optimizer.proto

syntax = "proto2";

package object_detection.protos;

// Messages for configuring the optimizing strategy for training object
// detection models.

// Top level optimizer message.
message Optimizer {
  oneof optimizer {
    RMSPropOptimizer rms_prop_optimizer = 1;
    MomentumOptimizer momentum_optimizer = 2;
    AdamOptimizer adam_optimizer = 3;
  }
  optional bool use_moving_average = 4 [default = true];
  optional float moving_average_decay = 5 [default = 0.9999];
}

// Configuration message for the RMSPropOptimizer
// See: https://www.tensorflow.org/api_docs/python/tf/train/RMSPropOptimizer
message RMSPropOptimizer {
  optional LearningRate learning_rate = 1;
  optional float momentum_optimizer_value = 2 [default = 0.9];
  optional float decay = 3 [default = 0.9];
  optional float epsilon = 4 [default = 1.0];
}

// Configuration message for the MomentumOptimizer
// See: https://www.tensorflow.org/api_docs/python/tf/train/MomentumOptimizer
message MomentumOptimizer {
  optional LearningRate learning_rate = 1;
  optional float momentum_optimizer_value = 2 [default = 0.9];
}

// Configuration message for the AdamOptimizer
// See: https://www.tensorflow.org/api_docs/python/tf/train/AdamOptimizer
message AdamOptimizer {
  optional LearningRate learning_rate = 1;
  // Default value for epsilon (1e-8) matches default value in
  // tf.compat.v1.train.AdamOptimizer. This differs from tf2 default of 1e-7
  // in tf.keras.optimizers.Adam .
  optional float epsilon = 2 [default = 1e-8];
}


// Configuration message for optimizer learning rate.
message LearningRate {
  oneof learning_rate {
    ConstantLearningRate constant_learning_rate = 1;
    ExponentialDecayLearningRate exponential_decay_learning_rate = 2;
    ManualStepLearningRate manual_step_learning_rate = 3;
    CosineDecayLearningRate cosine_decay_learning_rate = 4;
  }
}

// Configuration message for a constant learning rate.
message ConstantLearningRate {
  optional float learning_rate = 1 [default = 0.002];
}

// Configuration message for an exponentially decaying learning rate.
// See https://www.tensorflow.org/versions/master/api_docs/python/train/ \
//     decaying_the_learning_rate#exponential_decay
message ExponentialDecayLearningRate {
  optional float initial_learning_rate = 1 [default = 0.002];
  optional uint32 decay_steps = 2 [default = 4000000];
  optional float decay_factor = 3 [default = 0.95];
  optional bool staircase = 4 [default = true];
  optional float burnin_learning_rate = 5 [default = 0.0];
  optional uint32 burnin_steps =  6 [default = 0];
  optional float min_learning_rate =  7 [default = 0.0];
}

// Configuration message for a manually defined learning rate schedule.
message ManualStepLearningRate {
  optional float initial_learning_rate = 1 [default = 0.002];
  message LearningRateSchedule {
    optional uint32 step = 1;
    optional float learning_rate = 2 [default = 0.002];
  }
  repeated LearningRateSchedule schedule = 2;

  // Whether to linearly interpolate learning rates for steps in
  // [0, schedule[0].step].
  optional bool warmup = 3 [default = false];
}

// Configuration message for a cosine decaying learning rate as defined in
// object_detection/utils/learning_schedules.py
message CosineDecayLearningRate {
  optional float learning_rate_base = 1 [default = 0.002];
  optional uint32 total_steps = 2 [default = 4000000];
  optional float warmup_learning_rate = 3 [default = 0.0002];
  optional uint32 warmup_steps = 4 [default = 10000];
  optional uint32 hold_base_rate_steps = 5 [default = 0];
}