Skip to content

Latest commit

 

History

History
208 lines (183 loc) · 6.89 KB

pnn_sona_en.md

File metadata and controls

208 lines (183 loc) · 6.89 KB
Raw

PNN

1. Introduction of Algorithm

The algorithm is based on Embedding, perform a pair of inner or outer product on embedding result, and then the inner/outer product result is spliced together with the original Embedding result and is entered into the DNN to further extract the high-order crossed feature. It is worth noting that the PNN does not abandon the first-order feature, and finally first-order features are combined with higher-order features for prediction. The structure is as follows: PNN

ps: Angel currently only implements PNN in inner product version.

1.1 Description of the BiInnerCross layer

In implementation, it is stored by Embedding. After calling Embedding's 'calOutput', computing and output result together. So the Embedding output of a sample is

model

The inner product of each the Embedding feature pairs is:

model

The above is BiInnerCross's forward calculation method, which is implemented by Scala code:

(0 until batchSize).foreach { row =>
    val partitions = mat.getRow(row).getPartitions
    var opIdx = 0
    partitions.zipWithIndex.foreach { case (vector_outter, cidx_outter) =>
    if (cidx_outter != partitions.length - 1) {
        ((cidx_outter + 1) until partitions.length).foreach { cidx_inner =>
        data(row * outputDim + opIdx) = vector_outter.dot(partitions(cidx_inner))
        opIdx += 1
        }
    }
    }
}

The difference between BiInnerCross and BiInnerSumCross is that the latter sums the results of the inner products and outputs them as a scalar. The former is not added up, and the output is a vector. For BiInnerCross, the output dimension isis the number of fields, regardless of the dimension of the Embedding vector.

1.2 Description of other layers

  • SparseInputLayer: Sparse data input layer, specially optimized for sparse high-dimensional data, essentially a FClayer
  • Embedding: Implicit embedding layer, if the feature is not one-hot, multiply the feature value
  • FCLayer: The most common layer in DNN, linear transformation followed by transfer function
  • SumPooling: Adding multiple input data as element-wise, requiring inputs have the same shape
  • SimpleLossLayer: Loss layer, you can specify different loss functions

1.3 Building Network

  override def buildNetwork(): Unit = {
    val wide = new SparseInputLayer("input", 1, new Identity(),
      JsonUtils.getOptimizerByLayerType(jsonAst, "SparseInputLayer"))

    val embeddingParams = JsonUtils.getLayerParamsByLayerType(jsonAst, "Embedding")
      .asInstanceOf[EmbeddingParams]
    val embedding = new Embedding("embedding", embeddingParams.outputDim, embeddingParams.numFactors,
      embeddingParams.optimizer.build()
    )

    val crossOutputDim = numFields * (numFields - 1) / 2
    val innerCross = new BiInnerCross("innerPooling", crossOutputDim, embedding)

    val concatOutputDim = embeddingParams.outputDim + crossOutputDim
    val concatLayer = new ConcatLayer("concatMatrix", concatOutputDim, Array[Layer](embedding, innerCross))

    val hiddenLayers = JsonUtils.getFCLayer(jsonAst, concatLayer)

    val join = new SumPooling("sumPooling", 1, Array[Layer](wide, hiddenLayers))

    new SimpleLossLayer("simpleLossLayer", join, lossFunc)
  }

2. Running and performance

2.1 Explanation of Json configuration File

There are many parameters of PNN, which need to be specified by Json configuration file (for a complete description of Json configuration file, please refer toJson explanation), A typical example is:

{
  "data": {
    "format": "dummy",
    "indexrange": 148,
    "numfield": 13,
    "validateratio": 0.1
  },
  "model": {
    "modeltype": "T_FLOAT_SPARSE_LONGKEY",
    "modelsize": 148
  },
  "train": {
    "epoch": 10,
    "numupdateperepoch": 10,
    "lr": 0.01,
    "decay": 0.1
  },
  "default_optimizer": "Momentum",
  "layers": [
    {
      "name": "wide",
      "type": "sparseinputlayer",
      "outputdim": 1,
      "transfunc": "identity"
    },
    {
      "name": "embedding",
      "type": "embedding",
      "numfactors": 8,
      "outputdim": 104,
      "optimizer": {
        "type": "momentum",
        "momentum": 0.9,
        "reg2": 0.01
      }
    },
    {
      "name": "biInnerCross",
      "type": "BiInnerCross",
      "outputdim": 78,
      "inputlayer": "embedding"
    },
    {
      "name": "concatlayer",
      "type": "ConcatLayer",
      "outputdim": 182,
      "inputlayers": [
        "embedding",
        "biInnerCross"
      ]
    },
    {
      "name": "fclayer",
      "type": "FCLayer",
      "outputdims": [
        200,
        200,
        1
      ],
      "transfuncs": [
        "relu",
        "relu",
        "identity"
      ],
      "inputlayer": "concatlayer"
    },
    {
      "name": "sumPooling",
      "type": "SumPooling",
      "outputdim": 1,
      "inputlayers": [
        "wide",
        "fclayer"
      ]
    },
    {
      "name": "simplelosslayer",
      "type": "simplelosslayer",
      "lossfunc": "logloss",
      "inputlayer": "sumPooling"
    }
  ]
}

2.2 Submitting script

Several steps must be done before editing the submitting script and running.

  1. confirm Hadoop and Spark have ready in your environment
  2. unzip sona--bin.zip to local directory (SONA_HOME)
  3. upload sona--bin directory to HDFS (SONA_HDFS_HOME)
  4. Edit $SONA_HOME/bin/spark-on-angel-env.sh, set SPARK_HOME, SONA_HOME, SONA_HDFS_HOME and ANGEL_VERSION

Here's an example of submitting scripts, remember to adjust the parameters and fill in the paths according to your own task.

#test description
actionType=train or predict
jsonFile=path-to-jsons/pnn.json
modelPath=path-to-save-model
predictPath=path-to-save-predict-results
input=path-to-data
queue=your-queue

HADOOP_HOME=my-hadoop-home
source ./bin/spark-on-angel-env.sh
export HADOOP_HOME=$HADOOP_HOME

$SPARK_HOME/bin/spark-submit \
  --master yarn-cluster \
  --conf spark.ps.jars=$SONA_ANGEL_JARS \
  --conf spark.ps.instances=10 \
  --conf spark.ps.cores=2 \
  --conf spark.ps.memory=10g \
  --jars $SONA_SPARK_JARS \
  --files $jsonFile \
  --driver-memory 20g \
  --num-executors 20 \
  --executor-cores 5 \
  --executor-memory 30g \
  --queue $queue \
  --class org.apache.spark.angel.examples.JsonRunnerExamples \
  ./lib/angelml-$SONA_VERSION.jar \
  jsonFile:./pnn.json \
  dataFormat:libsvm \
  data:$input \
  modelPath:$modelPath \
  predictPath:$predictPath \
  actionType:$actionType \
  numBatch:500 \
  maxIter:2 \
  lr:4.0 \
  numField:39