com.intel.analytics.bigdl.nn.UpSampling2D.scala Maven / Gradle / Ivy
The newest version!
/*
* 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.nn
import com.intel.analytics.bigdl.nn.abstractnn.{DataFormat, TensorModule}
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric
import com.intel.analytics.bigdl.utils.Shape
import scala.reflect.ClassTag
/**
* Upsampling layer for 2D inputs.
* Repeats the heights and widths of the data by size(0) and size(1) respectively.
*
* If input's dataformat is NCHW, then the size of output is (N, C, H * size(0), W * size(1))
*
* @param size tuple of 2 integers. The upsampling factors for heights and widths.
* @param format DataFormat, NCHW or NHWC
* @tparam T The numeric type in the criterion, usually which are [[Float]] or [[Double]]
*/
class UpSampling2D[T: ClassTag] (val size: Array[Int], val format: DataFormat = DataFormat.NCHW)
(implicit ev: TensorNumeric[T]) extends TensorModule[T] {
require(size.length == 2, s"UpSampling2D's size should be an array containing" +
s" 2 elements, but got ${size.mkString("x")}")
require(size(0) > 0 && size(1) > 0, "UpSampling2D's size should be bigger than 0," +
s"but got ${size.mkString("x")}")
override def computeOutputShape(inputShape: Shape): Shape = {
val input = inputShape.toSingle().toArray
require(input.length == 4,
s"UpSampling2D requires 4D input, but got input dim ${input.length}")
format match {
case DataFormat.NCHW =>
Shape(input(0), input(1), input(2)*size(0), input(3)*size(1))
case DataFormat.NHWC =>
Shape(input(0), input(1)*size(0), input(2)*size(1), input(3))
}
}
override def updateOutput(input: Tensor[T]): Tensor[T] = {
require(input.dim() == 4, "UpSampling2D only supports 4D input")
require(input.isContiguous(), "input should be contiguous")
format match {
case DataFormat.NCHW =>
UpSampling2D.updateOutputNchw(input, output, size)
case DataFormat.NHWC =>
UpSampling2D.updateOutputNhwc(input, output, size)
case _ =>
throw new IllegalArgumentException("UpSampling2D: unsupported data format.")
}
output
}
override def updateGradInput(input: Tensor[T], gradOutput: Tensor[T]): Tensor[T] = {
require(gradOutput.dim() == 4, "UpSampling2D only supports 4D gradOutput")
require(gradOutput.isContiguous(), "gradOutput should be contiguous")
gradInput.resizeAs(input).zero()
format match {
case DataFormat.NCHW =>
UpSampling2D.updateGradInputNchw(gradInput, gradOutput, size)
case DataFormat.NHWC =>
UpSampling2D.updateGradInputNhwc(gradInput, gradOutput, size)
case _ =>
throw new IllegalArgumentException("UpSampling2D: unsupported data format.")
}
gradInput
}
}
object UpSampling2D {
def apply[T: ClassTag](size: Array[Int], format: DataFormat = DataFormat.NCHW)
(implicit ev: TensorNumeric[T]): UpSampling2D[T] = {
new UpSampling2D(size, format)
}
protected def updateOutputNchw[T: ClassTag](
input: Tensor[T],
output: Tensor[T],
size: Array[Int])(implicit ev: TensorNumeric[T]) : Tensor[T] = {
val inputHeight = input.size(3)
val inputWeight = input.size(4)
val outputHeight = inputHeight * size(0)
val outputWeight = inputWeight * size(1)
output.resize(input.size(1), input.size(2), outputHeight, outputWeight)
val inputData = input.storage().array()
var inputOffset = input.storageOffset() - 1
val outputData = output.storage().array()
var outputOffset = output.storageOffset() - 1
var i = 0
while (i < input.size(1) * input.size(2)) {
var rowIndex = 0
while (rowIndex < input.size(3)) {
var columnIndex = 0
// copy column
while (columnIndex < input.size(4)) {
var colReplicate = 0
while (colReplicate < size(1)) {
outputData(outputOffset) = inputData(inputOffset)
outputOffset += 1
colReplicate += 1
}
inputOffset += 1
columnIndex += 1
}
// copy row
var rowReplicate = 1
while (rowReplicate < size(0)) {
ev.arraycopy(outputData, outputOffset - outputWeight,
outputData, outputOffset + (rowReplicate - 1) * outputWeight, outputWeight)
rowReplicate += 1
}
outputOffset += outputWeight * (size(0) - 1)
rowIndex += 1
}
i += 1
}
output
}
protected def updateOutputNhwc[T: ClassTag](
input: Tensor[T],
output: Tensor[T],
size: Array[Int])(implicit ev: TensorNumeric[T]) : Tensor[T] = {
val inputHeight = input.size(2)
val inputWeight = input.size(3)
val outputHeight = inputHeight * size(0)
val outputWeight = inputWeight * size(1)
output.resize(input.size(1), outputHeight, outputWeight, input.size(4))
val channel = input.size(4)
val owc = outputWeight * channel
val inputData = input.storage().array()
var inputOffset = input.storageOffset() - 1
val outputData = output.storage().array()
var outputOffset = output.storageOffset() - 1
var i = 0
while (i < input.size(1)) {
var rowIndex = 0
while (rowIndex < input.size(2)) {
var columnIndex = 0
// copy column
while (columnIndex < input.size(3)) {
var colReplicate = 0
while (colReplicate < size(1)) {
ev.arraycopy(inputData, inputOffset,
outputData, outputOffset + colReplicate * channel, channel)
colReplicate += 1
}
outputOffset += channel * size(1)
inputOffset += channel
columnIndex += 1
}
// copy row
var rowReplicate = 1
while (rowReplicate < size(0)) {
ev.arraycopy(outputData, outputOffset - owc, outputData,
outputOffset + (rowReplicate - 1) * owc, owc)
rowReplicate += 1
}
outputOffset += owc * (size(0) - 1)
rowIndex += 1
}
i += 1
}
output
}
protected def updateGradInputNhwc[T: ClassTag](
gradInput: Tensor[T],
gradOutput: Tensor[T],
size: Array[Int])(implicit ev: TensorNumeric[T]) : Tensor[T] = {
val gradInputData = gradInput.storage().array()
var gradInputOffset = gradInput.storageOffset() - 1
val gradOutputData = gradOutput.storage().array()
var gradOutputOffset = gradOutput.storageOffset() - 1
val gradInputWidth = gradInput.size(4)
val gradOutputWidth = gradOutput.size(4)
val channel = gradInput.size(4)
val ocw = gradOutput.size(3) * gradOutput.size(4)
var i = 0
while (i < gradInput.size(1)) {
var rowIndex = 0
while (rowIndex < gradInput.size(2)) {
var colIndex = 0
while (colIndex < gradInput.size(3)) {
var rowReplicate = 0
while (rowReplicate < size(0)) {
var colReplicate = 0
while (colReplicate < size(1)) {
ev.axpy(channel, ev.one, gradOutputData,
gradOutputOffset + channel * colReplicate + rowReplicate * ocw, 1,
gradInputData, gradInputOffset, 1)
colReplicate += 1
}
rowReplicate += 1
}
gradInputOffset += channel
gradOutputOffset += size(1) * channel
colIndex += 1
}
gradOutputOffset += (size(0) - 1) * ocw
rowIndex += 1
}
i += 1
}
gradInput
}
protected def updateGradInputNchw[T: ClassTag](
gradInput: Tensor[T],
gradOutput: Tensor[T],
size: Array[Int])(implicit ev: TensorNumeric[T]) : Tensor[T] = {
val gradInputData = gradInput.storage().array()
var gradInputOffset = gradInput.storageOffset() - 1
val gradOutputData = gradOutput.storage().array()
var gradOutputOffset = gradOutput.storageOffset() - 1
val gradInputWidth = gradInput.size(4)
val gradOutputWidth = gradOutput.size(4)
var i = 0
while (i < gradInput.size(1) * gradInput.size(2) * gradInput.size(3)) {
var row = 0
while (row < size(0)) {
var col = 0
while (col < size(1)) {
ev.axpy(gradInputWidth, ev.one, gradOutputData,
gradOutputOffset + col, size(1),
gradInputData, gradInputOffset, 1)
col += 1
}
gradOutputOffset += gradOutputWidth
row += 1
}
gradInputOffset += gradInputWidth
i += 1
}
gradInput
}
}