Deep And Wide

1. 算法介绍

Deep and Wide算法是将Embedding的结果直接输入DNN进一步提取高阶特特交叉, 最后将一阶特征与高阶特征组合起来进行预测, 其构架如下:
DNN

1.1 Deep and Wide中的层

  • SparseInputLayer: 稀疏数据输入层, 对稀疏高维数据做了特别优化, 本质上是一个FCLayer
  • Embedding: 隐式嵌入层, 如果特征非one-hot, 则乘以特征值
  • FCLayer: DNN中最常见的层, 线性变换后接传递函数
  • SumPooling: 将多个输入的数据做element-wise的加和, 要求输入具本相同的shape
  • SimpleLossLayer: 损失层, 可以指定不同的损失函数

1.3 网络构建

  1.   override def buildNetwork(): Unit = {
  2.     val wide = new SparseInputLayer("input", 1, new Identity(),
  3.       JsonUtils.getOptimizerByLayerType(jsonAst, "SparseInputLayer"))
  5.     val embeddingParams = JsonUtils.getLayerParamsByLayerType(jsonAst, "Embedding")
  6.       .asInstanceOf[EmbeddingParams]
  7.     val embedding = new Embedding("embedding", embeddingParams.outputDim, embeddingParams.numFactors,
  8.       embeddingParams.optimizer.build()
  9.     )
  11.     val hiddenLayer = JsonUtils.getFCLayer(jsonAst, embedding)
  13.     val join = new SumPooling("sumPooling", 1, Array[Layer](wide, hiddenLayer))
  15.     new SimpleLossLayer("simpleLossLayer", join, lossFunc)
  16.   }

2. 运行与性能

2.1 Json配置文件说明

Deep and wide的参数较多, 需要用Json配置文件的方式指定(关于Json配置文件的完整说明请参考Json说明), 一个典型的例子如下:

  1. {
  2.   "data": {
  3.     "format": "dummy",
  4.     "indexrange": 148,
  5.     "numfield": 13,
  6.     "validateratio": 0.1
  7.   },
  8.   "model": {
  9.     "modeltype": "T_DOUBLE_SPARSE_LONGKEY",
  10.     "modelsize": 148
  11.   },
  12.   "train": {
  13.     "epoch": 10,
  14.     "numupdateperepoch": 10,
  15.     "lr": 0.1,
  16.     "decay": 0.8
  17.   },
  18.   "default_optimizer": {
  19.     "type": "momentum",
  20.     "momentum": 0.9,
  21.     "reg2": 0.01
  22.   },
  23.   "layers": [
  24.     {
  25.       "name": "wide",
  26.       "type": "sparseinputlayer",
  27.       "outputdim": 1,
  28.       "transfunc": "identity"
  29.     },
  30.     {
  31.       "name": "embedding",
  32.       "type": "embedding",
  33.       "numfactors": 8,
  34.       "outputdim": 104
  35.     },
  36.     {
  37.       "name": "fclayer",
  38.       "type": "FCLayer",
  39.       "inputlayer": "embedding",
  40.       "outputdims": [
  41.         100,
  42.         100,
  43.         1
  44.       ],
  45.       "transfuncs": [
  46.         "relu",
  47.         "relu",
  48.         "identity"
  49.       ]
  50.     },
  51.     {
  52.       "name": "sumPooling",
  53.       "type": "SumPooling",
  54.       "outputdim": 1,
  55.       "inputlayers": [
  56.         "wide",
  57.         "fclayer"
  58.       ]
  59.     },
  60.     {
  61.       "name": "simplelosslayer",
  62.       "type": "simplelosslayer",
  63.       "lossfunc": "logloss",
  64.       "inputlayer": "sumPooling"
  65.     }
  66.   ]
  67. }

2.2 提交脚本说明

  1. runner="com.tencent.angel.ml.core.graphsubmit.GraphRunner"
  2. modelClass="com.tencent.angel.ml.classification.WideAndDeep"
  4. $ANGEL_HOME/bin/angel-submit \
  5.     --angel.job.name DeepFM \
  6.     --action.type train \
  7.     --angel.app.submit.class $runner \
  8.     --ml.model.class.name $modelClass \
  9.     --angel.train.data.path $input_path \
  10.     --angel.workergroup.number $workerNumber \
  11.     --angel.worker.memory.gb $workerMemory  \
  12.     --angel.ps.number $PSNumber \
  13.     --angel.ps.memory.gb $PSMemory \  
  14.     --angel.task.data.storage.level $storageLevel \
  15.     --angel.task.memorystorage.max.gb $taskMemory

对深度学习模型, 其数据, 训练和网络的配置请优先使用Json文件指定.