.time() - batch_start, compute_end - batch_start, sync_time)) iter_counter += 1 val_acc = svm.evaluate() print("Epoch takes {}s, validation accuracy: {}".format( time.time() - epoch_start, val_acc))
def handler(event, context): start_time = time.time() # dataset setting train_file = event['train_file'] test_file = event['test_file'] data_bucket = event['data_bucket'] n_features = event['n_features'] n_classes = event['n_classes'] n_workers = event['n_workers'] worker_index = event['worker_index'] cp_bucket = event['cp_bucket'] # ps setting host = event['host'] port = event['port'] # training setting model_name = event['model'] optim = event['optim'] sync_mode = event['sync_mode'] assert model_name.lower() in MLModel.Deep_Models assert optim.lower() in Optimization.Grad_Avg assert sync_mode.lower() in Synchronization.Reduce # hyper-parameter learning_rate = event['lr'] batch_size = event['batch_size'] n_epochs = event['n_epochs'] start_epoch = event['start_epoch'] run_epochs = event['run_epochs'] function_name = event['function_name'] print('data bucket = {}'.format(data_bucket)) print("train file = {}".format(train_file)) print("test file = {}".format(test_file)) print('number of workers = {}'.format(n_workers)) print('worker index = {}'.format(worker_index)) print('model = {}'.format(model_name)) print('optimization = {}'.format(optim)) print('sync mode = {}'.format(sync_mode)) print('start epoch = {}'.format(start_epoch)) print('run epochs = {}'.format(run_epochs)) print('host = {}'.format(host)) print('port = {}'.format(port)) print("Run function {}, round: {}/{}, epoch: {}/{} to {}/{}".format( function_name, int(start_epoch / run_epochs) + 1, math.ceil(n_epochs / run_epochs), start_epoch + 1, n_epochs, start_epoch + run_epochs, n_epochs)) # download file from s3 storage = S3Storage() local_dir = "/tmp" read_start = time.time() storage.download(data_bucket, train_file, os.path.join(local_dir, train_file)) storage.download(data_bucket, test_file, os.path.join(local_dir, test_file)) print("download file from s3 cost {} s".format(time.time() - read_start)) train_set = torch.load(os.path.join(local_dir, train_file)) test_set = torch.load(os.path.join(local_dir, test_file)) train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True) n_train_batch = len(train_loader) test_loader = torch.utils.data.DataLoader(test_set, batch_size=100, shuffle=False) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') print("read data cost {} s".format(time.time() - read_start)) random_seed = 100 torch.manual_seed(random_seed) device = 'cpu' model = deep_models.get_models(model_name).to(device) optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate) # load checkpoint model if it is not the first round if start_epoch != 0: checked_file = 'checkpoint_{}.pt'.format(start_epoch - 1) storage.download(cp_bucket, checked_file, os.path.join(local_dir, checked_file)) checkpoint_model = torch.load(os.path.join(local_dir, checked_file)) model.load_state_dict(checkpoint_model['model_state_dict']) optimizer.load_state_dict(checkpoint_model['optimizer_state_dict']) print("load checkpoint model at epoch {}".format(start_epoch - 1)) # Set thrift connection # Make socket transport = TSocket.TSocket(host, port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( host, port)) # register model parameter_shape = [] parameter_length = [] model_length = 0 for param in model.parameters(): tmp_shape = 1 parameter_shape.append(param.data.numpy().shape) for w in param.data.numpy().shape: tmp_shape *= w parameter_length.append(tmp_shape) model_length += tmp_shape ps_client.register_model(t_client, worker_index, model_name, model_length, n_workers) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(start_epoch, min(start_epoch + run_epochs, n_epochs)): model.train() epoch_start = time.time() train_acc = Accuracy() train_loss = Average() for batch_idx, (inputs, targets) in enumerate(train_loader): batch_start = time.time() batch_cal_time = 0 batch_comm_time = 0 # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) pos = 0 for layer_index, param in enumerate(model.parameters()): param.data = Variable( torch.from_numpy( np.asarray(latest_model[pos:pos + parameter_length[layer_index]], dtype=np.float32).reshape( parameter_shape[layer_index]))) pos += parameter_length[layer_index] batch_comm_time += time.time() - batch_start batch_cal_start = time.time() outputs = model(inputs) loss = F.cross_entropy(outputs, targets) optimizer.zero_grad() loss.backward() # flatten and concat gradients of weight and bias param_grad = np.zeros((1)) for param in model.parameters(): # print("shape of layer = {}".format(param.data.numpy().flatten().shape)) param_grad = np.concatenate( (param_grad, param.data.numpy().flatten())) param_grad = np.delete(param_grad, 0) #print("model_length = {}".format(param_grad.shape)) batch_cal_time += time.time() - batch_cal_start # push gradient to PS batch_push_start = time.time() ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_grad(t_client, model_name, param_grad, -1. * learning_rate / n_workers, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index) # sync all workers batch_comm_time += time.time() - batch_push_start train_acc.update(outputs, targets) train_loss.update(loss.item(), inputs.size(0)) optimizer.step() iter_counter += 1 if batch_idx % 10 == 0: print( 'Epoch: [%d/%d], Batch: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, n_epochs, batch_idx + 1, n_train_batch, time.time() - train_start, loss.item(), time.time() - epoch_start, time.time() - batch_start, batch_cal_time, batch_comm_time)) test_loss, test_acc = test(epoch, model, test_loader) print( 'Epoch: {}/{},'.format(epoch + 1, n_epochs), 'train loss: {},'.format(train_loss), 'train acc: {},'.format(train_acc), 'test loss: {},'.format(test_loss), 'test acc: {}.'.format(test_acc), ) # training is not finished yet, invoke next round if epoch < n_epochs - 1: checkpoint_model = { 'epoch': epoch, 'model_state_dict': model.state_dict(), 'optimizer_state_dict': optimizer.state_dict(), 'loss': train_loss.average } checked_file = 'checkpoint_{}.pt'.format(epoch) if worker_index == 0: torch.save(checkpoint_model, os.path.join(local_dir, checked_file)) storage.upload(cp_bucket, checked_file, os.path.join(local_dir, checked_file)) print("checkpoint model at epoch {} saved!".format(epoch)) print( "Invoking the next round of functions. round: {}/{}, start epoch: {}, run epoch: {}" .format( int((epoch + 1) / run_epochs) + 1, math.ceil(n_epochs / run_epochs), epoch + 1, run_epochs)) lambda_client = boto3.client('lambda') payload = { 'train_file': event['train_file'], 'test_file': event['test_file'], 'data_bucket': event['data_bucket'], 'n_features': event['n_features'], 'n_classes': event['n_classes'], 'n_workers': event['n_workers'], 'worker_index': event['worker_index'], 'cp_bucket': event['cp_bucket'], 'host': event['host'], 'port': event['port'], 'model': event['model'], 'optim': event['optim'], 'sync_mode': event['sync_mode'], 'lr': event['lr'], 'batch_size': event['batch_size'], 'n_epochs': event['n_epochs'], 'start_epoch': epoch + 1, 'run_epochs': event['run_epochs'], 'function_name': event['function_name'] } lambda_client.invoke(FunctionName=function_name, InvocationType='Event', Payload=json.dumps(payload)) end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context): start_time = time.time() bucket = event['data_bucket'] worker_index = event['rank'] num_worker = event['num_workers'] key = event['key'] print('bucket = {}'.format(bucket)) print('number of workers = {}'.format(num_worker)) print('worker index = {}'.format(worker_index)) # Set thrift connection # Make socket transport = TSocket.TSocket(constants.HOST, constants.PORT) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( constants.HOST, constants.PORT)) #bucket = "cifar10dataset" print('data_bucket = {}\n worker_index:{}\n num_worker:{}\n key:{}'.format( bucket, worker_index, num_worker, key)) # read file from s3 readS3_start = time.time() train_path = download_file(bucket, key) trainset = torch.load(train_path) test_path = download_file(bucket, test_file) testset = torch.load(test_path) print("read data cost {} s".format(time.time() - readS3_start)) preprocess_start = time.time() batch_size = 200 train_loader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True) test_loader = torch.utils.data.DataLoader(testset, batch_size=batch_size, shuffle=False) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') device = 'cpu' print("preprocess data cost {} s".format(time.time() - preprocess_start)) model = MobileNet() model = model.to(device) # Loss and Optimizer # Softmax is internally computed. # Set parameters to be updated. criterion = torch.nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, weight_decay=5e-4) # register model model_name = "mobilenet" parameter_shape = [] parameter_length = [] model_length = 0 for param in model.parameters(): tmp_shape = 1 parameter_shape.append(param.data.numpy().shape) for w in param.data.numpy().shape: tmp_shape *= w parameter_length.append(tmp_shape) model_length += tmp_shape ps_client.register_model(t_client, worker_index, model_name, model_length, num_worker) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(num_epochs): epoch_start = time.time() model.train() for batch_index, (inputs, targets) in enumerate(train_loader): print("------worker {} epoch {} batch {}------".format( worker_index, epoch, batch_index)) batch_start = time.time() # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) pos = 0 for layer_index, param in enumerate(model.parameters()): param.data = Variable( torch.from_numpy( np.asarray(latest_model[pos:pos + parameter_length[layer_index]], dtype=np.float32).reshape( parameter_shape[layer_index]))) pos += parameter_length[layer_index] # Forward + Backward + Optimize inputs, targets = inputs.to(device), targets.to(device) outputs = model(inputs) loss = criterion(outputs, targets) optimizer.zero_grad() loss.backward() # flatten and concat gradients of weight and bias param_grad = np.zeros((1)) for param in model.parameters(): #print("shape of layer = {}".format(param.data.numpy().flatten().shape)) param_grad = np.concatenate( (param_grad, param.data.numpy().flatten())) param_grad = np.delete(param_grad, 0) print("model_length = {}".format(param_grad.shape)) # push gradient to PS sync_start = time.time() print( ps_client.can_push(t_client, model_name, iter_counter, worker_index)) print( ps_client.push_grad(t_client, model_name, param_grad, learning_rate, iter_counter, worker_index)) print( ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index)) # sync all workers sync_time = time.time() - sync_start print( 'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, num_epochs, batch_index + 1, len(train_indices) / batch_size, time.time() - train_start, loss.data, time.time() - epoch_start, time.time() - batch_start, cal_time, sync_time)) iter_counter += 1 test(epoch, model, test_loader, criterion, device) optimizer.step()
def main(): start_time = time.time() parser = argparse.ArgumentParser() parser.add_argument('--num-workers', type=int, default=1) parser.add_argument('--rank', type=int, default=0) parser.add_argument('--host', type=str, default=constants.HOST) parser.add_argument('--port', type=int, default=constants.PORT) parser.add_argument('--size', type=int, default=100) args = parser.parse_args() print(args) print("host = {}".format(args.host)) print("port = {}".format(args.port)) # Set thrift connection # Make socket transport = TSocket.TSocket(args.host, args.port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format(args.host, args.port)) # register model ps_client.register_model(t_client, args.rank, MODEL_NAME, args.size, args.num_workers) ps_client.exist_model(t_client, MODEL_NAME) print("register and check model >>> name = {}, length = {}".format(MODEL_NAME, args.size)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(NUM_EPOCHS): epoch_start = time.time() for batch_index in range(NUM_BATCHES): print("------worker {} epoch {} batch {}------" .format(args.rank, epoch, batch_index)) batch_start = time.time() loss = 0.0 # pull latest model ps_client.can_pull(t_client, MODEL_NAME, iter_counter, args.rank) pull_start = time.time() latest_model = ps_client.pull_model(t_client, MODEL_NAME, iter_counter, args.rank) pull_time = time.time() - pull_start cal_start = time.time() w_b_grad = np.random.rand(1, args.size).astype(np.double).flatten() cal_time = time.time() - cal_start # push gradient to PS ps_client.can_push(t_client, MODEL_NAME, iter_counter, args.rank) push_start = time.time() ps_client.push_grad(t_client, MODEL_NAME, w_b_grad, LEARNING_RATE, iter_counter, args.rank) push_time = time.time() - push_start ps_client.can_pull(t_client, MODEL_NAME, iter_counter + 1, args.rank) # sync all workers print('Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s, pull model cost %.4f s, push update cost %.4f s' % (epoch + 1, NUM_EPOCHS, batch_index, NUM_BATCHES, time.time() - train_start, loss, time.time() - epoch_start, time.time() - batch_start, cal_time, pull_time, push_time)) iter_counter += 1 end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context): start_time = time.time() worker_index = event['rank'] num_workers = event['num_workers'] # Make socket transport = TSocket.TSocket(constants.HOST, constants.PORT) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder client = ParameterServer.Client(protocol) # Connect! transport.open() client.ping() print('ping()') # register model if worker_index == 0: client.register_model(MODEL_ID, MODEL_LENGTH, num_workers) is_model_exist = False while is_model_exist is not True: is_model_exist = client.exist_model(MODEL_ID) # pull latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 0, worker_index) weight = client.pull_model(MODEL_ID, 0, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # push gradient try: can_push = False while can_push is not True: can_push = client.can_push(MODEL_ID, 0, worker_index) grad = Grad() grad.id = MODEL_ID grad.learning_rate = 0.01 grad.length = 10 grad.data = [i for i in range(MODEL_LENGTH)] grad.n_iter = 0 grad.worker_id = worker_index client.push_grad(grad) except InvalidOperation as e: print('InvalidOperation: %r' % e) # get latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 1, worker_index) weight = client.pull_model(MODEL_ID, 1, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # push update try: can_push = False while can_push is not True: can_push = client.can_push(MODEL_ID, 1, worker_index) update = Update() update.id = MODEL_ID update.length = MODEL_LENGTH update.data = [i for i in range(MODEL_LENGTH)] update.n_iter = 1 update.worker_id = worker_index client.push_update(update) except InvalidOperation as e: print('InvalidOperation: %r' % e) # get latest model try: can_pull = False while can_pull is not True: can_pull = client.can_pull(MODEL_ID, 2, worker_index) weight = client.pull_model(MODEL_ID, 2, worker_index) weight_data = weight.data print("current weight = {}".format(weight_data)) except InvalidOperation as e: print('InvalidOperation: %r' % e) # Close! transport.close()
def handler(event, context): start_time = time.time() # dataset setting file = event['file'] data_bucket = event['data_bucket'] dataset_type = event['dataset_type'] assert dataset_type == "sparse_libsvm" n_features = event['n_features'] n_classes = event['n_classes'] n_workers = event['n_workers'] worker_index = event['worker_index'] # ps setting host = event['host'] port = event['port'] # training setting model_name = event['model'] optim = event['optim'] sync_mode = event['sync_mode'] assert model_name.lower() in MLModel.Sparse_Linear_Models assert optim.lower() == Optimization.Grad_Avg assert sync_mode.lower() == Synchronization.Reduce # hyper-parameter learning_rate = event['lr'] batch_size = event['batch_size'] n_epochs = event['n_epochs'] valid_ratio = event['valid_ratio'] print('bucket = {}'.format(data_bucket)) print("file = {}".format(file)) print('number of workers = {}'.format(n_workers)) print('worker index = {}'.format(worker_index)) print('model = {}'.format(model_name)) print('host = {}'.format(host)) print('port = {}'.format(port)) # Set thrift connection # Make socket transport = TSocket.TSocket(host, port) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( host, port)) # Read file from s3 read_start = time.time() storage = S3Storage() lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n") print("read data cost {} s".format(time.time() - read_start)) parse_start = time.time() dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type) print("parse data cost {} s".format(time.time() - parse_start)) preprocess_start = time.time() # Creating data indices for training and validation splits: dataset_size = len(dataset) indices = list(range(dataset_size)) split = int(np.floor(valid_ratio * dataset_size)) shuffle_dataset = True random_seed = 100 if shuffle_dataset: np.random.seed(random_seed) np.random.shuffle(indices) train_indices, val_indices = indices[split:], indices[:split] # split train set and test set train_set = [dataset[i] for i in train_indices] n_train_batch = math.floor(len(train_set) / batch_size) val_set = [dataset[i] for i in val_indices] print("preprocess data cost {} s, dataset size = {}".format( time.time() - preprocess_start, dataset_size)) model = linear_models.get_sparse_model(model_name, train_set, val_set, n_features, n_epochs, learning_rate, batch_size) # register model model_name = "w.b" weight_length = n_features bias_length = 1 model_length = weight_length + bias_length ps_client.register_model(t_client, worker_index, model_name, model_length, n_workers) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 for epoch in range(n_epochs): epoch_start = time.time() epoch_cal_time = 0 epoch_comm_time = 0 epoch_loss = 0. for batch_idx in range(n_train_batch): batch_start = time.time() batch_comm_time = 0 # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) model.weight = torch.from_numpy( np.asarray(latest_model[:weight_length]).astype( np.float32).reshape(n_features, 1)) model.bias = float(latest_model[-1]) batch_comm_time += time.time() - batch_start batch_loss, batch_acc = model.one_batch() epoch_loss += batch_loss.average w_b = np.concatenate((model.weight.double().numpy().flatten(), np.array([model.bias]).astype(np.double))) w_b_update = np.subtract(w_b, latest_model) batch_cal_time = time.time() - batch_start # push gradient to PS batch_comm_start = time.time() ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_grad(t_client, model_name, w_b_update, 1.0 / n_workers, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index) # sync all workers batch_comm_time += time.time() - batch_comm_start epoch_cal_time += batch_cal_time epoch_comm_time += batch_comm_time if batch_idx % 10 == 0: print( 'Epoch: [%d/%d], Batch: [%d/%d], Time: %.4f, Loss: %.4f, Accuracy: %.4f,' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, n_epochs, batch_idx + 1, n_train_batch, time.time() - train_start, batch_loss.average, batch_acc.accuracy, time.time() - batch_start, batch_cal_time, batch_comm_time)) iter_counter += 1 # Test the Model test_start = time.time() test_loss, test_acc = model.evaluate() test_time = time.time() - test_start print( "Epoch: [{}/{}] finishes, Batch: [{}/{}], Time: {:.4f}, Loss: {:.4f}, epoch cost {:.4f} s, " "calculation cost = {:.4f} s, synchronization cost {:.4f} s, test cost {:.4f} s, " "accuracy of the model on the {} test samples: {}, loss = {}". format(epoch + 1, n_epochs, batch_idx + 1, n_train_batch, time.time() - train_start, epoch_loss, time.time() - epoch_start, epoch_cal_time, epoch_comm_time, test_time, len(val_set), test_acc.accuracy, test_loss.average)) end_time = time.time() print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context): startTs = time.time() num_features = event['num_features'] learning_rate = event["learning_rate"] batch_size = event["batch_size"] num_epochs = event["num_epochs"] validation_ratio = event["validation_ratio"] # Reading data from S3 bucket_name = event['bucket_name'] key = urllib.parse.unquote_plus(event['key'], encoding='utf-8') print(f"Reading training data from bucket = {bucket_name}, key = {key}") key_splits = key.split("_") worker_index = int(key_splits[0]) num_worker = int(key_splits[1]) # read file from s3 file = get_object(bucket_name, key).read().decode('utf-8').split("\n") print("read data cost {} s".format(time.time() - startTs)) parse_start = time.time() dataset = SparseDatasetWithLines(file, num_features) print("parse data cost {} s".format(time.time() - parse_start)) preprocess_start = time.time() dataset_size = len(dataset) indices = list(range(dataset_size)) split = int(np.floor(validation_ratio * dataset_size)) np.random.seed(42) np.random.shuffle(indices) train_indices, val_indices = indices[split:], indices[:split] train_set = [dataset[i] for i in train_indices] val_set = [dataset[i] for i in val_indices] print("preprocess data cost {} s".format(time.time() - preprocess_start)) # Set thrift connection # Make socket transport = TSocket.TSocket(constants.HOST, constants.PORT) # Buffering is critical. Raw sockets are very slow transport = TTransport.TBufferedTransport(transport) # Wrap in a protocol protocol = TBinaryProtocol.TBinaryProtocol(transport) # Create a client to use the protocol encoder t_client = ParameterServer.Client(protocol) # Connect! transport.open() # test thrift connection ps_client.ping(t_client) print("create and ping thrift server >>> HOST = {}, PORT = {}".format( constants.HOST, constants.PORT)) # register model model_name = "w.b" model_length = num_features + 1 ps_client.register_model(t_client, worker_index, model_name, model_length, num_worker) ps_client.exist_model(t_client, model_name) print("register and check model >>> name = {}, length = {}".format( model_name, model_length)) # Training the Model train_start = time.time() iter_counter = 0 lr = LogisticRegression(train_set, val_set, num_features, num_epochs, learning_rate, batch_size) # Training the Model for epoch in range(num_epochs): epoch_start = time.time() num_batches = math.floor(len(train_set) / batch_size) print(f"worker {worker_index} epoch {epoch}") for batch_idx in range(num_batches): batch_start = time.time() # pull latest model ps_client.can_pull(t_client, model_name, iter_counter, worker_index) latest_model = ps_client.pull_model(t_client, model_name, iter_counter, worker_index) lr.grad = torch.from_numpy(np.asarray(latest_model[:-1])).reshape( num_features, 1) lr.bias = float(latest_model[-1]) compute_start = time.time() batch_ins, batch_label = lr.next_batch(batch_idx) batch_grad = torch.zeros(lr.n_input, 1, requires_grad=False) batch_bias = np.float(0) train_loss = Loss() train_acc = Accuracy() for i in range(len(batch_ins)): z = lr.forward(batch_ins[i]) h = lr.sigmoid(z) loss = lr.loss(h, batch_label[i]) # print("z= {}, h= {}, loss = {}".format(z, h, loss)) train_loss.update(loss, 1) train_acc.update(h, batch_label[i]) g = lr.backward(batch_ins[i], h.item(), batch_label[i]) batch_grad.add_(g) batch_bias += np.sum(h.item() - batch_label[i]) batch_grad = batch_grad.div(len(batch_ins)) batch_bias = batch_bias / len(batch_ins) batch_grad.mul_(-1.0 * learning_rate) lr.grad.add_(batch_grad) lr.bias = lr.bias - batch_bias * learning_rate np_grad = lr.grad.numpy().flatten() w_b_grad = np.append(np_grad, lr.bias) compute_end = time.time() sync_start = time.time() ps_client.can_push(t_client, model_name, iter_counter, worker_index) ps_client.push_grad(t_client, model_name, w_b_grad, learning_rate, iter_counter, worker_index) ps_client.can_pull(t_client, model_name, iter_counter + 1, worker_index) # sync all workers sync_time = time.time() - sync_start print( 'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, ' 'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s' % (epoch + 1, NUM_EPOCHS, batch_idx + 1, len(train_indices) / batch_size, time.time() - train_start, train_loss, time.time() - epoch_start, time.time() - batch_start, compute_end - batch_start, sync_time)) iter_counter += 1 val_loss, val_acc = lr.evaluate() print(f"Validation loss: {val_loss}, validation accuracy: {val_acc}") print(f"Epoch takes {time.time() - epoch_start}s")