model_config_details_en.md

Task Types，worker，network configurations

Property Name	Default	Meaning
train	train	Angel task type，training model
predict	predict	use model to predict
inctrain	inctrain	incremental training of existing models
ml.matrix.dot.use.parallel.executor	false	whether to use parallel in dot of dense matrix
angel.worker.thread.num	1	the number of threads in a worker
angel.compress.bytes	8	low precision compression, the size of each floating point number can be set to [1,8]

Data Parameters

Property Name	Default	Meaning
ml.data.type	libsvm	Angel input data types，supports:libsvm，dense，dummy
ml.data.splitor	\s+	input data separator, customizable separator
ml.data.has.label	true	whether the input data has a label, the default has a label
ml.data.label.trans.class	NoTrans	trans the label of input data ，supports；NoTrans,PosNegTrans(threshold)<use threshold to convert the label +1, -1，when greater than the threshold return 1, otherwise -1>,ZeroOneTrans(threshold)<use threshold to convert the label 1, 0，when greater than the threshold return 0>,AddOneTrans<all labels +1>,SubOneTrans<all labels -1>
ml.data.label.trans.threshold	0	the threshold of label trans ，use with PosNegTrans and ZeroOneTrans in ml.data.label.trans.class
ml.data.validate.ratio	0.05	proportion of data used for validation (no validation when set to 0)
ml.feature.index.range	-1	input data dimension, because the feature hash, can not fill the entire hash space, there are a lot of gaps, the configuration is the size of the hash space. The maximum featureID value +1, when the -1 is selected, the feature dimension can be mapped to (0, Long.max)
ml.block.size	1000000	The size of each block after dividing the matrix, the number of rows * the number of columns <= block.size, the purpose is to make the matrix partition evenly
ml.data.use.shuffle	false	whether to shuffle the data
ml.data.posneg.ratio	-1	positive and negative sample sampling ratio, -1 means off sampling function, normal value is positive real number (0~1), useful for positive and negative samples with large difference (such as 5 times or more)

Data Types

name	Description
libsvm	each line of text represents a sample, and the format of each sample is "y index1:value1 index2:value2 index3:value3 ...". Where: index is the ID of the feature, and value is the corresponding feature value; y of the training data is the class of the sample, and two values of 1, 1 can be taken; y of the predicted data is the ID value of the sample. For example, the text of a sample [2.0, 3.1, 0.0, 0.0, -1, 2.2] belonging to a positive class is represented as "1 0:2.0 1:3.1 4:-1 5:2.2", where "1" is a category," 0: 2.0" indicates that the value of the 0th feature is 2.0. Similarly, the samples belonging to the negative class [2.0, 0.0, 0.1, 0.0, 0.0, 0.0] are expressed as "-1 0: 2.0 2: 0.1"
dense	each line of text represents a sample, and the format of each sample is "y value1 value2 value3 ...". The y of the training data is the category of the sample, which can take two values of 1, 1; the y of the predicted data is the ID value of the sample. For example, the text of the sample [2.0, 3.1, 0.0, 0.0, -1, 2.2] belonging to the positive class is represented as "1 2.0 3.1 -1 2.2", where "1" is the category and "2.0" is the 0th feature. The value is 2.0. Similarly, the samples belonging to the negative class [2.0, 0.0, 0.1, 0.0, 0.0, 0.0] are expressed as "-1 2.0 0.1"
dummy	each line of text represents a sample, and each sample has the format "y index1 index2 index3 ...". Where: the ID of the index feature; the y of the training data is the category of the sample, which can take two values of 1, 1; the y of the predicted data is the ID value of the sample. For example, the text of a sample [2.0, 3.1, 0.0, 0.0, -1, 2.2] belonging to a positive class is represented as "1 0 1 4 5", where "1" is a category and "0 1 4 5" represents a feature vector. The values of the 0th, 1st, 4th, and 5th dimensions are not 0. Similarly, the samples belonging to the negative class [2.0, 0.0, 0.1, 0.0, 0.0, 0.0] are represented as "-1 0 2"

Model Parameters

Property Name	Default	Meaning
ml.model.class.name	""	the name of Angel model
ml.model.size	-1	the size of model，when -1 is selected, the range is (0, +long.max)
ml.model.type	RowType.T_FLOAT_DENSE.toString	Angel model type，suports 28 kinds
ml.model.is.classification	true	whether the model belongs to the classification model
ml.epoch.num	30	Number of iterations
ml.batch.sample.ratio	1.0	Indicates the percentage of each batch as a percentage of the overall data
ml.learn.rate	0.5	learning rate
ml.num.update.per.epoch	10	the update number of parameter in one epoch
ml.opt.decay.class.name	StandardDecay	the name of decay class，optional：StandardDecay，WarmRestarts，CorrectionDecay，ConstantLearningRate
ml.opt.decay.on.batch	false	whether to decay in batches
ml.opt.decay.intervals	100	the intervals of decay
ml.opt.decay.alpha	0.001	alpha in decay
ml.opt.decay.beta	0.001	beta in decay

Model Types

Property Name	Default	Meaning
RowType.T_DOUBLE_DENSE	0	Indicates that the type of model data row is a dense Double type with an index value in the Int range
RowType.T_DOUBLE_DENSE_COMPONENT	1	Indicates that the type of the model data row is a combined dense Double type with an index value in the Int range
RowType.T_DOUBLE_DENSE_LONGKEY_COMPONENT	2	Indicates that the type of the model data row is a combined dense Double type with an index value in the Long range.
RowType.T_DOUBLE_SPARSE	3	Indicates that the type of the model data row is a sparse Double type with an index value in the Int range
RowType.T_DOUBLE_SPARSE_COMPONENT	4	Indicates that the type of the model data row is the combined sparse Double type with the index value in the Int range.
RowType.T_DOUBLE_SPARSE_LONGKEY	5	Indicates that the type of the model data row is a sparse Double type with an index value in the Long range.
RowType.T_DOUBLE_SPARSE_LONGKEY_COMPONENT	6	Indicates that the type of the model data row is the combined sparse Double type with the index value in the Int range.
RowType.T_FLOAT_DENSE	7	Indicates that the type of the model data row is a dense Float type with an index value in the Int range
RowType.T_FLOAT_DENSE_COMPONENT	8	Indicates that the type of the model data row is a combined dense Float type with an index value in the Int range.
RowType.T_FLOAT_DENSE_LONGKEY_COMPONENT	9	Indicates that the type of the model data row is a combined dense Float type with an index value in the Long range.
RowType.T_FLOAT_SPARSE	10	Indicates that the type of model data row is a sparse Float type with an index value in the Int range
RowType.T_FLOAT_SPARSE_COMPONENT	11	Indicates that the type of model data row is a combined sparse Float type with an index value in the Int range
RowType.T_FLOAT_SPARSE_LONGKEY	12	Indicates that the type of the model data row is a sparse Float type with an index value in the Long range
RowType.T_FLOAT_SPARSE_LONGKEY_COMPONENT	13	Indicates that the type of the model data row is a combined sparse Float type with an index value in the Long range
RowType.T_LONG_DENSE	14	Indicates that the type of the model data row is a dense Long type with an index value in the Int range
RowType.T_LONG_DENSE_COMPONENT	15	Indicates that the type of the model data row is the combined dense Long type with the index value in the Int range
RowType.T_LONG_DENSE_LONGKEY_COMPONENT	16	Indicates that the type of the model data row is the combined dense Long type with the index value in the Long range
RowType.T_LONG_SPARSE	17	Indicates that the type of the model data row is a sparse Long type with an index value in the Long range
RowType.T_LONG_SPARSE_COMPONENT	18	Indicates that the type of the model data row is the combined sparse Long type of the index value in the Int range
RowType.T_LONG_SPARSE_LONGKEY	19	Indicates that the type of the model data row is a sparse Long type with an index value in the Long range
RowType.T_LONG_SPARSE_LONGKEY_COMPONENT	20	Indicates that the type of the model data row is the combined sparse Long type with the index value in the Long range
RowType.T_INT_DENSE	21	Indicates that the type of model data row is a dense Int type with an index value in the Int range
RowType.T_INT_DENSE_COMPONENT	22	Indicates that the type of the model data row is the combined dense Int type with the index value in the Int range
RowType.T_INT_DENSE_LONGKEY_COMPONENT	23	Indicates that the type of the model data row is a combined dense Int type with an index value in the Long range
RowType.T_INT_SPARSE	24	Indicates that the type of the model data row is a sparse Int type with an index value in the Int range
RowType.T_INT_SPARSE_COMPONENT	25	Indicates that the type of the model data row is a sparse Int type with an index value in the Int range
RowType.T_INT_SPARSE_LONGKEY	26	Indicates that the type of the model data row is a sparse Int type with an index value in the Long range
RowType.T_INT_SPARSE_LONGKEY_COMPONENT	27	Indicates that the type of the model data row is the combined sparse Int type with the index value in the Long range

Optimizer Configuration

Property Name	Default	Meaning
ml.fclayer.optimizer	Momentum	full connection layer optimizer, optional optimizer: Momentum，AdaDelta，AdaGrad，Adam，FTRL
ml.embedding.optimizer	Momentum	embedding layer optimizer，optional optimizer: Momentum，AdaDelta，AdaGrad，Adam，FTRL
ml.inputlayer.optimizer	Momentum	input layer optimizer, optional optimizer: Momentum，AdaDelta，AdaGrad，Adam，FTRL
ml.fclayer.matrix.output.format	classOf[RowIdColIdValueTextRowFormat].getCanonicalName	the output format of full connection layer
ml.embedding.matrix.output.format	classOf[TextColumnFormat].getCanonicalName	the output format of embedding layer
ml.simpleinputlayer.matrix.output.format	classOf[ColIdValueTextRowFormat].getCanonicalName	the output format of simpleinput layer
ml.reg.l2	0.0	coefficient of the L2 penalty
ml.reg.l1	0.0	coefficient of the L1 penalty

Momentum

Property Name	Default	Meaning
ml.opt.momentum.momentum	0.9	momentum

AdaDelta

Property Name	Default	Meaning
ml.opt.adadelta.alpha	0.9	alpha
ml.opt.adadelta.beta	0.9	beta

AdaGrad

Property Name	Default	Meaning
ml.opt.adagrad.beta	0.9	beta

Adam

Property Name	Default	Meaning
ml.opt.adam.gamma	0.99	gamma
ml.opt.adam.beta	0.9	beta

FTRL

Property Name	Default	Meaning
ml.opt.ftrl.alpha	0.1	alpha
ml.opt.ftrl.beta	1.0	beta

Layers, models Parameter Configuration

Embedding Parameter Configuration

Property Name	Default	Meaning
ml.fm.field.num	-1	feature dimension, -1 for all features, and the range can be mapped to (0, +long.max)
ml.fm.rank	8	the length of vector in embedding

(MLP) Layer Parameter Configuration

Property Name	Default	Meaning
ml.num.class	2	the number of classification

(MLR) Layer Parameter Configuration

Property Name	Default	Meaning
ml.mlr.rank	5	the number of fields

RobustRegression Parameter Configuration

Property Name	Default	Meaning
ml.robustregression.loss.delta	1.0	residual difference section point

Kmeans Parameter Configuration

Property Name	Default	Meaning
ml.kmeans.center.num	5	the number of clusters
ml.kmeans.c	0.1	learning rate

GBDT Parameter Configuration

Property Name	Default	Meaning
ml.gbdt.task.type	classification	task type，Optional：classification, regression
ml.gbdt.class.num	2	the number of the classification
ml.gbdt.tree.num	10	the number of trees
ml.gbdt.tree.depth	5	maximum tree depth
ml.gbdt.max.node.num	(none)	the maximum number of nodes
ml.gbdt.split.num	5	maximum size of grad/hess histograms for each feature
ml.gbdt.sample.ratio	1	proportion of features selected for training; default is 1
ml.gbdt.min.child.weight	0.01	the minimum child weight
ml.gbdt.reg.alpha	0	L1 regular
ml.gbdt.reg.lambda	1.0	L2 regular
ml.gbdt.thread.num	20	the number of threads
ml.gbdt.batch.size	10000	the size of a batch
ml.gbdt.server.split	false	if true, use two-stage tree splitting; default is false
ml.gbdt.cate.feat	none	categorical features，with the format of "feature id : feature rang" ("0:2,1:3" for instance). "none"

Evaluation

Property Name	Default	Meaning
train.loss	(none)	the loss of training data
validate.loss	(none)	the loss of validating data
log.likelihood	(none)	log likelihood
train.error	(none)	the error of training data
validate.error	(none)	the error of validating data