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.
com.intel.analytics.bigdl.python.api.PythonBigDL.scala Maven / Gradle / Ivy
/*
* 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.python.api
import java.util.{ArrayList => JArrayList, HashMap => JHashMap, List => JList, Map => JMap}
import com.intel.analytics.bigdl._
import com.intel.analytics.bigdl.dataset.{Identity => DIdentity, Sample => JSample, _}
import com.intel.analytics.bigdl.nn.{PGCriterion, Zeros, _}
import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, _}
import com.intel.analytics.bigdl.numeric._
import com.intel.analytics.bigdl.optim.{Optimizer, _}
import com.intel.analytics.bigdl.tensor.{DenseType, SparseType, Storage, Tensor}
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric
import com.intel.analytics.bigdl.utils.{Table, _}
import com.intel.analytics.bigdl.visualization.{Summary, TrainSummary, ValidationSummary}
import org.apache.spark.api.java.{JavaRDD, JavaSparkContext}
import org.apache.spark.rdd.RDD
import java.lang.{Boolean => JBoolean}
import java.nio.ByteOrder
import com.intel.analytics.bigdl.dlframes._
import com.intel.analytics.bigdl.nn.Graph._
import com.intel.analytics.bigdl.nn.keras.{KerasLayer, KerasModel}
import com.intel.analytics.bigdl.optim.SGD.{LearningRateSchedule, SequentialSchedule}
import com.intel.analytics.bigdl.transform.vision.image._
import com.intel.analytics.bigdl.transform.vision.image.augmentation._
import com.intel.analytics.bigdl.transform.vision.image.label.roi._
import com.intel.analytics.bigdl.transform.vision.image.opencv.OpenCVMat
import com.intel.analytics.bigdl.utils.tf.TensorflowDataFormat
import com.intel.analytics.bigdl.utils.tf.TensorflowLoader.parse
import com.intel.analytics.bigdl.utils.tf._
import org.apache.spark.sql.{DataFrame, SQLContext}
import org.apache.log4j._
import org.opencv.imgproc.Imgproc
import scala.collection.JavaConverters._
import scala.collection.mutable.ArrayBuffer
import scala.language.existentials
import scala.reflect.ClassTag
/**
* [[com.intel.analytics.bigdl.dataset.Sample]] for python.
* @param features features
* @param labels labels
* @param bigdlType bigdl numeric type
*/
case class Sample(features: JList[JTensor],
labels: JList[JTensor],
bigdlType: String)
case class JTensor(storage: Array[Float], shape: Array[Int],
bigdlType: String, indices: Array[Array[Int]] = null)
case class JActivity(value: Activity)
/**
* [[ValidationResult]] for python
* @param result result
* @param totalNum total number
* @param method method name
*/
case class EvaluatedResult(val result: Float, totalNum: Int, method: String)
object PythonBigDL {
def ofFloat(): PythonBigDL[Float] = new PythonBigDL[Float]()
def ofDouble(): PythonBigDL[Double] = new PythonBigDL[Double]()
}
/**
* Implementation of Python API for BigDL
*/
class PythonBigDL[T: ClassTag](implicit ev: TensorNumeric[T]) extends Serializable {
private val typeName = {
val cls = implicitly[ClassTag[T]].runtimeClass
cls.getSimpleName
}
private def toTable(input: JList[JTensor]): Table = {
input.asScala.foldLeft(new Table())((t, jtensor) => t.insert(toTensor(jtensor)))
}
def jTensorsToActivity(input: JList[JTensor], isTable: Boolean): Activity = {
if (input.isEmpty) {
throw new IllegalArgumentException("Empty input")
}
if (isTable) {
toTable(input)
} else {
toTensor(input.iterator().next())
}
}
def activityToJTensors(outputActivity: Activity): JList[JTensor] = {
if (outputActivity.isInstanceOf[Tensor[T]]) {
List(toJTensor(outputActivity.toTensor)).asJava
} else {
outputActivity.toTable.getState().toList.map {
pair => (pair._1.asInstanceOf[Int], toJTensor(pair._2.asInstanceOf[Tensor[T]]))
}.sortWith(_._1 < _._1).map(pair => pair._2).asJava
}
}
def toPySample(sample: JSample[T]): Sample = {
val cls = implicitly[ClassTag[T]].runtimeClass
val features = new JArrayList[JTensor]()
features.add(toJTensor(sample.feature()))
val labels = new JArrayList[JTensor]()
labels.add(toJTensor(sample.label()))
Sample(features, labels, cls.getSimpleName)
}
def toTensor(jTensor: JTensor): Tensor[T] = {
if (jTensor == null) return null
this.typeName match {
case "float" =>
if (null == jTensor.indices) {
if (jTensor.shape == null || jTensor.shape.length == 0) {
Tensor()
} else {
Tensor(jTensor.storage.map(x => ev.fromType(x)), jTensor.shape)
}
} else {
Tensor.sparse(jTensor.indices, jTensor.storage.map(x => ev.fromType(x)), jTensor.shape)
}
case "double" =>
if (null == jTensor.indices) {
if (jTensor.shape == null || jTensor.shape.length == 0) {
Tensor()
} else {
Tensor(jTensor.storage.map(x => ev.fromType(x.toDouble)), jTensor.shape)
}
} else {
Tensor.sparse(jTensor.indices,
jTensor.storage.map(x => ev.fromType(x.toDouble)), jTensor.shape)
}
case t: String =>
throw new IllegalArgumentException(s"Not supported type: ${t}")
}
}
def toJTensor(tensor: Tensor[T]): JTensor = {
// clone here in case the the size of storage larger then the size of tensor.
require(tensor != null, "tensor cannot be null")
tensor.getTensorType match {
case SparseType =>
// Note: as SparseTensor's indices is inaccessible here,
// so we will transfer it to DenseTensor. Just for testing.
if (tensor.nElement() == 0) {
JTensor(Array(), Array(0), bigdlType = typeName)
} else {
val cloneTensor = Tensor.dense(tensor)
val result = JTensor(cloneTensor.storage().array().map(i => ev.toType[Float](i)),
cloneTensor.size(), bigdlType = typeName)
result
}
case DenseType =>
if (tensor.nElement() == 0) {
if (tensor.dim() == 0) {
JTensor(null, null, bigdlType = typeName)
} else {
JTensor(Array(), tensor.size(), bigdlType = typeName)
}
} else {
val cloneTensor = tensor.clone()
val result = JTensor(cloneTensor.storage().array().map(i => ev.toType[Float](i)),
cloneTensor.size(), bigdlType = typeName)
result
}
case _ =>
throw new IllegalArgumentException(s"toJTensor: Unsupported tensor type" +
s" ${tensor.getTensorType}")
}
}
def testTensor(jTensor: JTensor): JTensor = {
val tensor = toTensor(jTensor)
toJTensor(tensor)
}
def testSample(sample: Sample): Sample = {
val jsample = toJSample(sample)
toPySample(jsample)
}
def toJSample(record: Sample): JSample[T] = {
require(record.bigdlType == this.typeName,
s"record.bigdlType: ${record.bigdlType} == this.typeName: ${this.typeName}")
JSample[T](record.features.asScala.toArray.map(toTensor(_)),
record.labels.asScala.toArray.map(toTensor(_)))
}
def toJSample(psamples: RDD[Sample]): RDD[JSample[T]] = {
psamples.map(toJSample(_))
}
// The first dimension is batch for both X and y
def toSampleArray(Xs: List[Tensor[T]], y: Tensor[T] = null): Array[JSample[T]] = {
require(!Xs.isEmpty, "Xs should not be empty")
val totalNum = Xs(0).size()(0)
var i = 1
val samples = new Array[JSample[T]](totalNum)
if (y != null) {
require(Xs(0).size()(0) == y.size()(0),
s"The batch dim should be equal, but we got: ${Xs(0).size()(0)} vs ${y.size()(0)}")
while (i <= totalNum) {
samples(i-1) = JSample(Xs.map{X => X.select(1, i)}.toArray, y.select(1, i))
i += 1
}
} else {
val dummyTensor = Tensor[T](1).fill(ev.fromType(1))
while (i <= totalNum) {
samples(i-1) = JSample(Xs.map{X => X.select(1, i)}.toArray, dummyTensor)
i += 1
}
}
samples
}
def batching(dataset: DataSet[JSample[T]], batchSize: Int)
: DataSet[MiniBatch[T]] = {
dataset -> SampleToMiniBatch[T](batchSize)
}
private def enrichOptimizer[T](optimizer: Optimizer[T, MiniBatch[T]],
endTrigger: Trigger,
optimMethod: OptimMethod[T]): Optimizer[T, MiniBatch[T]] = {
optimizer.setEndWhen(endTrigger)
optimizer.setOptimMethod(optimMethod)
// TODO: remove this
optimizer.disableCheckSingleton()
optimizer
}
def createSequential(): Container[Activity, Activity, T] = {
Sequential[T]()
}
def createLinear(inputSize: Int, outputSize: Int,
withBias: Boolean,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null): Linear[T] = {
Linear[T](inputSize, outputSize, withBias, wRegularizer, bRegularizer,
toTensor(initWeight), toTensor(initBias), toTensor(initGradWeight), toTensor(initGradBias))
}
def createSparseLinear(inputSize: Int, outputSize: Int,
withBias: Boolean,
backwardStart: Int = -1,
backwardLength: Int = -1,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null): SparseLinear[T] = {
SparseLinear[T](inputSize, outputSize, withBias, backwardStart, backwardLength,
wRegularizer, bRegularizer, toTensor(initWeight), toTensor(initBias),
toTensor(initGradWeight), toTensor(initGradBias))
}
def createNegative(inplace: Boolean): Negative[T] = {
Negative[T](inplace)
}
def createDenseToSparse(): DenseToSparse[T] = {
DenseToSparse[T]()
}
def createReLU(ip: Boolean = false): ReLU[T] = {
ReLU[T](ip)
}
def createTanh(): Tanh[T] = {
Tanh[T]()
}
def createTimeDistributed(layer: TensorModule[T]): TimeDistributed[T] = {
TimeDistributed[T](layer)
}
def createSpatialWithinChannelLRN(size: Int = 5, alpha: Double = 1.0, beta: Double = 0.75)
: SpatialWithinChannelLRN[T] = {
SpatialWithinChannelLRN[T](size, alpha, beta)
}
def createRnnCell(inputSize: Int,
hiddenSize: Int,
activation: TensorModule[T],
isInputWithBias: Boolean = true,
isHiddenWithBias: Boolean = true,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null): RnnCell[T] = {
RnnCell[T](inputSize,
hiddenSize,
activation,
isInputWithBias,
isHiddenWithBias,
wRegularizer,
uRegularizer,
bRegularizer)
}
def createTimeDistributedMaskCriterion(critrn: TensorCriterion[T],
paddingValue: Int = 0): TimeDistributedMaskCriterion[T] = {
TimeDistributedMaskCriterion[T](critrn, paddingValue)
}
def createTimeDistributedCriterion(critrn: TensorCriterion[T],
sizeAverage: Boolean = false): TimeDistributedCriterion[T] = {
TimeDistributedCriterion[T](critrn, sizeAverage)
}
def createGRU(
inputSize: Int,
outputSize: Int,
p: Double = 0,
activation: TensorModule[T] = null,
innerActivation: TensorModule[T] = null,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null): GRU[T] = {
GRU[T](inputSize, outputSize, p, activation, innerActivation,
wRegularizer, uRegularizer, bRegularizer)
}
def createLSTM(
inputSize: Int,
hiddenSize: Int,
p: Double = 0,
activation: TensorModule[T] = null,
innerActivation: TensorModule[T] = null,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null): LSTM[T] = {
LSTM[T](inputSize, hiddenSize, p, activation, innerActivation,
wRegularizer, uRegularizer, bRegularizer)
}
def createLSTMPeephole(
inputSize: Int,
hiddenSize: Int,
p: Double = 0,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null): LSTMPeephole[T] = {
LSTMPeephole[T](inputSize, hiddenSize, p, wRegularizer, uRegularizer, bRegularizer)
}
def createRecurrent(): Recurrent[T] = {
Recurrent[T]()
}
def createRecurrentDecoder(outputLength: Int): RecurrentDecoder[T] = {
RecurrentDecoder[T](outputLength)
}
def createConvLSTMPeephole(
inputSize: Int,
outputSize: Int,
kernelI: Int,
kernelC: Int,
stride: Int = 1,
padding: Int = -1,
activation: TensorModule[T] = null,
innerActivation: TensorModule[T] = null,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
cRegularizer: Regularizer[T] = null,
withPeephole: Boolean = true): ConvLSTMPeephole[T] = {
ConvLSTMPeephole[T](inputSize, outputSize, kernelI, kernelC,
stride, padding, activation, innerActivation,
wRegularizer, uRegularizer, bRegularizer, cRegularizer, withPeephole)
}
def createConvLSTMPeephole3D(
inputSize: Int,
outputSize: Int,
kernelI: Int,
kernelC: Int,
stride: Int = 1,
padding: Int = -1,
wRegularizer: Regularizer[T] = null,
uRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
cRegularizer: Regularizer[T] = null,
withPeephole: Boolean = true): ConvLSTMPeephole3D[T] = {
ConvLSTMPeephole3D[T](inputSize, outputSize, kernelI, kernelC, stride, padding,
wRegularizer, uRegularizer, bRegularizer, cRegularizer, withPeephole)
}
def createEcho(): Echo[T] = {
Echo[T]()
}
def createLogSoftMax(): LogSoftMax[T] = {
LogSoftMax[T]()
}
def createTemporalMaxPooling(
kW: Int,
dW: Int)
: TemporalMaxPooling[T] = {
TemporalMaxPooling[T](
kW,
dW)
}
def createSpatialMaxPooling(kW: Int,
kH: Int,
dW: Int,
dH: Int,
padW: Int = 0,
padH: Int = 0,
ceilMode: Boolean = false,
format: String = "NCHW")
: SpatialMaxPooling[T] = {
val maxpooling = SpatialMaxPooling[T](kW,
kH,
dW,
dH,
padW,
padH,
format = DataFormat(format))
if (ceilMode) maxpooling.ceil()
else maxpooling
}
def createLocallyConnected2D(
nInputPlane: Int,
inputWidth: Int,
inputHeight: Int,
nOutputPlane: Int,
kernelW: Int,
kernelH: Int,
strideW: Int = 1,
strideH: Int = 1,
padW: Int = 0,
padH: Int = 0,
propagateBack: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null,
withBias: Boolean = true,
dataFormat: String = "NCHW"): LocallyConnected2D[T] = {
LocallyConnected2D[T](
nInputPlane,
inputWidth,
inputHeight,
nOutputPlane,
kernelW,
kernelH,
strideW,
strideH,
padW,
padH,
propagateBack,
wRegularizer,
bRegularizer,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias),
withBias,
DataFormat(dataFormat)
)
}
def createSpatialConvolution(nInputPlane: Int,
nOutputPlane: Int,
kernelW: Int,
kernelH: Int,
strideW: Int = 1,
strideH: Int = 1,
padW: Int = 0,
padH: Int = 0,
nGroup: Int = 1,
propagateBack: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null,
withBias: Boolean = true,
dataFormat: String = "NCHW"
)
: SpatialConvolution[T] = {
SpatialConvolution[T](nInputPlane,
nOutputPlane,
kernelW,
kernelH,
strideW,
strideH,
padW,
padH,
nGroup,
propagateBack,
wRegularizer,
bRegularizer,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias),
withBias,
DataFormat(dataFormat)
)
}
def createSpatialSeparableConvolution(
nInputChannel: Int,
nOutputChannel: Int,
depthMultiplier: Int,
kW: Int,
kH: Int,
sW: Int = 1,
sH: Int = 1,
pW: Int = 0,
pH: Int = 0,
withBias: Boolean = true,
dataFormat: String = "NCHW",
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
pRegularizer: Regularizer[T] = null
)
: SpatialSeparableConvolution[T] = {
SpatialSeparableConvolution[T](nInputChannel,
nOutputChannel,
depthMultiplier,
kW,
kH,
sW,
sH,
pW,
pH,
withBias,
DataFormat(dataFormat),
wRegularizer,
bRegularizer,
pRegularizer
)
}
def createReshape(size: JList[Int], batchMode: JBoolean = null): Reshape[T] = {
val mappedBatchMode = batchMode match {
case JBoolean.TRUE => Some(true)
case JBoolean.FALSE => Some(false)
case _ => None
}
Reshape(size.asScala.toArray, mappedBatchMode)
}
def createConcat(dimension: Int): Concat[T] = {
Concat[T](dimension)
}
def createSpatialAveragePooling(kW: Int,
kH: Int,
dW: Int = 1,
dH: Int = 1,
padW: Int = 0,
padH: Int = 0,
globalPooling: Boolean = false,
ceilMode: Boolean = false,
countIncludePad: Boolean = true,
divide: Boolean = true,
format: String = "NCHW")
: SpatialAveragePooling[T] = {
SpatialAveragePooling[T](kW, kH, dW, dH, padW, padH, globalPooling,
ceilMode, countIncludePad, divide, format = DataFormat(format))
}
def createSpatialBatchNormalization(nOutput: Int,
eps: Double = 1e-5,
momentum: Double = 0.1,
affine: Boolean = true,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null, dataFormat: String = "NCHW")
: SpatialBatchNormalization[T] = {
SpatialBatchNormalization[T](nOutput, eps, momentum, affine,
toTensor(initWeight), toTensor(initBias), toTensor(initGradWeight), toTensor(initBias),
DataFormat(dataFormat)
)
}
def createSpatialCrossMapLRN(size: Int = 5,
alpha: Double = 1.0,
beta: Double = 0.75,
k: Double = 1.0,
dataFormat: String = "NCHW")
: SpatialCrossMapLRN[T] = {
SpatialCrossMapLRN[T](size, alpha, beta, k, DataFormat(dataFormat))
}
def createDropout(initP: Double = 0.5,
inplace: Boolean = false,
scale: Boolean = true)
: Dropout[T] = {
Dropout[T](initP, inplace, scale)
}
def createGaussianDropout(rate: Double)
: GaussianDropout[T] = {
GaussianDropout[T](rate)
}
def createGaussianNoise(stddev: Double)
: GaussianNoise[T] = {
GaussianNoise[T](stddev)
}
def createView(sizes: JList[Int], num_input_dims: Int = 0): View[T] = {
View[T](sizes.asScala.toArray).setNumInputDims(num_input_dims)
}
def createAbs()
: Abs[T] = {
Abs[T]()
}
def createAdd(inputSize: Int)
: Add[T] = {
Add[T](inputSize)
}
def createAddConstant(constant_scalar: Double,
inplace: Boolean = false)
: AddConstant[T] = {
AddConstant[T](constant_scalar,
inplace)
}
def createBatchNormalization(nOutput: Int,
eps: Double = 1e-5,
momentum: Double = 0.1,
affine: Boolean = true,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null)
: BatchNormalization[T] = {
BatchNormalization[T](nOutput,
eps,
momentum,
affine,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias))
}
def createBilinear(inputSize1: Int,
inputSize2: Int,
outputSize: Int,
biasRes: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: Bilinear[T] = {
Bilinear[T](inputSize1,
inputSize2,
outputSize,
biasRes,
wRegularizer,
bRegularizer)
}
def createBottle(module: AbstractModule[Activity, Activity, T],
nInputDim: Int = 2,
nOutputDim1: Int = Int.MaxValue)
: Bottle[T] = {
Bottle[T](module,
nInputDim,
nOutputDim1)
}
def createCAdd(size: JList[Int],
bRegularizer: Regularizer[T] = null)
: CAdd[T] = {
CAdd[T](size.asScala.toArray, bRegularizer)
}
def createCAddTable(inplace: Boolean = false)
: CAddTable[T, T] = {
CAddTable[T](inplace)
}
def createCAveTable(inplace: Boolean = false)
: CAveTable[T] = {
CAveTable[T](inplace)
}
def createCDivTable()
: CDivTable[T] = {
CDivTable[T]()
}
def createCMaxTable()
: CMaxTable[T] = {
CMaxTable[T]()
}
def createCMinTable()
: CMinTable[T] = {
CMinTable[T]()
}
def createCMul(size: JList[Int],
wRegularizer: Regularizer[T] = null)
: CMul[T] = {
CMul[T](size.asScala.toArray, wRegularizer)
}
def createCMulTable()
: CMulTable[T] = {
CMulTable[T]()
}
def createCSubTable()
: CSubTable[T] = {
CSubTable[T]()
}
def createClamp(min: Int,
max: Int)
: Clamp[T] = {
Clamp[T](min,
max)
}
def createContiguous()
: Contiguous[T] = {
Contiguous[T]()
}
def createCosine(inputSize: Int,
outputSize: Int)
: Cosine[T] = {
Cosine[T](inputSize,
outputSize)
}
def createCosineDistance()
: CosineDistance[T] = {
CosineDistance[T]()
}
def createCosineDistanceCriterion(sizeAverage: Boolean = true)
: CosineDistanceCriterion[T] = {
CosineDistanceCriterion[T](sizeAverage)
}
def createCrossProduct(numTensor: Int = 0,
embeddingSize: Int = 0)
: CrossProduct[T] = {
CrossProduct[T](numTensor, embeddingSize)
}
def createDiceCoefficientCriterion(sizeAverage: Boolean = true,
epsilon: Float = 1.0f)
: DiceCoefficientCriterion[T] = {
DiceCoefficientCriterion[T](sizeAverage, epsilon)
}
def createDotProduct()
: DotProduct[T] = {
DotProduct[T]()
}
def createELU(alpha: Double = 1.0,
inplace: Boolean = false)
: ELU[T] = {
ELU[T](alpha,
inplace)
}
def createEuclidean(inputSize: Int,
outputSize: Int,
fastBackward: Boolean = true)
: Euclidean[T] = {
Euclidean[T](inputSize,
outputSize,
fastBackward)
}
def createExp()
: Exp[T] = {
Exp[T]()
}
def createFlattenTable()
: FlattenTable[T] = {
FlattenTable[T]()
}
def createGradientReversal(lambda: Double = 1)
: GradientReversal[T] = {
GradientReversal[T](lambda)
}
def createHardShrink(lambda: Double = 0.5)
: HardShrink[T] = {
HardShrink[T](lambda)
}
def createHardTanh(minValue: Double = -1,
maxValue: Double = 1,
inplace: Boolean = false)
: HardTanh[T] = {
HardTanh[T](minValue,
maxValue,
inplace)
}
def createIndex(dimension: Int)
: Index[T] = {
Index[T](dimension)
}
def createInferReshape(size: JList[Int], batchMode: Boolean = false)
: InferReshape[T] = {
InferReshape[T](size.asScala.toArray,
batchMode)
}
def createJoinTable(dimension: Int,
nInputDims: Int)
: JoinTable[T] = {
JoinTable[T](dimension,
nInputDims)
}
def createSparseJoinTable(dimension: Int): SparseJoinTable[T] = {
SparseJoinTable[T](dimension)
}
def createL1Cost()
: L1Cost[T] = {
L1Cost[T]()
}
def createUpSampling1D(length: Int): UpSampling1D[T] = {
UpSampling1D(length)
}
def createUpSampling2D(size: JList[Int], dataFormat: String): UpSampling2D[T] = {
UpSampling2D(size.asScala.toArray, DataFormat(dataFormat))
}
def createL1Penalty(l1weight: Int,
sizeAverage: Boolean = false,
provideOutput: Boolean = true)
: L1Penalty[T] = {
L1Penalty[T](l1weight,
sizeAverage,
provideOutput)
}
def createNegativeEntropyPenalty(beta: Double): NegativeEntropyPenalty[T] = {
NegativeEntropyPenalty(beta)
}
def createLeakyReLU(negval: Double = 0.01,
inplace: Boolean = false)
: LeakyReLU[T] = {
LeakyReLU[T](negval,
inplace)
}
def createLog()
: Log[T] = {
Log[T]()
}
def createLogSigmoid()
: LogSigmoid[T] = {
LogSigmoid[T]()
}
def createLookupTable(nIndex: Int, nOutput: Int,
paddingValue: Double = 0, maxNorm: Double = Double.MaxValue,
normType: Double = 2.0, shouldScaleGradByFreq: Boolean = false,
wRegularizer: Regularizer[T] = null)
: LookupTable[T] = {
LookupTable[T](nIndex,
nOutput,
paddingValue,
maxNorm,
normType,
shouldScaleGradByFreq,
wRegularizer)
}
def createLookupTableSparse(nIndex: Int, nOutput: Int,
combiner: String = "sum", maxNorm: Double = -1,
wRegularizer: Regularizer[T] = null)
: LookupTableSparse[T] = {
LookupTableSparse[T](nIndex,
nOutput,
combiner,
maxNorm,
wRegularizer)
}
def createMM(transA: Boolean = false,
transB: Boolean = false)
: MM[T] = {
MM[T](transA,
transB)
}
def createMV(trans: Boolean = false)
: MV[T] = {
MV[T](trans)
}
def createMapTable(module: AbstractModule[Activity, Activity, T] = null)
: MapTable[T] = {
MapTable[T](module)
}
def createMaskedSelect()
: MaskedSelect[T] = {
MaskedSelect[T]()
}
def createMax(dim: Int = 1,
numInputDims: Int = Int.MinValue)
: Max[T] = {
Max[T](dim,
numInputDims)
}
def createMean(dimension: Int = 1,
nInputDims: Int = -1,
squeeze: Boolean = true)
: Mean[T] = {
Mean[T](dimension,
nInputDims,
squeeze)
}
def createMin(dim: Int = 1,
numInputDims: Int = Int.MinValue)
: Min[T] = {
Min[T](dim,
numInputDims)
}
def createMixtureTable(dim: Int = Int.MaxValue)
: MixtureTable[T] = {
MixtureTable[T](dim)
}
def createMul()
: Mul[T] = {
Mul[T]()
}
def createMulConstant(scalar: Double,
inplace: Boolean = false)
: MulConstant[T] = {
MulConstant[T](scalar,
inplace)
}
def createNarrow(dimension: Int,
offset: Int,
length: Int = 1)
: Narrow[T] = {
Narrow[T](dimension,
offset,
length)
}
def createNarrowTable(offset: Int,
length: Int = 1)
: NarrowTable[T] = {
NarrowTable[T](offset,
length)
}
def createNormalize(p: Double,
eps: Double = 1e-10)
: Normalize[T] = {
Normalize[T](p,
eps)
}
def createPReLU(nOutputPlane: Int = 0)
: PReLU[T] = {
PReLU[T](nOutputPlane)
}
def createSReLU(shape: JArrayList[Int], shareAxes: JArrayList[Int] = null): SReLU[T] = {
val argv: Array[Int] = if (shareAxes == null) {
null
} else {
shareAxes.asScala.toArray
}
SReLU[T](shape.asScala.toArray, argv)
}
def createActivityRegularization(l1: Double, l2: Double): ActivityRegularization[T] = {
ActivityRegularization[T](l1, l2)
}
def createPadding(dim: Int,
pad: Int,
nInputDim: Int,
value: Double = 0.0,
nIndex: Int = 1)
: Padding[T] = {
Padding[T](dim,
pad,
nInputDim,
value,
nIndex)
}
def createPairwiseDistance(norm: Int = 2)
: PairwiseDistance[T] = {
PairwiseDistance[T](norm)
}
def createParallelTable()
: ParallelTable[T] = {
ParallelTable[T]()
}
def createPower(power: Double,
scale: Double = 1,
shift: Double = 0)
: Power[T] = {
Power[T](power,
scale,
shift)
}
def createRReLU(lower: Double = 1.0 / 8,
upper: Double = 1.0 / 3,
inplace: Boolean = false)
: RReLU[T] = {
RReLU[T](lower,
upper,
inplace)
}
def createReLU6(inplace: Boolean = false)
: ReLU6[T] = {
ReLU6[T](inplace)
}
def createReplicate(nFeatures: Int,
dim: Int = 1,
nDim: Int = Int.MaxValue)
: Replicate[T] = {
Replicate[T](nFeatures,
dim,
nDim)
}
def createRoiPooling(pooled_w: Int, pooled_h: Int, spatial_scale: Double)
: RoiPooling[T] = {
RoiPooling[T](pooled_w,
pooled_h,
ev.fromType(spatial_scale))
}
def createScale(size: JList[Int])
: Scale[T] = {
Scale[T](size.asScala.toArray)
}
def createSelect(dimension: Int,
index: Int)
: Select[T] = {
Select[T](dimension,
index)
}
def createSelectTable(dimension: Int)
: SelectTable[T] = {
SelectTable[T](dimension)
}
def createSigmoid()
: Sigmoid[T] = {
Sigmoid[T]()
}
def createSoftMax()
: SoftMax[T] = {
SoftMax[T]()
}
def createSoftMin()
: SoftMin[T] = {
SoftMin[T]()
}
def createSoftPlus(beta: Double = 1.0)
: SoftPlus[T] = {
SoftPlus[T](beta)
}
def createSoftShrink(lambda: Double = 0.5)
: SoftShrink[T] = {
SoftShrink[T](lambda)
}
def createSoftSign()
: SoftSign[T] = {
SoftSign[T]()
}
def createSpatialDropout1D(
initP: Double = 0.5
): SpatialDropout1D[T] = {
SpatialDropout1D[T](initP)
}
def createSpatialDropout2D(
initP: Double = 0.5,
dataFormat: String = "NCHW"
): SpatialDropout2D[T] = {
SpatialDropout2D[T](initP, DataFormat(dataFormat))
}
def createSpatialDropout3D(
initP: Double = 0.5,
dataFormat: String = "NCHW"
): SpatialDropout3D[T] = {
SpatialDropout3D[T](initP, DataFormat(dataFormat))
}
def createSpatialDilatedConvolution(nInputPlane: Int,
nOutputPlane: Int,
kW: Int,
kH: Int,
dW: Int = 1,
dH: Int = 1,
padW: Int = 0,
padH: Int = 0,
dilationW: Int = 1,
dilationH: Int = 1,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: SpatialDilatedConvolution[T] = {
SpatialDilatedConvolution[T](nInputPlane,
nOutputPlane,
kW,
kH,
dW,
dH,
padW,
padH,
dilationW,
dilationH,
wRegularizer,
bRegularizer)
}
def createTemporalConvolution(
inputFrameSize: Int,
outputFrameSize: Int,
kernelW: Int,
strideW: Int = 1,
propagateBack: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null
)
: TemporalConvolution[T] = {
TemporalConvolution[T](
inputFrameSize,
outputFrameSize,
kernelW,
strideW,
propagateBack,
wRegularizer,
bRegularizer,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias)
)
}
def createLocallyConnected1D(
nInputFrame: Int,
inputFrameSize: Int,
outputFrameSize: Int,
kernelW: Int,
strideW: Int = 1,
propagateBack: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null
)
: LocallyConnected1D[T] = {
LocallyConnected1D[T](
nInputFrame,
inputFrameSize,
outputFrameSize,
kernelW,
strideW,
propagateBack,
wRegularizer,
bRegularizer,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias)
)
}
def createBinaryTreeLSTM(
inputSize: Int,
hiddenSize: Int,
gateOutput: Boolean = true,
withGraph: Boolean = true)
: BinaryTreeLSTM[T] = {
BinaryTreeLSTM[T](
inputSize,
hiddenSize,
gateOutput,
withGraph)
}
def createVolumetricFullConvolution(nInputPlane: Int,
nOutputPlane: Int,
kT: Int,
kW: Int,
kH: Int,
dT: Int = 1,
dW: Int = 1,
dH: Int = 1,
padT: Int = 0,
padW: Int = 0,
padH: Int = 0,
adjT: Int = 0,
adjW: Int = 0,
adjH: Int = 0,
nGroup: Int = 1,
noBias: Boolean = false,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: VolumetricFullConvolution[T] = {
VolumetricFullConvolution[T](nInputPlane,
nOutputPlane,
kT,
kW,
kH,
dT,
dW,
dH,
padT,
padW,
padH,
adjT,
adjW,
adjH,
nGroup,
noBias,
wRegularizer,
bRegularizer)
}
def createSpatialFullConvolution(nInputPlane: Int,
nOutputPlane: Int,
kW: Int,
kH: Int,
dW: Int = 1,
dH: Int = 1,
padW: Int = 0,
padH: Int = 0,
adjW: Int = 0,
adjH: Int = 0,
nGroup: Int = 1,
noBias: Boolean = false,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: SpatialFullConvolution[T] = {
SpatialFullConvolution[T](nInputPlane,
nOutputPlane,
kW,
kH,
dW,
dH,
padW,
padH,
adjW,
adjH,
nGroup,
noBias,
wRegularizer,
bRegularizer)
}
def createSpatialShareConvolution(
nInputPlane: Int,
nOutputPlane: Int,
kernelW: Int,
kernelH: Int,
strideW: Int = 1,
strideH: Int = 1,
padW: Int = 0,
padH: Int = 0,
nGroup: Int = 1,
propagateBack: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null,
initWeight: JTensor = null,
initBias: JTensor = null,
initGradWeight: JTensor = null,
initGradBias: JTensor = null,
withBias: Boolean = true): SpatialShareConvolution[T] = {
SpatialShareConvolution[T](nInputPlane,
nOutputPlane,
kernelW,
kernelH,
strideW,
strideH,
padW,
padH,
nGroup,
propagateBack,
wRegularizer,
bRegularizer,
toTensor(initWeight),
toTensor(initBias),
toTensor(initGradWeight),
toTensor(initGradBias),
withBias
)
}
def createSpatialZeroPadding(padLeft: Int,
padRight: Int,
padTop: Int,
padBottom: Int)
: SpatialZeroPadding[T] = {
SpatialZeroPadding[T](padLeft,
padRight,
padTop,
padBottom)
}
def createBifurcateSplitTable(dimension: Int)
: BifurcateSplitTable[T] = {
BifurcateSplitTable[T](dimension)
}
def createSplitTable(dimension: Int,
nInputDims: Int = -1)
: SplitTable[T] = {
SplitTable[T](dimension,
nInputDims)
}
def createSqrt()
: Sqrt[T] = {
Sqrt[T]()
}
def createSquare()
: Square[T] = {
Square[T]()
}
def createSqueeze(dim: Int = Int.MinValue,
numInputDims: Int = Int.MinValue)
: Squeeze[T] = {
Squeeze[T](dim,
numInputDims)
}
def createSum(dimension: Int = 1,
nInputDims: Int = -1,
sizeAverage: Boolean = false,
squeeze: Boolean = true
)
: Sum[T] = {
Sum[T](dimension,
nInputDims,
sizeAverage,
squeeze
)
}
def createTanhShrink()
: TanhShrink[T] = {
TanhShrink[T]()
}
def createThreshold(th: Double = 1e-6,
v: Double = 0.0,
ip: Boolean = false)
: Threshold[T] = {
Threshold[T](th,
v,
ip)
}
def createUnsqueeze(pos: Int,
numInputDims: Int = Int.MinValue)
: Unsqueeze[T] = {
Unsqueeze[T](pos,
numInputDims)
}
def createBCECriterion(weights: JTensor = null,
sizeAverage: Boolean = true)
: BCECriterion[T] = {
BCECriterion[T](if (weights == null) null else toTensor(weights),
sizeAverage)
}
def createBiRecurrent(merge: AbstractModule[Table, Tensor[T], T] = null)
: BiRecurrent[T] = {
BiRecurrent[T](merge)
}
def createConcatTable()
: ConcatTable[T] = {
ConcatTable[Activity, T]()
}
def createIdentity()
: Identity[T] = {
Identity[T]()
}
def createGaussianSampler(): GaussianSampler[T] = {
GaussianSampler[T]()
}
def createMultiLabelSoftMarginCriterion(weights: JTensor = null,
sizeAverage: Boolean = true)
: MultiLabelSoftMarginCriterion[T] = {
MultiLabelSoftMarginCriterion[T](if (weights == null) null else toTensor(weights),
sizeAverage)
}
def createMultiMarginCriterion(p: Int = 1,
weights: JTensor = null,
margin: Double = 1.0,
sizeAverage: Boolean = true)
: MultiMarginCriterion[T] = {
MultiMarginCriterion[T](p,
if (weights == null) null else toTensor(weights),
margin,
sizeAverage)
}
def createReverse(dimension: Int = 1, isInplace: Boolean = false)
: Reverse[T] = {
Reverse[T](dimension, isInplace)
}
def createTranspose(permutations: JList[JList[Int]])
: Transpose[T] = {
Transpose[T](permutations.asScala.toArray.map { item =>
val itemArray = item.asScala.toArray
(itemArray(0), itemArray(1))
})
}
def createSpatialContrastiveNormalization(nInputPlane: Int = 1,
kernel: JTensor = null,
threshold: Double = 1e-4,
thresval: Double = 1e-4)
: SpatialContrastiveNormalization[T] = {
SpatialContrastiveNormalization[T](nInputPlane,
if (kernel == null) null else toTensor(kernel),
threshold,
thresval)
}
def createSpatialConvolutionMap(connTable: JTensor,
kW: Int,
kH: Int,
dW: Int = 1,
dH: Int = 1,
padW: Int = 0,
padH: Int = 0,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: SpatialConvolutionMap[T] = {
SpatialConvolutionMap[T](if (connTable == null) null else toTensor(connTable),
kW,
kH,
dW,
dH,
padW,
padH,
wRegularizer,
bRegularizer)
}
def createVolumetricConvolution(nInputPlane: Int,
nOutputPlane: Int,
kT: Int,
kW: Int,
kH: Int,
dT: Int = 1,
dW: Int = 1,
dH: Int = 1,
padT: Int = 0,
padW: Int = 0,
padH: Int = 0,
withBias: Boolean = true,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null)
: VolumetricConvolution[T] = {
VolumetricConvolution[T](nInputPlane,
nOutputPlane,
kT,
kW,
kH,
dT,
dW,
dH,
padT,
padW,
padH,
withBias,
wRegularizer,
bRegularizer)
}
def createVolumetricMaxPooling(kT: Int,
kW: Int,
kH: Int,
dT: Int,
dW: Int,
dH: Int,
padT: Int = 0,
padW: Int = 0,
padH: Int = 0): VolumetricMaxPooling[T] = {
VolumetricMaxPooling[T](kT, kW, kH, dT, dW, dH, padT, padW, padH)
}
def createVolumetricAveragePooling(kT: Int,
kW: Int,
kH: Int,
dT: Int,
dW: Int,
dH: Int,
padT: Int = 0,
padW: Int = 0,
padH: Int = 0,
countIncludePad: Boolean = true,
ceilMode: Boolean = false):
VolumetricAveragePooling[T] = {
VolumetricAveragePooling[T](kT, kW, kH, dT, dW, dH, padT, padW, padH, countIncludePad, ceilMode)
}
def createSpatialDivisiveNormalization(nInputPlane: Int = 1,
kernel: JTensor = null,
threshold: Double = 1e-4,
thresval: Double = 1e-4)
: SpatialDivisiveNormalization[T] = {
SpatialDivisiveNormalization[T](nInputPlane,
if (kernel == null) null else toTensor(kernel),
threshold,
thresval)
}
def createSpatialSubtractiveNormalization(nInputPlane: Int = 1,
kernel: JTensor = null)
: SpatialSubtractiveNormalization[T] = {
SpatialSubtractiveNormalization[T](nInputPlane,
if (kernel == null) null else toTensor(kernel))
}
def createSoftMarginCriterion(sizeAverage: Boolean = true)
: SoftMarginCriterion[T] = {
SoftMarginCriterion[T](sizeAverage)
}
def createCategoricalCrossEntropy(): CategoricalCrossEntropy[T] = {
CategoricalCrossEntropy[T]()
}
// Optimizer
def createPoly(power: Double, maxIteration: Int): SGD.Poly = {
SGD.Poly(power, maxIteration)
}
def createStep(stepSize: Int, gamma: Double): SGD.Step = {
SGD.Step(stepSize, gamma)
}
def createMultiStep(stepSizes: JList[Int], gamma: Double): SGD.MultiStep = {
SGD.MultiStep(stepSizes.asScala.toArray, gamma)
}
def createExponential(decayStep: Int, decayRate: Double,
stairCase: Boolean = false): SGD.Exponential = {
SGD.Exponential(decayStep, decayRate, stairCase)
}
def createDefault(): SGD.Default = {
SGD.Default()
}
def createPlateau(monitor: String, factor: Float = 0.1f,
patience: Int = 10, mode: String = "min", epsilon: Float = 1e-4f,
cooldown: Int = 0, minLr: Float = 0): SGD.Plateau = {
SGD.Plateau(monitor, factor, patience, mode, epsilon, cooldown, minLr)
}
def createWarmup(delta: Double): SGD.Warmup = {
SGD.Warmup(delta)
}
def createSequentialSchedule(iterationPerEpoch: Int): SGD.SequentialSchedule = {
SGD.SequentialSchedule(iterationPerEpoch)
}
def createClassNLLCriterion(weights: JTensor = null,
sizeAverage: Boolean = true, logProbAsInput: Boolean = true)
: ClassNLLCriterion[T] = {
ClassNLLCriterion[T](if (weights == null) null else toTensor(weights),
sizeAverage, logProbAsInput)
}
def createMSECriterion: MSECriterion[T] = {
MSECriterion[T]()
}
def createAbsCriterion(sizeAverage: Boolean = true)
: AbsCriterion[T] = {
AbsCriterion[T](sizeAverage)
}
def createClassSimplexCriterion(nClasses: Int)
: ClassSimplexCriterion[T] = {
ClassSimplexCriterion[T](nClasses)
}
def createCrossEntropyCriterion(weights: JTensor = null,
sizeAverage: Boolean = true): CrossEntropyCriterion[T] = {
new CrossEntropyCriterion[T](if (null == weights) null else toTensor(weights), sizeAverage)
}
def createCosineEmbeddingCriterion(margin: Double = 0.0,
sizeAverage: Boolean = true)
: CosineEmbeddingCriterion[T] = {
CosineEmbeddingCriterion[T](margin,
sizeAverage)
}
def createDistKLDivCriterion(sizeAverage: Boolean = true)
: DistKLDivCriterion[T] = {
DistKLDivCriterion[T](sizeAverage)
}
def createHingeEmbeddingCriterion(margin: Double = 1,
sizeAverage: Boolean = true)
: HingeEmbeddingCriterion[T] = {
HingeEmbeddingCriterion[T](margin,
sizeAverage)
}
def createL1HingeEmbeddingCriterion(margin: Double = 1)
: L1HingeEmbeddingCriterion[T] = {
L1HingeEmbeddingCriterion[T](margin)
}
def createMarginCriterion(margin: Double = 1.0,
sizeAverage: Boolean = true, squared: Boolean = false)
: MarginCriterion[T] = {
MarginCriterion[T](margin,
sizeAverage, squared)
}
def createMarginRankingCriterion(margin: Double = 1.0,
sizeAverage: Boolean = true)
: MarginRankingCriterion[T] = {
MarginRankingCriterion[T](margin,
sizeAverage)
}
def createMultiCriterion()
: MultiCriterion[T] = {
MultiCriterion[T]()
}
def createMultiLabelMarginCriterion(sizeAverage: Boolean = true)
: MultiLabelMarginCriterion[T] = {
MultiLabelMarginCriterion[T](sizeAverage)
}
def createParallelCriterion(repeatTarget: Boolean = false)
: ParallelCriterion[T] = {
ParallelCriterion[T](repeatTarget)
}
def createKLDCriterion(sizeAverage: Boolean): KLDCriterion[T] = {
KLDCriterion[T](sizeAverage)
}
def createGaussianCriterion(): GaussianCriterion[T] = {
GaussianCriterion[T]()
}
def createSmoothL1Criterion(sizeAverage: Boolean = true)
: SmoothL1Criterion[T] = {
SmoothL1Criterion[T](sizeAverage)
}
def createSmoothL1CriterionWithWeights(sigma: Double, num: Int = 0)
: SmoothL1CriterionWithWeights[T] = {
SmoothL1CriterionWithWeights[T](sigma,
num)
}
def createSoftmaxWithCriterion(ignoreLabel: Integer = null,
normalizeMode: String = "VALID")
: SoftmaxWithCriterion[T] = {
val normM = normalizeMode match {
case "FULL" => NormMode.FULL
case "VALID" => NormMode.VALID
case "BATCH_SIZE" => NormMode.BATCH_SIZE
case "NONE" => NormMode.NONE
case n: String =>
throw new IllegalArgumentException(s"Only support 'FULL', " +
s"'VALID', 'BATCH_SIZE' and 'NONE': $n")
}
val labelToIgnore = ignoreLabel match {
case i: Integer => Some(i.toInt)
case null => None
}
SoftmaxWithCriterion[T](labelToIgnore, normM)
}
def createTransformerCriterion(
criterion: AbstractCriterion[Activity, Activity, T],
inputTransformer: AbstractModule[Activity, Activity, T] = null,
targetTransformer: AbstractModule[Activity, Activity, T] = null
): TransformerCriterion[T] = {
TransformerCriterion(criterion, Option(inputTransformer), Option(targetTransformer))
}
def createDotProductCriterion(
sizeAverage: Boolean = false): DotProductCriterion[T] = {
DotProductCriterion[T](sizeAverage)
}
def createPGCriterion(
sizeAverage: Boolean = false): PGCriterion[T] = {
PGCriterion(sizeAverage)
}
def createPack(dimension: Int): Pack[T] = {
Pack(dimension)
}
def createTile(dim : Int, copies : Int): Tile[T] = {
Tile(dim, copies)
}
def createBinaryThreshold(th: Double, ip: Boolean): BinaryThreshold[T] = {
BinaryThreshold(th, ip)
}
def setModelSeed(seed: Long): Unit = {
RandomGenerator.RNG.setSeed(seed)
}
def modelEvaluate(model: AbstractModule[Activity, Activity, T],
valRDD: JavaRDD[Sample],
batchSize: Int,
valMethods: JList[ValidationMethod[T]])
: JList[EvaluatedResult] = {
val resultArray = model.evaluate(valRDD.rdd.map(toJSample(_)),
valMethods.asScala.toArray, Some(batchSize))
val testResultArray = resultArray.map { result =>
EvaluatedResult(result._1.result()._1, result._1.result()._2,
result._2.toString())
}
testResultArray.toList.asJava
}
def loadBigDL(path: String): AbstractModule[Activity, Activity, T] = {
Module.load[T](path)
}
def loadBigDLModule(modulePath: String,
weightPath : String): AbstractModule[Activity, Activity, T] = {
Module.loadModule[T](modulePath, weightPath)
}
def loadTorch(path: String): AbstractModule[Activity, Activity, T] = {
Module.loadTorch[T](path)
}
def loadCaffe(model: AbstractModule[Activity, Activity, T],
defPath: String,
modelPath: String,
matchAll: Boolean = true): AbstractModule[Activity, Activity, T] = {
Module.loadCaffe[T](model, defPath, modelPath, matchAll)
}
def loadCaffeModel(defPath: String, modelPath: String): AbstractModule[Activity, Activity, T] = {
Module.loadCaffeModel[T](defPath, modelPath)
}
def loadTF(path: String, inputs: JList[String], outputs: JList[String],
byteOrder: String, binFile: String = null): AbstractModule[Activity, Activity, T] = {
val order = byteOrder match {
case "little_endian" => ByteOrder.LITTLE_ENDIAN
case "big_endian" => ByteOrder.BIG_ENDIAN
case _ => throw new IllegalArgumentException(s"No support byte order $byteOrder")
}
Module.loadTF[T](path, inputs.asScala, outputs.asScala, order, Option(binFile))
}
def saveTF(model: AbstractModule[Activity, Activity, T],
inputs: JList[Any],
path: String,
byteOrder: String,
dataFormat: String): Unit = {
val order = byteOrder.toLowerCase match {
case "little_endian" => ByteOrder.LITTLE_ENDIAN
case "big_endian" => ByteOrder.BIG_ENDIAN
case _ => throw new IllegalArgumentException(s"Unknown byte order $byteOrder")
}
val format = dataFormat.toLowerCase match {
case "nhwc" => TensorflowDataFormat.NHWC
case "nchw" => TensorflowDataFormat.NCHW
case _ => throw new IllegalArgumentException(s"Unknown format $dataFormat")
}
val scalaInputs = inputs.asScala.map { elem =>
val array = elem.asInstanceOf[JList[Any]]
val name = array.get(0).asInstanceOf[String]
val shape = array.get(1).asInstanceOf[JList[Int]]
(name, shape.asScala)
}
model.saveTF(scalaInputs, path, order, format)
}
def predictLocal(model: AbstractModule[Activity, Activity, T],
features: JList[JTensor]): JList[JTensor] = {
val sampleArray = toSampleArray(features.asScala.toList.map{f => toTensor(f)})
val localPredictor = LocalPredictor(model)
val result = localPredictor.predict(sampleArray)
result.map{a => toJTensor(a.asInstanceOf[Tensor[T]])}.toList.asJava
}
def predictLocalClass(model: AbstractModule[Activity, Activity, T],
features: JList[JTensor]): JList[Int] = {
val sampleArray = toSampleArray(features.asScala.toList.map{f => toTensor(f)})
val localPredictor = LocalPredictor(model)
val result = localPredictor.predictClass(sampleArray)
result.toList.asJava
}
def modelPredictRDD(model: AbstractModule[Activity, Activity, T],
dataRdd: JavaRDD[Sample]): JavaRDD[JTensor] = {
val tensorRDD = model.predict(dataRdd.rdd.map(toJSample(_)))
val listRDD = tensorRDD.map { res =>
val tensor = res.asInstanceOf[Tensor[T]]
val cloneTensor = tensor.clone()
toJTensor(cloneTensor)
}
new JavaRDD[JTensor](listRDD)
}
def modelPredictImage(model: AbstractModule[Activity, Activity, T],
imageFrame: ImageFrame,
featLayerName: String,
shareBuffer: Boolean,
batchPerPartition: Int,
predictKey: String)
: ImageFrame = {
model.predictImage(imageFrame,
featLayerName, shareBuffer, batchPerPartition, predictKey)
}
def evaluate(module: AbstractModule[Activity, Activity, T]):
AbstractModule[Activity, Activity, T] = {
module.evaluate()
}
def modelPredictClass(model: AbstractModule[Activity, Activity, T],
dataRdd: JavaRDD[Sample]): JavaRDD[Int] = {
val sampleRdd = toJSample(dataRdd)
val tensorRDD = model.predictClass(sampleRdd)
new JavaRDD[Int](tensorRDD)
}
def modelForward(model: AbstractModule[Activity, Activity, T],
input: JList[JTensor],
inputIsTable: Boolean): JList[JTensor] = {
val inputActivity = jTensorsToActivity(input, inputIsTable)
val outputActivity = model.forward(inputActivity)
activityToJTensors(outputActivity)
}
def modelBackward(model: AbstractModule[Activity, Activity, T],
input: JList[JTensor],
inputIsTable: Boolean,
gradOutput: JList[JTensor],
gradOutputIsTable: Boolean): JList[JTensor] = {
val inputActivity = jTensorsToActivity(input, inputIsTable)
val gradOutputActivity = jTensorsToActivity(gradOutput, gradOutputIsTable)
val outputActivity = model.backward(inputActivity, gradOutputActivity)
activityToJTensors(outputActivity)
}
def modelSave(module: AbstractModule[Activity, Activity, T],
path: String, overWrite: Boolean): Unit = {
module.save(path, overWrite)
}
def saveBigDLModule(module: AbstractModule[Activity, Activity, T],
modulePath: String, weightPath: String, overWrite: Boolean): Unit = {
module.saveModule(modulePath, weightPath, overWrite)
}
def saveCaffe(module: AbstractModule[Activity, Activity, T],
prototxtPath: String, modelPath: String,
useV2: Boolean = true, overwrite: Boolean = false): Unit = {
module.saveCaffe(prototxtPath, modelPath, useV2, overwrite)
}
def criterionForward(criterion: AbstractCriterion[Activity, Activity, T],
input: JList[JTensor],
inputIsTable: Boolean,
target: JList[JTensor],
targetIsTable: Boolean): T = {
val inputActivity = jTensorsToActivity(input, inputIsTable)
val targetActivity = jTensorsToActivity(target, targetIsTable)
return criterion.forward(inputActivity, targetActivity)
}
def criterionBackward(criterion: AbstractCriterion[Activity, Activity, T],
input: JList[JTensor],
inputIsTable: Boolean,
target: JList[JTensor],
targetIsTable: Boolean): JList[JTensor] = {
val inputActivity = jTensorsToActivity(input, inputIsTable)
val targetActivity = jTensorsToActivity(target, targetIsTable)
val outputActivity = criterion.backward(inputActivity, targetActivity)
activityToJTensors(outputActivity)
}
def modelGetParameters(model: AbstractModule[Activity, Activity, T])
: JMap[Any, JMap[Any, JList[JList[Any]]]] = {
model.getParametersTable().getState().mapValues {
case name2Values: Table =>
name2Values.getState().mapValues {
case t: Tensor[T] =>
val tensorClone = t.clone()
val item = List(tensorClone.storage().toList.asJava.asInstanceOf[JList[Any]],
tensorClone.size().toList.asJava.asInstanceOf[JList[Any]]).asJava
item
}.asJava
}.asJava
}
def createMaxEpoch(max: Int): Trigger = {
Trigger.maxEpoch(max)
}
def createEveryEpoch(): Trigger = {
Trigger.everyEpoch
}
def createSeveralIteration(interval: Int): Trigger = {
Trigger.severalIteration(interval)
}
def createMaxIteration(max: Int): Trigger = {
Trigger.maxIteration(max)
}
def createMaxScore(max: Float): Trigger = {
Trigger.maxScore(max)
}
def createMinLoss(min: Float): Trigger = {
Trigger.minLoss(min)
}
def createTop1Accuracy(): ValidationMethod[T] = {
new Top1Accuracy()
}
def createTreeNNAccuracy(): ValidationMethod[T] = {
new TreeNNAccuracy()
}
def createTop5Accuracy(): ValidationMethod[T] = {
new Top5Accuracy()
}
def createLoss(criterion: Criterion[T]): ValidationMethod[T] = {
new Loss(criterion)
}
def createMAE(): ValidationMethod[T] = {
new MAE()
}
def createSGD(learningRate: Double = 1e-3,
learningRateDecay: Double = 0.0,
weightDecay: Double = 0.0,
momentum: Double = 0.0,
dampening: Double = Double.MaxValue,
nesterov: Boolean = false,
leaningRateSchedule: SGD.LearningRateSchedule = SGD.Default(),
learningRates: JTensor = null,
weightDecays: JTensor = null): SGD[T] = {
val p1 = if (learningRates == null) null else toTensor(learningRates)
val p2 = if (weightDecays == null) null else toTensor(weightDecays)
new SGD[T](learningRate, learningRateDecay, weightDecay, momentum, dampening,
nesterov, leaningRateSchedule, p1, p2)
}
def createAdagrad(learningRate: Double = 1e-3,
learningRateDecay: Double = 0.0,
weightDecay: Double = 0.0): Adagrad[T] = {
new Adagrad[T](learningRate, learningRateDecay, weightDecay)
}
def createLBFGS(maxIter: Int = 20,
maxEval: Double = Double.MaxValue,
tolFun: Double = 1e-5,
tolX: Double = 1e-9,
nCorrection: Int = 100,
learningRate: Double = 1.0,
verbose: Boolean = false,
lineSearch: LineSearch[T] = null,
lineSearchOptions: JMap[Any, Any] = null): LBFGS[T] = {
val p1 = if (lineSearch == null) None else Option(lineSearch)
val p2 = if (lineSearchOptions == null) None else Option(T(lineSearchOptions))
new LBFGS[T](maxIter, maxEval, tolFun, tolX, nCorrection, learningRate, verbose, p1, p2)
}
def createAdadelta(decayRate: Double = 0.9, Epsilon: Double = 1e-10): Adadelta[T] = {
new Adadelta[T](decayRate, Epsilon)
}
def createAdam(
learningRate: Double = 1e-3,
learningRateDecay: Double = 0.0,
beta1: Double = 0.9,
beta2: Double = 0.999,
Epsilon: Double = 1e-8): Adam[T] = {
new Adam[T](learningRate, learningRateDecay, beta1, beta2, Epsilon)
}
def createAdamax(
learningRate: Double = 0.002,
beta1: Double = 0.9,
beta2: Double = 0.999,
Epsilon: Double = 1e-38): Adamax[T] = {
new Adamax(learningRate, beta1, beta2, Epsilon)
}
def createRMSprop(
learningRate: Double = 1e-2,
learningRateDecay: Double = 0.0,
decayRate: Double = 0.99,
Epsilon: Double = 1e-8): RMSprop[T] = {
new RMSprop[T](learningRate, learningRateDecay, decayRate, Epsilon)
}
def loadOptimMethod(path: String): OptimMethod[T] = {
OptimMethod.load[T](path)
}
def saveOptimMethod(method: OptimMethod[T], path: String,
overWrite: Boolean = false): Unit = {
method.save(path, overWrite)
}
/**
* Save tensor dictionary to a Java hashmap object file
*/
def saveTensorDictionary(tensors: JHashMap[String, JTensor], path: String): Unit = {
File.save(tensors, path, true)
}
def trainTF(
modelPath: String,
output: String,
samples: JavaRDD[Sample],
optMethod: OptimMethod[T],
criterion: Criterion[T],
batchSize: Int,
endWhen: Trigger): AbstractModule[Activity, Activity, T] = {
val nodeList = parse(modelPath)
val context = new Context[T]()
val session = new BigDLSessionImpl[T](nodeList.asScala, context, ByteOrder.LITTLE_ENDIAN)
val dataset = batching(DataSet.rdd(toJSample(samples)),
batchSize).asInstanceOf[DistributedDataSet[MiniBatch[T]]]
val model = session.train(Seq(output), dataset,
optMethod, criterion, endWhen)
model
}
def createLocalOptimizer(features: JList[JTensor],
y: JTensor,
model: AbstractModule[Activity, Activity, T],
criterion: Criterion[T],
optimMethod: OptimMethod[T],
endTrigger: Trigger,
batchSize: Int,
localCores: Int): Optimizer[T, MiniBatch[T]] = {
val sampleArray = toSampleArray(features.asScala.toList.map{f => toTensor(f)}, toTensor(y))
val optimizer = new LocalOptimizer[T](
model,
batching(DataSet.array(sampleArray), batchSize)
.asInstanceOf[LocalDataSet[MiniBatch[T]]],
criterion
).asInstanceOf[Optimizer[T, MiniBatch[T]]]
Engine.setNodeAndCore(1, localCores)
enrichOptimizer(optimizer, endTrigger, optimMethod)
}
def createDistriOptimizer(model: AbstractModule[Activity, Activity, T],
trainingRdd: JavaRDD[Sample],
criterion: Criterion[T],
optimMethod: OptimMethod[T],
endTrigger: Trigger,
batchSize: Int): Optimizer[T, MiniBatch[T]] = {
val sampleRDD = toJSample(trainingRdd)
val optimizer = new DistriOptimizer(
_model = model,
_dataset = batching(DataSet.rdd(sampleRDD), batchSize)
.asInstanceOf[DistributedDataSet[MiniBatch[T]]],
_criterion = criterion
).asInstanceOf[Optimizer[T, MiniBatch[T]]]
enrichOptimizer(optimizer, endTrigger, optimMethod)
}
def createDistriOptimizerFromDataSet(model: AbstractModule[Activity, Activity, T],
trainDataSet: DataSet[ImageFeature],
criterion: Criterion[T],
optimMethod: OptimMethod[T],
endTrigger: Trigger,
batchSize: Int): Optimizer[T, MiniBatch[T]] = {
val dataSet = trainDataSet -> ImageFeatureToMiniBatch[T](batchSize)
val optimizer = new DistriOptimizer(
_model = model,
_dataset = dataSet.asInstanceOf[DistributedDataSet[MiniBatch[T]]],
_criterion = criterion
).asInstanceOf[Optimizer[T, MiniBatch[T]]]
enrichOptimizer(optimizer, endTrigger, optimMethod)
}
def featureTransformDataset(dataset: DataSet[ImageFeature],
transformer: FeatureTransformer): DataSet[ImageFeature] = {
dataset -> transformer
}
def createL1L2Regularizer(l1: Double, l2: Double): L1L2Regularizer[T] = {
L1L2Regularizer[T](l1, l2)
}
def createL1Regularizer(l1: Double): L1Regularizer[T] = {
L1Regularizer[T](l1)
}
def createL2Regularizer(l2: Double): L2Regularizer[T] = {
L2Regularizer[T](l2)
}
def setValidation(optimizer: Optimizer[T, MiniBatch[T]],
batchSize: Int,
trigger: Trigger,
valRdd: JavaRDD[Sample],
vMethods: JList[ValidationMethod[T]]): Unit = {
val sampleRDD = toJSample(valRdd)
optimizer.setValidation(trigger, batching(DataSet.rdd(sampleRDD), batchSize.toInt),
vMethods.asScala.toArray)
}
def setValidationFromDataSet(optimizer: Optimizer[T, MiniBatch[T]],
batchSize: Int,
trigger: Trigger,
valDataSet: DataSet[ImageFeature],
vMethods: JList[ValidationMethod[T]]): Unit = {
val dataSet = valDataSet -> ImageFeatureToMiniBatch[T](batchSize)
optimizer.setValidation(trigger, dataSet,
vMethods.asScala.toArray)
}
def setValidation(optimizer: Optimizer[T, MiniBatch[T]],
batchSize: Int,
trigger: Trigger,
xVal: JList[JTensor],
yVal: JTensor,
vMethods: JList[ValidationMethod[T]]): Unit = {
val sampleArray = toSampleArray(xVal.asScala.toList.map{f => toTensor(f)}, toTensor(yVal))
optimizer.setValidation(trigger, batching(DataSet.array(sampleArray), batchSize),
vMethods.asScala.toArray)
}
def setTrainData(optimizer: Optimizer[T, MiniBatch[T]],
trainingRdd: JavaRDD[Sample],
batchSize: Int): Unit = {
val sampleRDD = toJSample(trainingRdd)
optimizer.setTrainData(sampleRDD, batchSize)
}
def setCriterion(optimizer: Optimizer[T, MiniBatch[T]],
criterion: Criterion[T]): Unit = {
optimizer.setCriterion(criterion)
}
def setCheckPoint(optimizer: Optimizer[T, MiniBatch[T]],
trigger: Trigger,
checkPointPath: String,
isOverwrite: Boolean): Unit = {
optimizer.setCheckpoint(checkPointPath, trigger)
if (isOverwrite) {
optimizer.overWriteCheckpoint()
}
}
def setTrainSummary(optimizer: Optimizer[T, MiniBatch[T]], summary: TrainSummary): Unit = {
optimizer.setTrainSummary(summary)
}
def setValSummary(optimizer: Optimizer[T, MiniBatch[T]], summary: ValidationSummary): Unit = {
optimizer.setValidationSummary(summary)
}
def summaryReadScalar(summary: Summary, tag: String): JList[JList[Any]] = {
val result = summary.readScalar(tag)
result.toList.map { item =>
List(item._1, item._2, item._3).asJava.asInstanceOf[JList[Any]]
}.asJava
}
def summarySetTrigger(
summary: TrainSummary,
summaryName: String,
trigger: Trigger): TrainSummary = {
summary.setSummaryTrigger(summaryName, trigger)
summary
}
def createTrainSummary(logDir: String,
appName: String): TrainSummary = {
new TrainSummary(logDir, appName)
}
def createValidationSummary(logDir: String,
appName: String): ValidationSummary = {
new ValidationSummary(logDir, appName)
}
def createModel(input: JList[ModuleNode[T]],
output: JList[ModuleNode[T]]): Graph[T] = {
Graph(input.asScala.toArray, output.asScala.toArray)
}
def createNode(module: AbstractModule[Activity, Activity, T],
x: JList[ModuleNode[T]]): ModuleNode[T] = {
if (null == x || x.isEmpty) {
module.inputs()
} else {
module.inputs(x.asScala: _*)
}
}
def createInput(): ModuleNode[T] = {
Input()
}
def initEngine(): Unit = {
Engine.init
}
def getNodeAndCoreNumber(): Array[Int] = {
Array(Engine.nodeNumber(), Engine.coreNumber())
}
def setWeights(model: AbstractModule[Activity, Activity, T], weights: JList[JTensor]): Unit = {
val weightTensor = weights.asScala.toArray.map(toTensor(_))
model.setWeightsBias(weightTensor)
}
def getWeights(model: AbstractModule[Activity, Activity, T]): JList[JTensor] = {
val weights = model.getWeightsBias()
if (weights != null) {
weights.map(toJTensor(_)).toList.asJava
} else {
null
}
}
def updateParameters(model: AbstractModule[Activity, Activity, T], lr: Double): Unit = {
val (w, g) = model.getParameters()
w.add(ev.negative(ev.fromType(lr)), g)
}
def uniform(a: Double, b: Double, size: JList[Int]): JTensor = {
val result = Tensor[T]().resize(size.asScala.toArray)
result.apply1(i => ev.fromType(RandomGenerator.RNG.uniform(a, b)))
toJTensor(result)
}
def createZeros(): Zeros.type = {
Zeros
}
def createOnes(): Ones.type = {
Ones
}
def createConstInitMethod(value: Double): ConstInitMethod = {
ConstInitMethod(value)
}
def createRandomUniform(lower: Double, upper: Double): InitializationMethod = {
RandomUniform(lower, upper)
}
def createRandomUniform(): InitializationMethod = {
RandomUniform
}
def createRandomNormal(mean: Double, stdv: Double): RandomNormal = {
RandomNormal(mean, stdv)
}
def createXavier(): Xavier.type = {
Xavier
}
def createMsraFiller(varianceNormAverage: Boolean = true): MsraFiller = {
MsraFiller(varianceNormAverage)
}
def createBilinearFiller(): BilinearFiller.type = {
BilinearFiller
}
def createHardSigmoid : HardSigmoid[T] = {
HardSigmoid()
}
def createMeanAbsolutePercentageCriterion: MeanAbsolutePercentageCriterion[T] = {
MeanAbsolutePercentageCriterion()
}
def createMeanSquaredLogarithmicCriterion: MeanSquaredLogarithmicCriterion[T] = {
MeanSquaredLogarithmicCriterion()
}
def createKullbackLeiblerDivergenceCriterion: KullbackLeiblerDivergenceCriterion[T] = {
KullbackLeiblerDivergenceCriterion()
}
def createPoissonCriterion: PoissonCriterion[T] = {
PoissonCriterion()
}
def setInitMethod(layer: Initializable, weightInitMethod: InitializationMethod,
biasInitMethod: InitializationMethod): layer.type = {
layer.setInitMethod(weightInitMethod, biasInitMethod)
}
def setInitMethod(layer: Initializable,
initMethods: JArrayList[InitializationMethod]): layer.type = {
layer.setInitMethod(initMethods.asScala.toArray)
}
def getHiddenState(rec: Recurrent[T]): JActivity = {
JActivity(rec.getHiddenState())
}
def freeze(model: AbstractModule[Activity, Activity, T], freezeLayers: JList[String])
: AbstractModule[Activity, Activity, T] = {
if (null == freezeLayers) model.freeze() else model.freeze(freezeLayers.asScala: _*)
}
def unFreeze(model: AbstractModule[Activity, Activity, T],
names: JList[String]): AbstractModule[Activity, Activity, T] = {
if (names == null) {
model.unFreeze()
} else {
model.unFreeze(names.asScala: _*)
}
}
def setStopGradient(model: Graph[T], layers: JList[String]): Graph[T] = {
model.stopGradient(layers.asScala.toArray)
}
def saveGraphTopology(model: Graph[T], logPath: String): Graph[T] = {
model.saveGraphTopology(logPath)
}
def createResizeBilinear(
outputHeight: Int,
outputWidth: Int,
alignCorner: Boolean,
dataFormat: String
): ResizeBilinear[T] = {
ResizeBilinear[T](outputHeight,
outputWidth,
alignCorner, DataFormat.apply(dataFormat))
}
def createMultiRNNCell(cells: JList[Cell[T]]): MultiRNNCell[T] = {
MultiRNNCell(cells.asScala.toArray)
}
def createHighway(size: Int, withBias: Boolean,
activation: TensorModule[T] = null,
wRegularizer: Regularizer[T] = null,
bRegularizer: Regularizer[T] = null): Graph[T] = {
Highway(size, withBias, activation, wRegularizer, bRegularizer)
}
def createUpSampling3D(size: JList[Int]): UpSampling3D[T] = {
UpSampling3D(size.asScala.toArray)
}
def createCropping2D(
heightCrop: JList[Int],
widthCrop: JList[Int],
dataFormat: String = "NCHW"): Cropping2D[T] = {
Cropping2D(heightCrop.asScala.toArray, widthCrop.asScala.toArray, DataFormat(dataFormat))
}
def createCropping3D(
dim1Crop: JList[Int],
dim2Crop: JList[Int],
dim3Crop: JList[Int],
dataFormat: String = Cropping3D.CHANNEL_FIRST): Cropping3D[T] = {
Cropping3D(
dim1Crop.asScala.toArray, dim2Crop.asScala.toArray, dim3Crop.asScala.toArray, dataFormat)
}
def redirectSparkLogs(logPath: String): Unit = {
LoggerFilter.redirectSparkInfoLogs(logPath)
}
def showBigDlInfoLogs(): Unit = {
Logger.getLogger("com.intel.analytics.bigdl.optim").setLevel(Level.INFO)
}
def quantize(module: AbstractModule[Activity, Activity, T]): Module[T] = {
module.quantize()
}
def createDLEstimator(model: Module[T], criterion: Criterion[T],
featureSize: JArrayList[Int],
labelSize: JArrayList[Int]): DLEstimator[T] = {
new DLEstimator[T](model, criterion, featureSize.asScala.toArray, labelSize.asScala.toArray)
}
def createDLClassifier(model: Module[T], criterion: Criterion[T],
featureSize: JArrayList[Int],
labelSize: JArrayList[Int]): DLClassifier[T] = {
new DLClassifier[T](model, criterion, featureSize.asScala.toArray)
}
def fitEstimator(estimator: DLEstimator[T], dataSet: DataFrame): DLModel[T] = {
estimator.fit(dataSet)
}
def fitClassifier(classifier: DLClassifier[T], dataSet: DataFrame): DLModel[T] = {
classifier.fit(dataSet)
}
def setBatchSizeDLEstimator(estimator: DLEstimator[T], batchSize: Int): DLEstimator[T] = {
estimator.setBatchSize(batchSize)
}
def setBatchSizeDLClassifier(classifier: DLClassifier[T], batchSize: Int): DLClassifier[T] = {
classifier.setBatchSize(batchSize)
}
def setMaxEpochDLEstimator(estimator: DLEstimator[T], maxEpoch: Int): DLEstimator[T] = {
estimator.setMaxEpoch(maxEpoch)
}
def setMaxEpochDLClassifier(classifier: DLClassifier[T], maxEpoch: Int): DLClassifier[T] = {
classifier.setMaxEpoch(maxEpoch)
}
def setLearningRateDLEstimator(estimator: DLEstimator[T], lr: Double): DLEstimator[T] = {
estimator.setLearningRate(lr)
}
def setLearningRateDLClassifier(classifier: DLClassifier[T], lr: Double): DLClassifier[T] = {
classifier.setLearningRate(lr)
}
def createDLModel(model: Module[T], featureSize: JArrayList[Int]): DLModel[T] = {
new DLModel[T](model, featureSize.asScala.toArray)
}
def createDLClassifierModel(model: Module[T],
featureSize: JArrayList[Int]): DLClassifierModel[T] = {
new DLClassifierModel[T](model, featureSize.asScala.toArray)
}
def dlModelTransform(dlModel: DLModel[T], dataSet: DataFrame): DataFrame = {
dlModel.transform(dataSet)
}
def dlClassifierModelTransform(dlClassifierModel: DLClassifierModel[T],
dataSet: DataFrame): DataFrame = {
dlClassifierModel.transform(dataSet)
}
def setFeatureSizeDLModel(dlModel: DLModel[T], featureSize: JArrayList[Int]): DLModel[T] = {
dlModel.setFeatureSize(featureSize.asScala.toArray)
}
def setFeatureSizeDLClassifierModel(dlClassifierModel: DLClassifierModel[T],
featureSize: JArrayList[Int]): DLClassifierModel[T] = {
dlClassifierModel.setFeatureSize(featureSize.asScala.toArray)
}
def setBatchSizeDLModel(dlModel: DLModel[T], batchSize: Int): DLModel[T] = {
dlModel.setBatchSize(batchSize)
}
def setBatchSizeDLClassifierModel(dlClassifierModel: DLClassifierModel[T],
batchSize: Int): DLClassifierModel[T] = {
dlClassifierModel.setBatchSize(batchSize)
}
def findGraphNode(model: Graph[T], name: String): ModuleNode[T] = {
model.node(name)
}
def getContainerModules(module: Container[Activity, Activity, T])
: JList[AbstractModule[Activity, Activity, T]] = {
module match {
case m: KerasModel[T] =>
m.getSubModules().asJava
case kl: KerasLayer[Activity, Activity, T] =>
throw new RuntimeException(s"There's no sub modules for ${kl}")
case _ =>
module.modules.toList.asJava
}
}
def getFlattenModules(module: Container[Activity, Activity, T],
includeContainer: Boolean)
: JList[AbstractModule[Activity, Activity, T]] = {
val result = ArrayBuffer[AbstractModule[Activity, Activity, T]]()
doGetFlattenModules(module, includeContainer, result)
result.toList.asJava
}
// TODO: refactor Container and KerasLayer to simplify this logic
private def hasSubModules(module: AbstractModule[Activity, Activity, T]) = {
module match {
case km: KerasModel[T] => true
case kl: KerasLayer[Activity, Activity, T] => false
case c: Container[_, _, _] => true
case _ => false
}
}
private def doGetFlattenModules(module: Container[Activity, Activity, T],
includeContainer: Boolean,
result: ArrayBuffer[AbstractModule[Activity, Activity, T]]): Unit = {
getContainerModules(module).asScala.foreach {m =>
if (hasSubModules(m)) {
doGetFlattenModules(m.asInstanceOf[Container[Activity, Activity, T]],
includeContainer,
result)
} else {
result.append(m)
}
}
if (includeContainer) {
result.append(module)
}
}
def isWithWeights(module: Module[T]): Boolean = {
val weights = module.getWeightsBias()
return weights != null && !weights.isEmpty
}
def setRunningMean(module: BatchNormalization[T], runningMean: JTensor): Unit = {
module.runningMean.set(toTensor(runningMean))
}
def setRunningStd(module: BatchNormalization[T], runningStd: JTensor): Unit = {
module.runningVar.set(toTensor(runningStd))
}
def getRunningMean(module: BatchNormalization[T]): JTensor = {
toJTensor(module.runningMean)
}
def getRunningStd(module: BatchNormalization[T]): JTensor = {
toJTensor(module.runningVar)
}
def createMasking(maskValue: Double)
: Masking[T] = {
Masking[T](maskValue)
}
def createMaxout(inputSize: Int, outputSize: Int, maxoutNumber: Int, withBias: Boolean = true,
wRegularizer: Regularizer[T] = null, bRegularizer: Regularizer[T] = null,
initWeight: Tensor[T] = null, initBias: Tensor[T] = null)
: Maxout[T] = {
Maxout[T](inputSize, outputSize, maxoutNumber, withBias, wRegularizer, bRegularizer,
initWeight, initBias)
}
def createCosineProximityCriterion(): CosineProximityCriterion[T] = {
CosineProximityCriterion[T]()
}
def createPriorBox(minSizes: JList[Double], maxSizes: JList[Double] = null,
aspectRatios: JList[Double] = null, isFlip: Boolean = true, isClip: Boolean = false,
variances: JList[Double] = null, offset: Float = 0.5f,
imgH: Int = 0, imgW: Int = 0, imgSize: Int = 0,
stepH: Float = 0, stepW: Float = 0, step: Float = 0): PriorBox[T] = {
val maxS = if (maxSizes == null) null else maxSizes.asScala.toArray.map(_.toFloat)
val aspectR = if (aspectRatios == null) null else aspectRatios.asScala.toArray.map(_.toFloat)
val vars = if (variances == null) null else variances.asScala.toArray.map(_.toFloat)
new PriorBox[T](minSizes.asScala.toArray.map(_.toFloat),
maxS, aspectR, isFlip, isClip, vars, offset, imgH, imgW, imgSize, stepH, stepW, step)
}
def createNormalizeScale(p: Double, eps: Double = 1e-10, scale: Double, size: JList[Int],
wRegularizer: Regularizer[T] = null): NormalizeScale[T] =
new NormalizeScale[T](p, eps, scale, size.asScala.toArray, wRegularizer)
def createDetectionOutputSSD(nClasses: Int,
shareLocation: Boolean,
bgLabel: Int,
nmsThresh: Double,
nmsTopk: Int,
keepTopK: Int,
confThresh: Double,
varianceEncodedInTarget: Boolean,
confPostProcess: Boolean): DetectionOutputSSD[T] =
new DetectionOutputSSD[T](nClasses, shareLocation, bgLabel, nmsThresh.toFloat,
nmsTopk, keepTopK, confThresh.toFloat, varianceEncodedInTarget, confPostProcess)
def createDetectionOutputFrcnn(nmsThresh: Float = 0.3f, nClasses: Int,
bboxVote: Boolean, maxPerImage: Int = 100, thresh: Double = 0.05): DetectionOutputFrcnn = {
new DetectionOutputFrcnn(nmsThresh, nClasses, bboxVote, maxPerImage, thresh)
}
def createProposal(preNmsTopN: Int, postNmsTopN: Int,
ratios: JList[Double], scales: JList[Double],
rpnPreNmsTopNTrain: Int = 12000, rpnPostNmsTopNTrain: Int = 2000): Proposal = {
new Proposal(preNmsTopN, postNmsTopN, ratios.asScala.toArray.map(_.toFloat),
scales.asScala.toArray.map(_.toFloat), rpnPreNmsTopNTrain, rpnPostNmsTopNTrain)
}
def createHFlip(): HFlip = {
HFlip()
}
def createResize(resizeH: Int, resizeW: Int, resizeMode: Int = Imgproc.INTER_LINEAR,
useScaleFactor: Boolean): Resize = {
Resize(resizeH, resizeW, resizeMode, useScaleFactor)
}
def createColorJitter(brightnessProb: Double = 0.5, brightnessDelta: Double = 32,
contrastProb: Double = 0.5, contrastLower: Double = 0.5, contrastUpper: Double = 1.5,
hueProb: Double = 0.5, hueDelta: Double = 18,
saturationProb: Double = 0.5, saturationLower: Double = 0.5, saturationUpper: Double = 1.5,
randomOrderProb: Double = 0, shuffle: Boolean = false): ColorJitter = {
ColorJitter(brightnessProb, brightnessDelta, contrastProb,
contrastLower, contrastUpper, hueProb, hueDelta, saturationProb,
saturationLower, saturationUpper, randomOrderProb, shuffle)
}
def createBrightness(deltaLow: Double, deltaHigh: Double): Brightness = {
Brightness(deltaLow, deltaHigh)
}
def createChannelOrder(): ChannelOrder = {
ChannelOrder()
}
def createContrast(deltaLow: Double, deltaHigh: Double): Contrast = {
Contrast(deltaLow, deltaHigh)
}
def createRandomCrop(cropWidth: Int, cropHeight: Int, isClip: Boolean): RandomCrop = {
RandomCrop(cropWidth, cropHeight, isClip)
}
def createCenterCrop(cropWidth: Int, cropHeight: Int, isClip: Boolean): CenterCrop = {
CenterCrop(cropWidth, cropHeight, isClip)
}
def createFixedCrop(wStart: Double,
hStart: Double, wEnd: Double, hEnd: Double, normalized: Boolean,
isClip: Boolean): FixedCrop = {
FixedCrop(wStart.toFloat, hStart.toFloat, wEnd.toFloat, hEnd.toFloat, normalized, isClip)
}
def createDetectionCrop(roiKey: String, normalized: Boolean): DetectionCrop = {
DetectionCrop(roiKey, normalized)
}
def createExpand(meansR: Int = 123, meansG: Int = 117, meansB: Int = 104,
minExpandRatio: Double = 1.0,
maxExpandRatio: Double = 4.0): Expand = {
Expand(meansR, meansG, meansB, minExpandRatio, maxExpandRatio)
}
def createRandomAspectScale(scales: JList[Int], scaleMultipleOf: Int = 1,
maxSize: Int = 1000): RandomAspectScale = {
RandomAspectScale(scales.asScala.toArray, scaleMultipleOf, maxSize)
}
def createHue(deltaLow: Double, deltaHigh: Double): Hue = {
Hue(deltaLow, deltaHigh)
}
def createRandomTransformer(transformer: FeatureTransformer, prob: Double): RandomTransformer = {
RandomTransformer(transformer, prob)
}
def createSaturation(deltaLow: Double, deltaHigh: Double): Saturation = {
Saturation(deltaLow, deltaHigh)
}
def createRandomSampler(): FeatureTransformer = {
RandomSampler()
}
def createChannelNormalize(meanR: Double, meanG: Double, meanB: Double,
stdR: Double = 1, stdG: Double = 1, stdB: Double = 1): FeatureTransformer = {
ChannelNormalize(meanR.toFloat, meanG.toFloat, meanB.toFloat,
stdR.toFloat, stdG.toFloat, stdB.toFloat)
}
def createAspectScale(scale: Int,
scaleMultipleOf: Int,
maxSize: Int,
resizeMode: Int = 1,
useScaleFactor: Boolean = true,
minScale: Double = -1): FeatureTransformer = {
val minS = if (minScale == -1) None else Some(minScale.toFloat)
AspectScale(scale, scaleMultipleOf, maxSize, resizeMode, useScaleFactor, minS)
}
def createFiller(startX: Double, startY: Double, endX: Double, endY: Double,
value: Int = 255): Filler = {
Filler(startX.toFloat, startY.toFloat, endX.toFloat, endY.toFloat, value)
}
def createPixelNormalize(means: JList[Double]): PixelNormalizer = {
PixelNormalizer(means.asScala.toArray.map(_.toFloat))
}
def createRoiProject(needMeetCenterConstraint: Boolean): RoiProject = {
RoiProject(needMeetCenterConstraint)
}
def createRoiResize(normalized: Boolean): RoiResize = {
RoiResize(normalized)
}
def createRoiHFlip(normalized: Boolean = true): RoiHFlip = {
RoiHFlip(normalized)
}
def createRoiNormalize(): RoiNormalize = {
RoiNormalize()
}
def createFixExpand(eh: Int, ew: Int): FixExpand = {
FixExpand(eh, ew)
}
def transformImageFeature(transformer: FeatureTransformer, feature: ImageFeature)
: ImageFeature = {
transformer.transform(feature)
}
def transformImageFrame(transformer: FeatureTransformer,
imageFrame: ImageFrame): ImageFrame = {
imageFrame.transform(transformer)
}
def createDistributedImageFrame(imageRdd: JavaRDD[JTensor], labelRdd: JavaRDD[JTensor])
: DistributedImageFrame = {
require(null != imageRdd, "imageRdd cannot be null")
val featureRdd = if (null != labelRdd) {
imageRdd.rdd.zip(labelRdd.rdd).map(data => {
createImageFeature(data._1, data._2)
})
} else {
imageRdd.rdd.map(image => {
createImageFeature(image, null)
})
}
new DistributedImageFrame(featureRdd)
}
def createLocalImageFrame(images: JList[JTensor], labels: JList[JTensor])
: LocalImageFrame = {
require(null != images, "images cannot be null")
val features = if (null != labels) {
(0 until images.size()).map(i => {
createImageFeature(images.get(i), labels.get(i))
})
} else {
(0 until images.size()).map(i => {
createImageFeature(images.get(i), null)
})
}
new LocalImageFrame(features.toArray)
}
def createPipeline(list: JList[FeatureTransformer]): FeatureTransformer = {
var cur = list.get(0)
(1 until list.size()).foreach(t => cur = cur -> list.get(t))
cur
}
def createImageFeature(data: JTensor = null, label: JTensor = null, uri: String = null)
: ImageFeature = {
val feature = new ImageFeature()
if (null != data) {
val mat = OpenCVMat.fromFloats(data.storage, data.shape(0), data.shape(1), data.shape(2))
feature(ImageFeature.bytes) = OpenCVMat.imencode(mat)
feature(ImageFeature.mat) = mat
feature(ImageFeature.originalSize) = mat.shape()
}
if (null != label) {
// todo: may need a method to change label format if needed
feature(ImageFeature.label) = toTensor(label)
}
if (null != uri) {
feature(ImageFeature.uri) = uri
}
feature
}
def imageFeatureGetKeys(imageFeature: ImageFeature): JList[String] = {
imageFeature.keys().toList.asJava
}
def distributedImageFrameToImageTensorRdd(imageFrame: DistributedImageFrame,
floatKey: String = ImageFeature.floats, toChw: Boolean = true): JavaRDD[JTensor] = {
imageFrame.rdd.map(imageFeatureToImageTensor(_, floatKey, toChw)).toJavaRDD()
}
def distributedImageFrameToLabelTensorRdd(imageFrame: DistributedImageFrame): JavaRDD[JTensor] = {
imageFrame.rdd.map(imageFeatureToLabelTensor).toJavaRDD()
}
def distributedImageFrameToPredict(imageFrame: DistributedImageFrame, key: String)
: JavaRDD[JList[Any]] = {
imageFrame.rdd.map(x => {
if (x.isValid && x.contains(key)) {
List[Any](x.uri(), toJTensor(x[Tensor[T]](key))).asJava
} else {
List[Any](x.uri(), null).asJava
}
})
}
def localImageFrameToPredict(imageFrame: LocalImageFrame, key: String)
: JList[JList[Any]] = {
imageFrame.array.map(x =>
if (x.isValid && x.contains(key)) {
List[Any](x.uri(), toJTensor(x[Tensor[T]](key))).asJava
} else {
List[Any](x.uri(), null).asJava
}).toList.asJava
}
def localImageFrameToImageTensor(imageFrame: LocalImageFrame,
floatKey: String = ImageFeature.floats, toChw: Boolean = true): JList[JTensor] = {
imageFrame.array.map(imageFeatureToImageTensor(_, floatKey, toChw)).toList.asJava
}
def localImageFrameToLabelTensor(imageFrame: LocalImageFrame): JList[JTensor] = {
imageFrame.array.map(imageFeatureToLabelTensor).toList.asJava
}
def imageFeatureToImageTensor(imageFeature: ImageFeature,
floatKey: String = ImageFeature.floats, toChw: Boolean = true): JTensor = {
toJTensor(imageFeature.toTensor(floatKey, toChw).asInstanceOf[Tensor[T]])
}
def imageFeatureToLabelTensor(imageFeature: ImageFeature): JTensor = {
val label = if (imageFeature.hasLabel()) {
imageFeature.getLabel[Tensor[T]]
} else {
Tensor[T](1).fill(ev.fromType[Float](-1f))
}
toJTensor(label)
}
def read(path: String, sc: JavaSparkContext, minPartitions: Int): ImageFrame = {
if (sc == null) {
ImageFrame.read(path, null, minPartitions)
} else {
ImageFrame.read(path, sc.sc, minPartitions)
}
}
def readParquet(path: String, sqlContext: SQLContext): DistributedImageFrame = {
ImageFrame.readParquet(path, sqlContext)
}
def createBytesToMat(byteKey: String): BytesToMat = {
BytesToMat(byteKey)
}
def createPixelBytesToMat(byteKey: String): PixelBytesToMat = {
PixelBytesToMat(byteKey)
}
def createMatToFloats(validHeight: Int = 300, validWidth: Int = 300, validChannels: Int = 3,
outKey: String = ImageFeature.floats, shareBuffer: Boolean = true): MatToFloats =
new MatToFloats(validHeight, validWidth, validChannels, outKey, shareBuffer)
def createMatToTensor(toRGB: Boolean = false, tensorKey: String = ImageFeature.imageTensor)
: MatToTensor[T] = new MatToTensor[T](toRGB, tensorKey)
def isLocal(imageFrame: ImageFrame): Boolean = imageFrame.isLocal()
def isDistributed(imageFrame: ImageFrame): Boolean = imageFrame.isDistributed()
def createImageFrameToSample(inputKeys: JList[String],
targetKeys: JList[String], sampleKey: String): ImageFrameToSample[T] = {
val targets = if (targetKeys == null) null else targetKeys.asScala.toArray
ImageFrameToSample[T](inputKeys.asScala.toArray, targets, sampleKey)
}
def seqFilesToImageFrame(url: String, sc: JavaSparkContext,
classNum: Int, partitionNum: Int): ImageFrame = {
val pn = if (partitionNum <= 0) None else Some(partitionNum)
DataSet.SeqFileFolder.filesToImageFrame(url, sc, classNum, pn)
}
def setConstantClip(optimizer: Optimizer[T, MiniBatch[T]],
min: Float, max: Float): Unit = {
optimizer.setConstantGradientClipping(min, max)
}
def setL2NormClip(optimizer: Optimizer[T, MiniBatch[T]],
normValue: Float): Unit = {
optimizer.setGradientClippingByl2Norm(normValue)
}
def disableClip(optimizer: Optimizer[T, MiniBatch[T]]): Unit = {
optimizer.disableGradientClipping()
}
def addScheduler(seq: SequentialSchedule, scheduler: LearningRateSchedule,
maxIteration: Int): SequentialSchedule = {
seq.add(scheduler, maxIteration)
}
private[bigdl] def initExecutorGateway(sc: JavaSparkContext, driverPort: Int): Unit = {
sc.parallelize(Seq(""), Engine.coreNumber() * Engine.nodeNumber())
.foreachPartition(_ => Engine.createJavaGateway(driverPort))
}
def createDatasetFromImageFrame(imageFrame: ImageFrame): DataSet[ImageFeature] = {
DataSet.imageFrame(imageFrame)
}
def dlReadImage(path: String, sc: JavaSparkContext, minParitions: Int): DataFrame = {
val df = DLImageReader.readImages(path, sc.sc, minParitions)
df
}
def createDLImageTransformer(transformer: FeatureTransformer): DLImageTransformer = {
new DLImageTransformer(transformer)
}
def dlImageTransform(dlImageTransformer: DLImageTransformer, dataSet: DataFrame): DataFrame = {
dlImageTransformer.transform(dataSet)
}
def getRealClassNameOfJValue(module: AbstractModule[Activity, Activity, T]): String = {
module.getClass.getCanonicalName
}
}
object PythonBigDLUtils {
def toTensor[T: ClassTag](jTensor: JTensor, typeName: String)
(implicit ev: TensorNumeric[T]): Tensor[T] = {
if (jTensor == null) return null
typeName match {
case "float" =>
Tensor(jTensor.storage.map(x => ev.fromType(x.toFloat)), jTensor.shape)
case "double" =>
Tensor(jTensor.storage.map(x => ev.fromType(x.toDouble)), jTensor.shape)
case t: String =>
throw new IllegalArgumentException(s"Not supported type: ${t}")
}
}
}
