- PNN
- 1. 算法介绍
- 1.1 BiInnerCross层的说明
- 1.2 其它层说明
- 1.3 网络构建
- 2. 运行与性能
- 2.1 Json配置文件说明
- 2.2 提交脚本说明
- 1. 算法介绍
PNN
1. 算法介绍
FNN(Product-Based Neural Networks)算法是在Embedding的基础上, 对Embedding的结果进行两两内积或外积, 然后将内/外积结果与原始的Embedding结果拼接起来输入DNN进一步提取高阶特特交叉. 值得注意的是, PNN并没有放弃一阶特征, 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:
注: 目前Angel只实现了内积形式的PNN.
1.1 BiInnerCross层的说明
在实现中, 用Embedding的方式存储, 调用Embedding的calOutput
后, 将计算后一起输出, 所以一个样本的Embedding output结果为:
=(\bold{u}_1,\bold{u}_2,\bold{u}_3,\cdots,\bold{u}_k))
对Embedding特征两两做内积有:
)
以上即是BiInnerCross的前向计算方式, 用Scala代码实现为:
(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
}
}
}
}
BiInnerCross与BiInnerSumCross的区别在于后者将两两内积的结果加和起来输出为一个标量, 前者没有加和起来, 输出是一个向量. 对于BiInnerCross, 输出的维数为为field的个数, 与Embedding向量的维数无关.
1.2 其它层说明
- 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
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. 运行与性能
2.1 Json配置文件说明
PNN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:
{
"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 提交脚本说明
runner="com.tencent.angel.ml.core.graphsubmit.GraphRunner"
modelClass="com.tencent.angel.ml.classification.ProductNeuralNetwork"
$ANGEL_HOME/bin/angel-submit \
--angel.job.name DeepFM \
--action.type train \
--angel.app.submit.class $runner \
--ml.model.class.name $modelClass \
--angel.train.data.path $input_path \
--angel.workergroup.number $workerNumber \
--angel.worker.memory.gb $workerMemory \
--angel.ps.number $PSNumber \
--angel.ps.memory.gb $PSMemory \
--angel.task.data.storage.level $storageLevel \
--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.