Save Keras-like model to pure keras or tensorflow protobuf. (intel-an…

…alytics#1600)

* checkpoint

* some update

* refine api

* some update

* fix build fail

* meet code review

* style check

* fix typo

* fix style check

layers/Dense.scala

@@ -59,6 +59,24 @@ class Dense[T: ClassTag](
   override val inputShape: Shape = null)(implicit ev: TensorNumeric[T])
   extends BigDLDense[T](outputDim, init, activation, wRegularizer, bRegularizer, bias,
     inputShape) with Net {
+
+  override private[zoo] def toKeras2(dir: String): String = {
+    val params = Net.inputShapeToString(inputShape) ++
+      Net.activationToString(activation) ++
+      Net.param(getName()) ++
+      Net.param(bias, "use_bias") ++
+      Net.param(outputDim, "units")
+    Net.kerasDef(this, params)
+  }
+
+  override private[zoo] def getKerasWeights(): Array[Tensor[Float]] = {
+    val weights = this.parameters()._1
+    val kWeights = Array.tabulate(weights.length)(_ => Tensor[Float]())
+    weights(0) = weights(0).t().contiguous()
+    weights(0).cast[Float](kWeights(0).resizeAs(weights(0)))
+    weights(1).cast[Float](kWeights(1).resizeAs(weights(1)))
+    kWeights
+  }
 }
 
 object Dense {

layers/Dropout.scala

@@ -36,7 +36,16 @@ class Dropout[T: ClassTag](
   override val p: Double,
   override val inputShape: Shape = null)
   (implicit ev: TensorNumeric[T])
-  extends com.intel.analytics.bigdl.nn.keras.Dropout[T](p, inputShape) with Net {}
+  extends com.intel.analytics.bigdl.nn.keras.Dropout[T](p, inputShape) with Net {
+
+  override private[zoo] def toKeras2(dir: String): String = {
+    val params = Net.inputShapeToString(inputShape) ++
+      Net.param(getName()) ++
+      Net.param(p, "rate")
+    Net.kerasDef(this, params)
+  }
+
+}
 
 object Dropout {
   def apply[@specialized(Float, Double) T: ClassTag](

layers/LSTM.scala

@@ -78,6 +78,41 @@ class LSTM[T: ClassTag](
       uRegularizer = uRegularizer,
       bRegularizer = bRegularizer)
   }
+
+  override private[zoo] def toKeras2(dir: String): String = {
+    val params = Net.inputShapeToString(inputShape) ++
+      Net.activationToString(activation) ++
+      Net.activationToString(innerActivation, "recurrent_activation") ++
+      Net.param(returnSeq, "return_sequences") ++
+      Net.param(outputDimension, "units")
+      Net.param(getName())
+    Net.kerasDef(this, params)
+  }
+
+  override private[zoo] def getKerasWeights(): Array[Tensor[Float]] = {
+    val weights = this.parameters()._1
+    val kWeights = Array.tabulate(weights.length)(_ => Tensor[Float]())
+    weights(0) = weights(0).t().contiguous()
+    weights(2) = weights(2).t().contiguous()
+    weights(0).cast[Float](kWeights(0).resizeAs(weights(0)))
+    weights(2).cast[Float](kWeights(1).resizeAs(weights(2)))
+    weights(1).cast[Float](kWeights(2).resizeAs(weights(1)))
+    // map to keras's weight
+    switch(kWeights(0), 2)
+    switch(kWeights(1), 2)
+    switch(kWeights(2), 1)
+
+    kWeights
+  }
+
+  private def switch(t: Tensor[Float], dim: Int): Unit = {
+    val tmpWeight = t.narrow(dim, 1, outputDimension).clone()
+    tmpWeight.copy(t.narrow(dim, 1 + outputDimension, outputDimension))
+    t.narrow(dim, 1 + outputDimension, outputDimension)
+      .copy(t.narrow(dim, 2 * outputDimension + 1, outputDimension))
+    t.narrow(dim, 2 * outputDimension + 1, outputDimension).copy(tmpWeight)
+  }
+
 }
 
 object LSTM {

layers/Permute.scala

@@ -40,6 +40,13 @@ class Permute[T: ClassTag](
     override val inputShape: Shape = null)(implicit ev: TensorNumeric[T])
   extends BigDLPermute[T](
     dims, inputShape) with Net {
+
+  override private[zoo] def toKeras2(dir: String): String = {
+    val params = Net.inputShapeToString(inputShape) ++
+      Net.param(getName()) ++
+      Net.arrayToString(dims, "dims")
+    Net.kerasDef(this, params)
+  }
 }
 
 object Permute {

layers/Reshape.scala

@@ -98,6 +98,13 @@ class Reshape[T: ClassTag](
     }
     layer.asInstanceOf[AbstractModule[Tensor[T], Tensor[T], T]]
   }
+
+  override private[zoo] def toKeras2(dir: String): String = {
+    val params = Net.inputShapeToString(inputShape) ++
+      Net.param(getName()) ++
+      Net.arrayToString(targetShape, "target_shape")
+    Net.kerasDef(this, params)
+  }
 }
 
 object Reshape {

layers/utils/KerasUtils.scala

@@ -20,7 +20,7 @@ import com.intel.analytics.bigdl.Criterion
 import com.intel.analytics.bigdl.nn.Graph.ModuleNode
 import com.intel.analytics.bigdl.nn._
 import com.intel.analytics.bigdl.nn.keras.{KerasIdentityWrapper, KerasLayer, KerasLayerWrapper, Sequential => KSequential, SoftMax => KSoftMax}
-import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, Activity, DataFormat}
+import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, Activity, DataFormat, TensorModule}
 import com.intel.analytics.bigdl.optim._
 import com.intel.analytics.bigdl.tensor.Tensor
 import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric
@@ -84,6 +84,32 @@ object KerasUtils {
     }
   }
 
+  def getActivationName[T: ClassTag](activation: AbstractModule[_, _, T]): String = {
+    if (activation == null) {
+      throw new IllegalArgumentException("activation is null")
+    } else {
+      activation match {
+        case _: Tanh[T] => "tanh"
+        case _: Sigmoid[T] => "sigmoid"
+        case _: ReLU[T] => "relu"
+        case _: com.intel.analytics.bigdl.nn.SoftMax[T] => "softmax"
+        case _: SoftPlus[T] => "softplus"
+        case _: SoftSign[T] => "softsign"
+        case _: HardSigmoid[T] => "hard_sigmoid"
+        case _: ReLU6[T] => "relu6"
+        case _: TanhShrink[T] => "tanh_shrink"
+        case _: SoftMin[T] => "softmin"
+        case _: LogSigmoid[T] => "log_sigmoid"
+        case _: LogSoftMax[T] => "log_softmax"
+        case _: Identity[T] => "linear"
+        case _: com.intel.analytics.zoo.pipeline.api.keras.layers.SoftMax[T] => "softmax"
+        case _ => throw new IllegalArgumentException("unkown activation"
+          + activation.getClass.getName)
+      }
+    }
+
+  }
+
   def getTorchActivation[T : ClassTag] (activation: String)
     (implicit ev: TensorNumeric[T]): AbstractModule[Tensor[T], Tensor[T], T] = {
     if (activation == null) null

models/Topology.scala

@@ -547,6 +547,29 @@ abstract class KerasNet[T](implicit val tag: ClassTag[T], implicit val ev: Tenso
 
   def toModel(): Model[T]
 
+
+  /**
+   * Save model to keras2 h5 file. Only for inference
+   * @param filePath path to save model.
+   * @param python python path, need analytics-zoo and tensorflow installed.
+   */
+  def saveToKeras2[T: ClassTag](
+        filePath: String,
+        python: String = "python")(implicit ev: TensorNumeric[T]): Unit = {
+    Net.saveToKeras2[T](this, filePath, python)
+  }
+
+  /**
+   * Save model to tensorflow protobuf. Only for inference.
+   * @param dir directory to save model.
+   * @param python python path, need analytics-zoo and tensorflow installed.
+   */
+  def saveToTf[T: ClassTag](
+        dir: String,
+        python: String = "python")(implicit ev: TensorNumeric[T]): Unit = {
+    Net.saveToTf[T](this, dir, python)
+  }
+
   /**
    * Print out the summary information of an Analytics Zoo Keras Model.
    *
@@ -892,6 +915,19 @@ class Sequential[T: ClassTag] private ()
     val graph = this.toModel()
     graph.summary(lineLength, positions)
   }
+
+  override private[zoo] def getKerasWeights(): Array[Tensor[Float]] = {
+    val weights = new ArrayBuffer[Tensor[Float]]()
+    modules(0).asInstanceOf[TSequential[T]].modules.foreach(m => {
+      val params = m.asInstanceOf[Net].getKerasWeights()
+      if (params != null) {
+        params.foreach{p =>
+          weights += p
+        }
+      }
+    })
+    weights.toArray
+  }
 }
 
 object Sequential extends KerasLayerSerializable {