OptimizerConfig.proto

syntax = "proto2";
 
option optimize_for = LITE_RUNTIME;

package paddle;

message SGDConfig {
  // SGD
  // momentum: float >= 0. Parameter updates momentum.
  // decay: float >= 0. Learning rate decay over each update.
  // nesterov: boolean. Whether to apply Nesterov momentum.
  optional double momentum = 21 [default = 0.0];
  optional double decay = 23 [default = 0.0];
  optional bool nesterov =24 [default = false];

}


message AdadeltaConfig {
  // Adadelta
  // It is recommended to leave it at the default value.
  // rho: float >= 0.
  // epsilon: float >= 0. Fuzz factor.
  // decay: float >= 0. Learning rate decay over each update.

  // reference : [Adadelta - an adaptive learning rate method](http://arxiv.org/abs/1212.5701)
  optional double rho = 33 [default = 0.90];
  optional double epsilon = 31 [default = 1e-5];
  optional double decay = 32 [default = 0.0];

}

message AdagradConfig {
// Adagrad
// epsilon: float >= 0.
// decay: float >= 0. Learning rate decay over each update.

// reference : [Adaptive Subgradient Methods for Online Learning and Stochastic Optimization](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
  optional double epsilon = 41 [default = 1e-5];
  optional double decay = 42 [default = 0.0];
}

message AdamConfig {
  // Adaj
  // beta_1: float, 0 < beta < 1. Generally close to 1.
  // beta_2: float, 0 < beta < 1. Generally close to 1.
  // epsilon: float >= 0. Fuzz factor.
  // decay: float >= 0. Learning rate decay over each update.
  // reference : [Adam - A Method for Stochastic Optimization](http://arxiv.org/abs/1412.6980v8)
  optional double beta_1 = 41;
  optional double beta_2 = 42;
  optional double epsilon = 43;
  optional double decay = 44;
}

message ConstLr {
  // learninRate Policy
  required double learning_rate = 1 [default = 1.0];
}

message LinearLr {
  // learninRate Policy
  required double learning_rate = 1 [default = 1.0];
  optional double lr_decay_a = 2;
  optional double lr_decay_b = 3;
}

message TensorProto {
enum DataType {
  PADDLE_ELEMENT_TYPE_INT32 = 0;
  PADDLE_ELEMENT_TYPE_UINT32 = 1;
  PADDLE_ELEMENT_TYPE_INT64 = 2;
  PADDLE_ELEMENT_TYPE_UINT64 = 3;
  PADDLE_ELEMENT_TYPE_FLOAT32 = 4;
  PADDLE_ELEMENT_TYPE_FLOAT64 = 5;
}
  required DataType data_type = 1;
  repeated bytes content = 2;
}

message OptimizerState {
  optional double learning_rate = 101;
  optional double lr_decay_a = 102;
  optional double lr_decay_b = 103;
  optional double num_sample_passed = 104;
  // momentum
  optional TensorProto parameter = 105;
  optional TensorProto momentums = 1;

  // adadelta
  optional TensorProto accum_gradient = 2;
  optional TensorProto accum_delta = 3;
  optional TensorProto update_delta = 4;

  // adam
  optional TensorProto velocitys = 5;

  // momentum
  optional double momentum = 6;
  optional double decay = 7;
  optional bool nesterov = 8;

  // adadelta
  optional double rho = 9;
  optional double epsilon = 10;

  // adam
  optional double beta_1 = 11;
  optional double beta_2 = 12;

}

message OptimizerConfig {
  enum Optimizer {
   SGD = 1;
   Adadelta = 2;
   Adagrad = 3;
   Adam = 4;
  }
  required Optimizer optimizer = 1;
  optional SGDConfig sgd = 3;
  optional AdadeltaConfig adadelta = 4;
  optional AdagradConfig adagrad = 5;
  optional AdamConfig adam = 6;

  enum LrPolicy {
   ConstLr = 0;
   LinearLr = 1;
  }
  required LrPolicy lr_policy = 11;
  optional ConstLr const_lr = 12;
  optional LinearLr linear_lr = 13;

  // common config of optimizer
  // gradient clip when L2 exceeding value
  optional double clip_norm = 101;
  // gradient clip when L1 exceeding value
  optional double clip_value = 102;
}