package com.nus.dbsytem.falcon.v0

Get desktop application:
View/edit binary Protocol Buffers messages

This message denotes the decision tree builder parameters

Used in: GbdtParams, RandomForestParams

• string tree_type = 1

  type of the tree, 'classification' or 'regression'

• string criterion = 2

  the function to measure the quality of a split 'gini' or 'entropy'

• string split_strategy = 3

  the strategy used to choose a split at each node, 'best' or 'random'

• int32 class_num = 4

  the number of classes in the dataset, if regression, set to 1

• int32 max_depth = 5

  the maximum depth of the tree

• int32 max_bins = 6

  the maximum number of bins to split a feature

• int32 min_samples_split = 7

  the minimum number of samples required to split an internal node

• int32 min_samples_leaf = 8

  the minimum number of samples required to be at a leaf node

• int32 max_leaf_nodes = 9

  the maximum number of leaf nodes

• double min_impurity_decrease = 10

  a node will be split if this split induces a decrease of impurity >= this value

• double min_impurity_split = 11

  threshold for early stopping in tree growth

• double dp_budget = 12

  differential privacy (DP) budget, 0 denotes not use DP

This message is used when transmitting the training labels and testing labels

Used in: DoubleMatrix

• repeated double item = 1

This message is used when transmitting the training data and testing data

• repeated DoubleArray array = 1

Used in: EncodedNumberMatrix, Layer

• repeated FixedPointEncodedNumber encoded_number = 1

Used in: Layer

• repeated EncodedNumberArray encoded_array = 1

Used in: EncryptedStatistics

• repeated FixedPointEncodedNumber encrypted_stat = 1

  encrypted stat for each branch of each class

• int32 client_id = 1

  client id

• int32 node_index = 2

  node index

• int32 local_split_num = 3

  total split num on a client

• int32 classes_num = 4

  classes_num * 2 (if regression, classes_num = 2)

• repeated FixedPointEncodedNumber left_sample_nums_of_splits = 5

  sample_nums of left branch of total splits

• repeated FixedPointEncodedNumber right_sample_nums_of_splits = 6

  sample_nums of right branch of total splits

• repeated EncryptedStatPerSplit encrypted_stats_of_splits = 7

  two-dimensional encrypted statistics

This message denotes the feature selection parameters

• int32 num_samples = 1

  number of samples generated or selected

• string feature_selection = 2

  feature selection method,

• string selected_features_file = 3

  selected features to be saved

This message is used to serialize an EncodedNumber value

Used in: EncodedNumberArray, EncryptedStatPerSplit, EncryptedStatistics, LrModel, Node, NodeUpdateInfo

• string n = 1

  maximum value

• string value = 2

  encoded value

• int32 exponent = 3

  fixed point precision

• int32 type = 4

  value type

ensemble tree model: random forest

• int32 tree_size = 1

  the number of estimator stored

• int32 tree_type = 2

  the model type: classification or regression

• repeated TreeModel trees = 3

  the stored trees

ensemble tree model: gbdt

• int32 tree_size = 1

  the number of total trees stored, n_estimator * class_num

• int32 tree_type = 2

  the model type: classification or regression

• int32 n_estimator = 3

  the number of estimator for gbdt model

• int32 class_num = 4

  the number of classes, 1 for regression

• double learning_rate = 5

  the shrinkage rate during the training

• repeated double dummy_predictors = 6

  the dummy predictors for the trees

• repeated TreeModel trees = 7

  the stored trees

This message denotes the gradient boosting decision tree builder parameters

• int32 n_estimator = 1

  number of estimators (note that the number of total trees in the model does not necessarily equal to the number of estimators for classification)

• string loss = 2

  loss function to be optimized

• double learning_rate = 3

  learning rate shrinks the contribution of each tree

• double subsample = 4

  the fraction of samples to be used for fitting individual base learners default 1.0, reserved here for future usage

• optional DecisionTreeParams dt_param = 5

  decision tree builder params (note that the tree type here may be changed when building the gbdt model as they are all regression trees in gbdt)

This message is used when initializing the training data and testing data, as well as transmitting batch ids in each training iteration

• repeated int32 int_item = 1

Used in: MlpModel

• int32 num_inputs = 1

  the number of inputs for each neuron in this layer in fact, it is the number of neurons (number of outputs) of the previous layer

• int32 num_outputs = 2

  the number of outputs for this layer

• bool fit_bias = 3

  whether the neurons of this layer have bias (default true)

• string activation_func_str = 4

  the activation function string

• optional EncodedNumberMatrix weight_mat = 5

  the weight matrix, encrypted values during training, dimension = (m_num_inputs, m_num_outputs)

• optional EncodedNumberArray bias_vec = 6

  the bias vector, encrypted values during training, dimension = m_num_outputs

This message denotes the LIME interpretability compute predictions parameters

• string original_model_name = 1

  vertical original model name

• string original_model_saved_file = 2

  vertical original model saved file

• string generated_sample_file = 3

  generated samples save file

• string model_type = 4

  type of model task, 'regression' or 'classification'

• int32 class_num = 5

  number of classes in classification, set to 1 if regression

• string computed_prediction_file = 6

  prediction save file

This message denotes the compute sample weights parameters

• int32 explain_instance_idx = 1

  the instance index for explain

• string generated_sample_file = 2

  generated samples save file

• string computed_prediction_file = 3

  prediction save file

• bool is_precompute = 4

  whether it is pre-computed

• int32 num_samples = 5

  number of samples to be generated or selected

• int32 class_num = 6

  number of classes in classification, set to 1 if regression

• string distance_metric = 7

  the metric for computing the distance, only "euclidean"

• string kernel = 8

  kernel, similarity kernel that takes euclidean distances and kernel width as input and outputs weights in (0,1). If not specified, default is exponential kernel

• double kernel_width = 9

  width for the kernel

• string sample_weights_file = 10

  sample weights file to be saved

• string selected_samples_file = 11

  selected samples to be saved if is_precompute = true

• string selected_predictions_file = 12

  selected predictions to be saved if is_precompute = true

This message denotes the LIME feature selection parameters

• string selected_samples_file = 1

  selected samples file

• string selected_predictions_file = 2

  selected predictions file

• string sample_weights_file = 3

  the sample weights file

• int32 num_samples = 4

  number of samples generated or selected

• int32 class_num = 5

  number of classes in classification, set to 1 if regression

• int32 class_id = 6

  the label id to be explained

• string feature_selection = 7

  feature selection method, current options are 'pearson', 'lasso_path',

• string feature_selection_param = 8

  feature selection model params, should be serialized LinearRegressionParams or null for pearson

• int32 num_explained_features = 9

  number of features to be explained in the interpret model

• string selected_features_file = 10

  selected features to be saved

This message denotes the LIME interpret model training parameters

• string selected_data_file = 1

  selected data file, either selected_samples_file or selected_features_file

• string selected_predictions_file = 2

  selected predictions saved

• string sample_weights_file = 3

  sample weights file saved

• int32 num_samples = 4

  number of samples generated or selected

• int32 class_num = 5

  number of classes in classification, set to 1 if regression

• int32 class_id = 6

  the label id to be explained

• string interpret_model_name = 7

  interpretable model name, linear_regression or decision_tree

• string interpret_model_param = 8

  interpretable model params, should be serialized LinearRegressionParams or DecisionTreeParams

• string explanation_report = 9

  explanation report

This message denotes the LIME interpretability sampling parameters

• int32 explain_instance_idx = 1

  the instance index for explain

• bool sample_around_instance = 2

  whether sampling around the above instance

• int32 num_total_samples = 3

  number of total samples to be generated

• string sampling_method = 4

  the sampling method, now only support "gaussian"

• string generated_sample_file = 5

  generated samples save file

This message denotes the linear regression parameters

• int32 batch_size = 1

  batch size in each iteration

• int32 max_iteration = 2

  maximum number of iterations for training

• double converge_threshold = 3

  tolerance of convergence

• bool with_regularization = 4

  whether use regularization or not

• double alpha = 5

  regularization parameter

• double learning_rate = 6

  learning rate for parameter updating

• double decay = 7

  decay rate for learning rate, following lr = lr0 / (1 + decay*t), t is #iteration

• string penalty = 8

  penalty method used, 'l1' or 'l2', default l2, currently support 'l2'

• string optimizer = 9

  optimization method, default 'sgd', currently support 'sgd'

• string metric = 10

  evaluation metric for training and testing, 'mse'

• double differential_privacy_budget = 11

  differential privacy budget

• bool fit_bias = 12

  whether to fit the bias the term

This message denotes the logistic regression parameters

• int32 batch_size = 1

  batch size in each iteration

• int32 max_iteration = 2

  maximum number of iterations for training

• double converge_threshold = 3

  tolerance of convergence

• bool with_regularization = 4

  whether use regularization or not

• double alpha = 5

  regularization parameter

• double learning_rate = 6

  learning rate for parameter updating

• double decay = 7

  decay rate for learning rate, following lr = lr0 / (1 + decay*t), t is #iteration

• string penalty = 8

  penalty method used, 'l1' or 'l2', default l2, currently support 'l2'

• string optimizer = 9

  optimization method, default 'sgd', currently support 'sgd'

• string multi_class = 10

  strategy for handling multi-class classification, default 'ovr', currently support 'ovr'

• string metric = 11

  evaluation metric for training and testing, 'acc', 'auc', or 'ks', currently support 'acc'

• double differential_privacy_budget = 12

  differential privacy budget

• bool fit_bias = 13

  whether to fit the bias the term

• int32 weight_size = 1

  model weight size

• repeated FixedPointEncodedNumber model_weights = 2

  model weights vector

• bool is_classification = 1

  whether classification or regression

• int32 num_inputs = 2

  the number of inputs (input layer size)

• int32 num_outputs = 3

  the number of outputs (output layer size)

• int32 num_hidden_layers = 4

  the number of hidden_layers

• repeated int32 num_layers_neurons = 5

  the number of neurons in each layer

• repeated Layer layers = 6

  the vector of layers

• int32 n_layers = 7

  the number of layers

This message denotes the multi-layer perceptron builder parameters

• bool is_classification = 1

  whether classification or regression

• int32 batch_size = 2

  size of mini-batch in each iteration

• int32 max_iteration = 3

  maximum number of iterations for training

• double converge_threshold = 4

  tolerance of convergence

• bool with_regularization = 5

  whether use regularization or not

• double alpha = 6

  regularization parameter

• double learning_rate = 7

  learning rate for parameter updating

• double decay = 8

  decay rate for learning rate, following lr = lr0 / (1 + decay*t), t is #iteration

• string penalty = 9

  penalty method used, 'l1' or 'l2', default l2, currently support 'l2'

• string optimizer = 10

  optimization method, default 'sgd', currently support 'sgd'

• string metric = 11

  evaluation metric for training and testing, 'mse'

• double dp_budget = 12

  differential privacy (DP) budget, 0 denotes not use DP

• bool fit_bias = 13

  whether to fit the bias term

• repeated int32 num_layers_outputs = 14

  the number of neurons in each layer

• repeated string layers_activation_funcs = 15

  the vector of layers activation functions

• repeated string ips = 1

  ips of different parties

• repeated PortArray executor_executor_port_arrays = 2

  an array of executor to executor PortArray

• optional PortArray executor_mpc_port_array = 3

  PortArray for executor to mpc ports

store the tree model on disk

Used in: TreeModel

• int32 node_type = 1

  node type, default is internal node

• int32 depth = 2

  the depth of the current node, root node is 0, -1: not decided

• int32 is_self_feature = 3

  if the node belongs to the party itself, 0: no, 1: yes, -1: not decided

• int32 best_party_id = 4

  the party that owns the selected feature on this node, -1: not decided

• int32 best_feature_id = 5

  the feature on this node, -1: not self feature, 0 -- d_i: self feature id

• int32 best_split_id = 6

  the split of the feature on this node, -1: not decided

• double split_threshold = 7

  the split threshold if it is its own feature

• int32 node_sample_num = 8

  the number of samples where the element in sample_iv is [1]

• repeated int32 node_sample_distribution = 9

  the number of samples for each class on the node

• optional FixedPointEncodedNumber impurity = 10

  node impurity, Gini index for classification, variance for regression

• optional FixedPointEncodedNumber label = 11

  if is_leaf is true, a label is assigned

• int32 left_child = 12

  left branch id of the current node, if not a leaf node, -1: not decided

• int32 right_child = 13

  right branch id of the current node, if not a leaf node, -1: not decided

• int32 source_client_id = 1

  client who sends the message

• int32 best_client_id = 2

  client owns the best split in the current node, should be same as first parameter

• int32 best_feature_id = 3

  best feature id on the best_client_id

• int32 best_split_id = 4

  best split id for the best_feature_id

• optional FixedPointEncodedNumber left_branch_impurity = 5

  impurity of the left child node

• optional FixedPointEncodedNumber right_branch_impurity = 6

  impurity of the right child node

• repeated FixedPointEncodedNumber left_branch_sample_ivs = 7

  sample iv of the left child node

• repeated FixedPointEncodedNumber right_branch_sample_ivs = 8

  sample iv of the right child node

other workers will send requests to this address

Used in: PSNetworkConfig

• string ps_ip = 1

  ip of parameter server

• int32 ps_port = 2

  port of parameter server

one ps and many workers,

• repeated Worker workers = 1

  an array of followers

• repeated PS ps = 2

  ps information, ps need multiple port where each is corresponding to one worker

Used in: NetworkConfig

• repeated int32 ports = 1

  ports of each part

This message denotes the random forest builder parameters

• int32 n_estimator = 1

  number of trees in the forest

• double sample_rate = 2

  sample rate for each tree in the forest

• optional DecisionTreeParams dt_param = 3

  decision tree builder params

• int32 global_split_num = 1

  total split num of all clients

• repeated int32 split_num_vec = 2

  split_num of each client

decision tree model

Used in: ForestModel, GbdtModel

• int32 tree_type = 1

  classification or regression

• int32 class_num = 2

  number of classes if classification

• int32 max_depth = 3

  maximum tree depth

• repeated Node nodes = 4

  array of Node

• int32 internal_node_num = 5

  internal node count

• int32 total_node_num = 6

  total node count

• int32 capacity = 7

  tree capacity

ps will read those and send message to the worker address

Used in: PSNetworkConfig

• string worker_ip = 1

  ip of follower

• int32 worker_port = 2

  port of follower