Many resources are needed to download a project. Please understand that we have to compensate our server costs. Thank you in advance. Project price only 1 $
You can buy this project and download/modify it how often you want.
/*
* Copyright 2016 The BigDL Authors.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package com.intel.analytics.bigdl.utils.tf
import java.io.{DataInputStream, InputStream, FileReader => JFileReader}
import java.nio.ByteOrder
import java.util
import java.util.List
import java.util.{HashMap => JHashMap}
import com.google.protobuf.{CodedInputStream, TextFormat}
import com.intel.analytics.bigdl.Module
import com.intel.analytics.bigdl.nn.Graph
import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, Activity}
import com.intel.analytics.bigdl.nn.tf.AssignGrad
import com.intel.analytics.bigdl.python.api.{JTensor, PythonBigDL, PythonBigDLUtils}
import com.intel.analytics.bigdl.tensor.{DoubleType, FloatType, Tensor}
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric
import com.intel.analytics.bigdl.utils._
import com.intel.analytics.bigdl.nn.tf.{SwitchControlNode, SwitchOps}
import com.intel.analytics.bigdl.utils.tf.TensorflowToBigDL._
import com.intel.analytics.bigdl.utils.tf.loaders.TensorflowOpsLoader
import org.tensorflow.framework.{GraphDef, NodeDef}
import scala.collection.mutable
import scala.collection.mutable.ArrayBuffer
import scala.reflect.ClassTag
import scala.collection.JavaConverters._
object TensorflowLoader{
/**
* Load tensorflow model from a prototxt file
* @param graphPrototxt where is the tensorflow protobuf file
* @param inputs input node names, where feed in the data. You can feed data into an internal
* node. If the internal node has multiple dependency, you can use name:n to specify
* which dependency you choose to feed
* @param outputs output node names
* @param byteOrder file byteOrder
* @param generatedBackward if generate backward graph
* @return
*/
def load[T: ClassTag](graphPrototxt: String, inputs: Seq[String], outputs: Seq[String],
byteOrder: ByteOrder, binFile: Option[String] = None,
generatedBackward: Boolean = true)(
implicit ev: TensorNumeric[T]): Module[T] = {
// Get node list
val nodeList = parse(graphPrototxt)
// Input name remove the port
val realInputNames = inputs.map(i => if (i.split(":").length == 2) i.split(":")(0) else i)
.distinct
// Construct tf node graph
val (tfGraph, newInputMap, _) =
buildTFGraph(nodeList, outputs, (node: NodeDef) => realInputNames.contains(node.getName),
Some(getInputPorts(inputs)))
// If you choose an internal node with multiple inputs, extra placeholder will be insert into
// the model
// Keep the order with the inputs list
val newInputs = ArrayBuffer[String]()
realInputNames.foreach(i => {
if (newInputMap.isDefinedAt(i)) {
newInputMap(i).foreach(n => newInputs.append(n))
}
})
// Try to load variables
val context = binFile.map(loadBinFiles(_))
// Build BigDL model from the tf node graph
buildBigDLModel(tfGraph, newInputs, outputs, byteOrder, graphPrototxt,
context, generatedBackward)
}
def checkpoints[T: ClassTag](graphFile: String, binFile: String, byteOrder: ByteOrder)(
implicit ev: TensorNumeric[T]): Session[T] = {
// Get node list
val nodeList = parse(graphFile)
new BigDLSessionImpl[T](nodeList.asScala, loadBinFiles(binFile), byteOrder)
}
/**
* Get the input ports if user specify. It will also check the input name list.
* @param inputs
* @return
*/
private def getInputPorts(inputs: Seq[String]): mutable.Map[String, ArrayBuffer[Int]] = {
require(inputs.distinct.length == inputs.length,
"input should not contain duplicated names")
val inputPorts = inputs.filter(_.split(":").length == 2)
val result = mutable.HashMap[String, ArrayBuffer[Int]]()
inputPorts.foreach(s => {
val name = s.split(":")(0)
val pos = s.split(":")(1)
require(!inputs.contains(name), "You should not specify node name and node name " +
"with port at same time")
if (!result.isDefinedAt(name)) {
result(name) = ArrayBuffer[Int]()
}
result(name).append(pos.toInt)
})
result
}
/**
* Parse a tensorflow model protobuf binary file, read a list of op nodes from it
* @param graphProtoTxt where is the tf protobuf file
* @return
*/
private[bigdl] def parse(graphProtoTxt: String) : List[NodeDef] = {
var fr: FileReader = null
var in: InputStream = null
try {
fr = FileReader(graphProtoTxt)
in = fr.open()
val reader = CodedInputStream.newInstance(new DataInputStream(in))
reader.setSizeLimit(0x7fffffff)
val graph = GraphDef.parseFrom(reader)
graph.getNodeList
} finally {
if (fr != null) fr.close()
if (in != null) in.close()
}
}
private[bigdl] def saveBinFile[T: ClassTag](file: String, context: Context[T])
(implicit ev: TensorNumeric[T]): Unit = {
val save = new JHashMap[String, JTensor]()
context.tensorNames().foreach(n => {
val tensor = context(n)._1
val saveTensor = ev.getType() match {
case FloatType => new JTensor(tensor.asInstanceOf[Tensor[Float]].storage().array(),
tensor.size(), "float")
case DoubleType => new JTensor(tensor.asInstanceOf[Tensor[Double]].storage().array()
.map(_.toFloat), tensor.size(), "double")
case t => throw new NotImplementedError(s"$t is not supported")
}
save.put(n, saveTensor)
})
File.save(save, file, true)
}
private def loadBinFiles[T: ClassTag](file: String)(implicit ev: TensorNumeric[T]): Context[T] = {
val m = File.load(file).asInstanceOf[JHashMap[String, JTensor]].asScala
val map = new mutable.HashMap[String, (Tensor[T], Tensor[T], Option[Seq[(Int, Int)]])]()
for (k <- m.keys) {
val tensor = ev.getType() match {
case FloatType => PythonBigDLUtils.toTensor(m(k), "float")
case DoubleType => PythonBigDLUtils.toTensor(m(k), "double")
case t => throw new NotImplementedError(s"$t is not supported")
}
map(k) = (tensor, tensor.clone(), None)
}
new Context[T](map)
}
/**
* Parse a tensorflow model protobuf text file, read a list of op nodes from it
* @param graphProtoTxt where is the tf protobuf file
* @return
*/
private[bigdl] def parseTxt(graphProtoTxt: String) : List[NodeDef] = {
val f = new java.io.File(graphProtoTxt)
require(f.exists(), graphProtoTxt + " does not exists")
val reader = new JFileReader(f)
val graph = GraphDef.newBuilder()
TextFormat.merge(reader, graph)
graph.build().getNodeList
}
/**
* Build tf ops graph from a given node list. The graph output is the outputs.
*
* @param nodes Tensorflow NodeDefs
* @param outputs output node names
* @param isInput check if a node is input
* @param inputPorts if user want to use a part of the node input as model input
* @return
*/
private[bigdl] def buildTFGraph(
nodes : List[NodeDef], outputs: Seq[String],
isInput: (NodeDef) => Boolean = (_: NodeDef) => false,
inputPorts: Option[mutable.Map[String, ArrayBuffer[Int]]] = None
): (DirectedGraph[NodeDef], mutable.HashMap[String, ArrayBuffer[String]], Seq[String]) = {
val name2Node = nodes.asScala.map(n => n.getName -> new Node(n)).toMap
// Build graph
val outputNodes = if (outputs == null) {
name2Node.valuesIterator.filter(_.nextNodes.isEmpty).toArray
} else {
val results = name2Node.valuesIterator.toArray.filter(n =>
outputs.contains(n.element.getName))
require(results.length == outputs.length, "Invalid outputNode names")
results
}
val (inputs, originInputs) = connect(outputNodes, name2Node, isInput, inputPorts)
val dummyOutput = new Node[NodeDef](null)
outputNodes.foreach(_ -> dummyOutput)
(dummyOutput.graph(reverse = true), inputs, originInputs)
}
/**
* Build a graph from the output. The build process will stop when met node without inputs
* or specified input node
*/
private[bigdl] def connect(
nodes: Seq[Node[NodeDef]],
name2Node: Map[String, Node[NodeDef]],
isInput: (NodeDef) => Boolean,
inputPorts: Option[mutable.Map[String, ArrayBuffer[Int]]]
): (mutable.HashMap[String, ArrayBuffer[String]], Seq[String]) = {
var inputCounter = 0
var depCounter = 0
val queue = new mutable.Queue[Node[NodeDef]]()
val visited = mutable.Set[Node[NodeDef]]()
val newInputs = new mutable.HashMap[String, ArrayBuffer[String]]()
val originInputs = new mutable.ArrayBuffer[String]()
// Do a BFS to connect the nodes
queue.enqueue(nodes: _*)
while(queue.nonEmpty) {
val node = queue.dequeue()
if (!visited(node)) {
visited += node
if (!isInput(node.element) && !node.element.getInputList.isEmpty) {
// continue to traverse
node.element.getInputList.asScala.foreach { preNodeName =>
depCounter = pushPreNode(preNodeName, name2Node, depCounter, node, queue)
}
} else {
if (newInputs.get(node.element.getName).isEmpty) {
newInputs(node.element.getName) = new ArrayBuffer[String]()
}
if (isInput(node.element) && node.element.getOp != "Placeholder") {
// if the predefined input node is not a Placeholder, add one to match the Input node
val inputNum = getInputNumber(node.element)
if (inputNum == 0) {
require(!inputPorts.isDefined ||
!inputPorts.get.isDefinedAt(node.element.getName),
s"node ${node.element.getName} has no input")
newInputs(node.element.getName).append(node.element.getName)
} else {
if (inputPorts.isDefined &&
inputPorts.get.isDefinedAt(node.element.getName)) {
val selectInputs = inputPorts.get(node.element.getName)
selectInputs.foreach(i => require(i < inputNum && i >= 0,
s"invalid input port $i at ${node.element.getName}, it should between 0 and" +
s" ${inputNum - 1}"))
var i = 0
while (i < inputNum) {
if (selectInputs.contains(i)) {
val name = s"input$inputCounter"
val placeholder = NodeDef.newBuilder()
.setName(name)
.setOp("Placeholder").build()
inputCounter = inputCounter + 1
val n = Node(placeholder)
n -> node
newInputs(node.element.getName).append(name)
} else {
val preNodeName = node.element.getInputList.asScala.apply(i)
depCounter = pushPreNode(preNodeName, name2Node, depCounter, node, queue)
}
i = i + 1
}
} else {
val name = s"input$inputCounter"
val placeholder = NodeDef.newBuilder()
.setName(name)
.setOp("Placeholder").build()
inputCounter = inputCounter + 1
val n = Node(placeholder)
n -> node
newInputs(node.element.getName).append(name)
}
}
originInputs += node.element.getName
} else if (node.element.getOp == "Placeholder") {
newInputs(node.element.getName).append(node.element.getName)
originInputs += node.element.getName
}
}
}
}
(newInputs, originInputs)
}
private def pushPreNode(name: String, name2Node: Map[String, Node[NodeDef]],
depCounter: Int, node: Node[NodeDef], queue: mutable.Queue[Node[NodeDef]]): Int = {
// It is tricky here, remove the first char in the name of control dep node
var realName = name
var controlDep = false
var channel = 0
var _depCounter = depCounter
if (realName.charAt(0) == '^') {
realName = realName.substring(1)
controlDep = true
}
if (realName.split(":").length > 1) {
val pair = realName.split(":")
realName = pair(0)
channel = pair(1).toInt
}
val preNode = name2Node(realName)
val curNode = if (controlDep) {
val dependencyNode = Node(NodeDef.newBuilder()
.setOp("DependencyNode")
.addInput(preNode.element.getName)
.setName(s"depends_on_${preNode.element.getName}$depCounter")
.build())
_depCounter += 1
dependencyNode -> node
dependencyNode
} else {
node
}
preNode.add(curNode, Edge(channel + 1))
queue.enqueue(preNode)
_depCounter
}
private def getInputNumber(nodeDef: NodeDef): Int = {
import scala.collection.JavaConverters._
nodeDef.getOp match {
case "QueueDequeueV2" => nodeDef.getAttrOrThrow("component_types").getList.getTypeCount
case "QueueDequeueManyV2" => nodeDef.getAttrOrThrow("component_types").getList.getTypeCount
case _ => nodeDef.getInputList.asScala.filterNot(_.charAt(0) == '^').length
}
}
private[bigdl] def buildBigDLModel[T: ClassTag](
tfGraph: DirectedGraph[NodeDef],
inputs: Seq[String],
outputs: Seq[String],
byteOrder: ByteOrder,
graphPrototxt: String,
ctx: Option[Context[T]] = None,
generatedBackward: Boolean = true
)(implicit ev: TensorNumeric[T]): Module[T] = {
import scala.collection.JavaConverters._
// Map from tensorflow node to the converted BigDL node
val convertedNode = new mutable.HashMap[Node[NodeDef],
Node[AbstractModule[Activity, Activity, T]]]()
val nameToNode =
new mutable.HashMap[String, Node[AbstractModule[Activity, Activity, T]]]()
val moduleToInputNodes =
new mutable.HashMap[Node[AbstractModule[Activity, Activity, T]], Seq[Node[NodeDef]]]()
val moduleToAllNodes =
new mutable.HashMap[Node[AbstractModule[Activity, Activity, T]], Set[Node[NodeDef]]]()
val context = ctx.getOrElse(new Context[T])
// BFS to keep the input order same
tfGraph.BFS.foreach(n => {
if (n.element == null) {
// Dummy node, skip
} else if (convertedNode.get(n).isDefined) {
// converted node, skip
} else {
val errorMsg =
s"""
| Cannot convert the given tensorflow operation graph to BigDL model. The convert fails
| at node ${n.element.getName}. Operation type is ${n.element.getOp}
""".stripMargin
val (module, nodes, inputNodes) =
extract[T](n.graph(reverse = true), context, byteOrder).getOrElse({
try {
val cls = Class.forName("com.intel.analytics.bigdl.utils.tf.loaders." +
n.element.getOp)
val builder = cls.getConstructors()(0).newInstance().asInstanceOf[TensorflowOpsLoader]
(builder.build[T](n.element, byteOrder, context), Seq(n).asJava, Seq(n))
} catch {
case e: Throwable =>
throw new UnsupportedOperationException(errorMsg, e)
}
})
// set name
if (nodes.size() == 1) {
// Use tf operation name if one to one map
module.setName(removeColon(nodes.get(0).element.getName()))
} else {
// Many to one map
val name = removeColon(findCommonPrefix(nodes.asScala.map(_.element.getName)))
if (name == "") {
// Use a name combine nodes
module.setName(s"[${nodes.asScala.map(_.element.getName).map(_.replaceAll("/", "\\\\"))
.map(removeColon(_)).mkString(", ")}]")
} else {
// Use the common name
module.setName(name + "/" + module.getName())
}
}
val node = module match {
case _: SwitchOps[_] => new SwitchControlNode(module)
case _ => Node(module)
}
nodes.asScala.foreach(m => {
convertedNode(m) = node
nameToNode(m.element.getName) = node
})
moduleToInputNodes(node) = inputNodes
moduleToAllNodes(node) = nodes.asScala.toSet
}
})
/**
* Go through all tensorflow nodes
* @param outputModuleNode
*/
def connect(outputModuleNode: Seq[Node[AbstractModule[Activity, Activity, T]]]) = {
val queue = new mutable.Queue[Node[AbstractModule[Activity, Activity, T]]]()
val visited = mutable.Set[Node[AbstractModule[Activity, Activity, T]]]()
queue.enqueue(outputModuleNode: _*)
while (queue.nonEmpty) {
val currNode = queue.dequeue()
if (!visited(currNode)) {
visited += currNode
val inputNodes = moduleToInputNodes(currNode)
val allNodes = moduleToAllNodes(currNode)
val inputModuleNodes = inputNodes.flatMap(_.prevNodesAndEdges)
.filterNot(n => context.containsTensor(n._1.element.getName) &&
n._1.element.getOp() != "VariableV2")
.filterNot(n => allNodes(n._1))
.map(n => (convertedNode(n._1), n._2.newInstance())).filter(n => n._1 != currNode)
inputModuleNodes.foreach(n => n._1.add(currNode, n._2))
queue.enqueue(inputModuleNodes.map(_._1): _*)
}
}
}
val outputModules = tfGraph.source.prevNodes.map(_.element.getName).map(nameToNode)
connect(outputModules)
val inputNodes = inputs
.map(n => nameToNode.getOrElse(n, throw new IllegalArgumentException(s"Can't find node $n")))
val outputNodes = outputs
.map(n => nameToNode.getOrElse(n, throw new IllegalArgumentException(s"Can't find node $n")))
val weights = ArrayBuffer[Tensor[T]]()
val gradients = ArrayBuffer[Tensor[T]]()
for ((weight, grad, _) <- context.tensors) {
weights += weight
gradients += grad
}
// Append assign nodes
val adjustOutputs = if (context.assignGrads.isDefined) {
outputNodes.map(n => {
val matchNode = context.assignGrads.get.filter(_._2 == n.element.getName())
require(matchNode.size <= 1, "Invalid gradients output")
if (matchNode.size == 1) {
new AssignGrad[T](context(matchNode.head._1)._2).inputs(n)
} else {
n
}
})
} else {
outputNodes
}
Graph.dynamic(inputNodes.toArray, adjustOutputs.toArray,
Some((weights.toArray, gradients.toArray)),
generatedBackward)
}
/**
* Extract one module and the corresponding node list from the given graph
* @param graph
* @return
*/
private[bigdl] def extract[T: ClassTag](graph: DirectedGraph[NodeDef],
context: Context[T], byteOrder: ByteOrder)(
implicit ev: TensorNumeric[T]): Option[(
AbstractModule[Activity, Activity, T],
List[Node[NodeDef]],
Seq[Node[NodeDef]]
)] = {
var i = 0
while(i < patterns.length) {
val (result, inputs) = matchGraph(graph, patterns(i).topology)
if (result.size != 0) {
// get model
return Some(patterns(i).layer(graph, context, byteOrder), result, inputs)
}
i += 1
}
None
}
private def matchGraph(graph: DirectedGraph[NodeDef], pattern: DirectedGraph[String])
: (List[Node[NodeDef]], Seq[Node[NodeDef]]) = {
require(graph.reverse && pattern.reverse, "Must pass in reversed graph")
val patternToGraph = new mutable.HashMap[Node[String], Node[NodeDef]]()
val inputs = new ArrayBuffer[Node[NodeDef]]()
patternToGraph(pattern.source) = graph.source
inputs.append(graph.source)
pattern.BFS.foreach(patternNode => {
if (patternNode.element != N_INPUT_PLACEHOLDER && patternNode.element != INPUT_PLACEHOLDER) {
// Normal operation node
if (patternToGraph.get(patternNode).isEmpty) return (util.Collections.emptyList(), Seq())
val graphNode = patternToGraph(patternNode)
// Operation type should match
if (patternNode.element != graphNode.element.getOp) return (
util.Collections.emptyList(), Seq())
// Prev nodes number should be same except for the Ninput case
val patternInputLength = patternNode.prevNodes.length
val graphInputLength = graphNode.prevNodes.
filterNot(_.element.getOp == "DependencyNode").length
if (patternInputLength != graphInputLength &&
!patternNode.prevNodes.exists(_.element == N_INPUT_PLACEHOLDER)) {
return (util.Collections.emptyList(), Seq())
}
var i = 0
var direction = 0
var j = 0
while (i < patternNode.prevNodes.length) {
if (patternNode.prevNodes(i).element == N_INPUT_PLACEHOLDER) {
require(patternNode.prevNodes.count(_.element == N_INPUT_PLACEHOLDER) == 1,
s"only support one $N_INPUT_PLACEHOLDER ")
direction = 1
// skip the left input nodes of graphNode,
// once we find the placeholder, we start from another side
if (!inputs.contains(graphNode)) {
inputs.append(graphNode)
}
} else if (patternNode.prevNodes(i).element == INPUT_PLACEHOLDER) {
// skip input placeholder
if (!inputs.contains(graphNode)) {
inputs.append(graphNode)
}
} else {
val posPattern = { if (direction == 0) i else patternNode.prevNodes.length - 1 - j}
val posGraph = { if (direction == 0) i else graphNode.prevNodes.length - 1 - j}
val pn = patternNode.prevNodes(posPattern)
val gn = graphNode.prevNodes(posGraph)
if (patternToGraph.contains(pn)) {
if (!patternToGraph(pn).eq(gn)) return (util.Collections.emptyList(), Seq())
} else {
patternToGraph(pn) = gn
}
if (direction == 1) j += 1
}
i += 1
}
}
})
import scala.collection.JavaConverters._
return (patternToGraph.valuesIterator.toList.asJava, inputs)
}
private def findCommonPrefix(data: Seq[String]): String = {
if (data.length == 0) return ""
var shortest = data(0).length
data.foreach(s => if (s.length < shortest) shortest = s.length)
var prefix = ""
var i = 0
while(i < shortest) {
var c = data(0).charAt(i)
data.foreach(s => if (c != s.charAt(i)) return removeLast(prefix))
prefix += c
i += 1
}
return removeLast(prefix)
}
private def removeLast(s: String): String = {
if (s.length == 0) return s
if (s.charAt(s.length - 1) == '/') s.substring(0, s.length - 1) else s
}
private def removeColon(s: String): String = {
s.replaceAll(":", "")
}
}
