package PS.LM

Get desktop application:
View/edit binary Protocol Buffers messages

• optional DataConfig training_data = 1

• optional DataConfig validation_data = 2

• optional DataConfig model_output = 4

• optional DataConfig model_input = 5

• optional LossConfig loss = 10

• optional PenaltyConfig penalty = 11

• optional LearningRateConfig learning_rate = 12

• optional SGDConfig async_sgd = 17

• optional BCDConfig darlin = 15

Used in: Config

• optional LearningRateConfig.Type type = 1

• optional double alpha = 2

• optional double beta = 3

Used in: LearningRateConfig

• CONSTANT = 1

  = alpha

• DECAY = 2

  = alpha / (beta + x), where x is user-defined, such as sqrt(iter)

Used in: Config

• required LossConfig.Type type = 1

Used in: LossConfig

• SQUARE = 1

• LOGIT = 2

• HINGE = 3

• SQUARE_HINGE = 4

Used in: Config

• required PenaltyConfig.Type type = 1

• repeated double lambda = 2

Used in: PenaltyConfig

• L1 = 1

  lambda(0) * ||w||_1 + lambda(1) * ||w||_F^2

• L2 = 2

  lambda(0) * ||w||_F^2

Used in: Config

• required SGDConfig.Algo algo = 1

• optional int32 minibatch = 2

  The size of minibatch

• optional int32 data_buf = 12

  in mb

• optional bool ada_grad = 5

• optional int32 max_delay = 4

• optional int32 num_data_pass = 11

  The number of data passes

• optional int32 report_interval = 3

  in sec

• optional int32 tail_feature_freq = 6

  features which occurs <= *tail_feature_freq* will be filtered before training. it save both memory and bandwidth.

• optional float countmin_n = 8

  It controls the countmin size. We filter the tail features by countmin, which is more efficient than hash, but still is the memory bottleneck for servers. A smaller ratio reduces the memory footprint, but may increase the size of filtered feature.

• optional int32 countmin_k = 7

• repeated FilterConfig push_filter = 13

• repeated FilterConfig pull_filter = 14

Used in: SGDConfig

• STANDARD = 1

• FTRL = 2