CNN.sc

README.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## CNN\n",
    "\n",
    "This plugin provides a standalone implementation of a Convolutional Neural Network."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Dependencies"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import $ivy.`com.thoughtworks.each::each:3.3.1`\n",
    "import $ivy.`com.thoughtworks.deeplearning::plugins-builtins:2.0.0`\n",
    "\n",
    "import scala.concurrent.ExecutionContext.Implicits.global\n",
    "import org.nd4j.linalg.api.ndarray.INDArray\n",
    "import org.nd4j.linalg.convolution.Convolution\n",
    "import org.nd4j.linalg.util.ArrayUtil\n",
    "import org.nd4j.linalg.factory.Nd4j\n",
    "import org.nd4j.linalg.api.ops.impl.transforms.IsMax\n",
    "import scalaz.syntax.all._\n",
    "import com.thoughtworks.raii.shared._\n",
    "import com.thoughtworks.raii.asynchronous._\n",
    "import com.thoughtworks.feature.ImplicitApply\n",
    "import com.thoughtworks.each.Monadic._\n",
    "import com.thoughtworks.feature.Factory\n",
    "import com.thoughtworks.deeplearning.DeepLearning\n",
    "import com.thoughtworks.deeplearning.plugins.ImplicitsSingleton\n",
    "import com.thoughtworks.deeplearning.plugins.INDArrayLayers\n",
    "import com.thoughtworks.deeplearning.plugins.Operators\n",
    "import com.thoughtworks.deeplearning.plugins.Training\n",
    "import com.thoughtworks.deeplearning.plugins.Builtins"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "interp.load(scala.io.Source.fromURL(new java.net.URL(\"https://gist.githubusercontent.com/TerrorJack/cdd9cc5adc82fc86abf8b4c72cd26e76/raw/1f15523ee4b5a7fcc7e7317ae34ba09be207a62a/CNN.sc\")).mkString)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "interp.load(\"\"\"\n",
    "    val hyperparameters = Factory[Builtins with CNNs].newInstance()\n",
    "\"\"\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### For sbt projects"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "// build.sbt\n",
    "libraryDependencies += \"com.thoughtworks.deeplearning\" %% \"plugins-builtins\" % \"latest.release\"\n",
    "\n",
    "addCompilerPlugin(\"com.thoughtworks.import\" %% \"import\" % \"latest.release\")\n",
    "\n",
    "// XXX.scala\n",
    "import $exec.`https://gist.githubusercontent.com/TerrorJack/cdd9cc5adc82fc86abf8b4c72cd26e76/raw/1f15523ee4b5a7fcc7e7317ae34ba09be207a62a/CNN.sc`\n",
    "\n",
    "import scala.concurrent.ExecutionContext.Implicits.global\n",
    "import org.nd4j.linalg.api.ndarray.INDArray\n",
    "import org.nd4j.linalg.convolution.Convolution\n",
    "import org.nd4j.linalg.util.ArrayUtil\n",
    "import org.nd4j.linalg.factory.Nd4j\n",
    "import org.nd4j.linalg.api.ops.impl.transforms.IsMax\n",
    "import scalaz.syntax.all._\n",
    "import com.thoughtworks.raii.shared._\n",
    "import com.thoughtworks.raii.asynchronous._\n",
    "import com.thoughtworks.feature.ImplicitApply\n",
    "import com.thoughtworks.each.Monadic._\n",
    "import com.thoughtworks.feature.Factory\n",
    "import com.thoughtworks.deeplearning.DeepLearning\n",
    "import com.thoughtworks.deeplearning.plugins.ImplicitsSingleton\n",
    "import com.thoughtworks.deeplearning.plugins.INDArrayLayers\n",
    "import com.thoughtworks.deeplearning.plugins.Operators\n",
    "import com.thoughtworks.deeplearning.plugins.Training\n",
    "import com.thoughtworks.deeplearning.plugins.Builtins\n",
    "             \n",
    "val hyperparameters = Factory[Builtins with CNNs].newInstance()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Scala",
   "language": "scala",
   "name": "scala"
  },
  "language_info": {
   "codemirror_mode": "text/x-scala",
   "file_extension": ".scala",
   "mimetype": "text/x-scala",
   "name": "scala211",
   "nbconvert_exporter": "script",
   "pygments_lexer": "scala",
   "version": "2.11.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}

CNN.sc

trait CNNs extends INDArrayLayers with ImplicitsSingleton with Training with Operators {

  trait ImplicitsApi
    extends super[INDArrayLayers].ImplicitsApi
      with super[Training].ImplicitsApi
      with super[Operators].ImplicitsApi

  type Implicits <: ImplicitsApi

  private def toArray(tuple2: (Int, Int)): Array[Int] = {
    val (one, two) = tuple2
    Array(one, two)
  }

  def im2col[Operand0, Out <: INDArrayLayer](operand0: Operand0,
                                             kernel: (Int, Int),
                                             stride: (Int, Int),
                                             padding: (Int, Int))(
                                              implicit deepLearning: DeepLearning.Aux[Operand0, INDArray, INDArray],
                                              layerImplicits: ImplicitApply.Aux[indArrayPartialApplyRawForward.Rest, Out]): Out = {
    INDArrayLayer.unary(operand0) { data0: INDArray =>
      val shape0 = data0.shape
      val strideArray = toArray(stride)
      val paddingArray = toArray(padding)
      val outputData = Convolution.im2col(data0, toArray(kernel), strideArray, paddingArray)
      val delta0 = { outputDelta: INDArray =>
        Convolution.col2im(outputDelta, strideArray, paddingArray, shape0(2), shape0(3))
      }
      (outputData, delta0)
    }
  }

  @inline
  def conv2d[Input, Weight, Bias, Out <: INDArrayLayer](input: Input,
                                                        weight: Weight,
                                                        bias: Bias,
                                                        kernel: (Int, Int),
                                                        stride: (Int, Int),
                                                        padding: (Int, Int))(
                                                         implicit inputDeepLearning: DeepLearning.Aux[Input, INDArray, INDArray],
                                                         weightDeepLearning: DeepLearning.Aux[Weight, INDArray, INDArray],
                                                         biasDeepLearning: DeepLearning.Aux[Bias, INDArray, INDArray],
                                                         layerImplicits: ImplicitApply.Aux[indArrayPartialApplyRawForward.Rest, Out]): Out = {
    import implicits._
    INDArrayLayer(monadic[Do] {
      val inputShape = input.forward.each.data.shape
      val numberOfImages = inputShape(0)
      val depth = inputShape(1)
      val height = inputShape(2)
      val width = inputShape(3)
      val numberOfKernels = weight.forward.each.data.shape.head
      val col = im2col(input, kernel, stride, padding)
      val permutedCol = col.permute(0, 4, 5, 1, 2, 3)
      val depthKernelKernel = depth * kernel._1 * kernel._2
      val operandCol2d = permutedCol.reshape(numberOfImages * height * width, depthKernelKernel)
      val reshapedWeight = weight.reshape(numberOfKernels, depthKernelKernel)
      val permutedWeight = reshapedWeight.permute(1, 0)
      val dotResult = operandCol2d dot permutedWeight
      val plusResult = dotResult + bias
      val reshapeResult = plusResult.reshape(numberOfImages, height, width, numberOfKernels)
      reshapeResult.permute(0, 3, 1, 2).forward.each
    })
  }

  @inline
  def maxPool[Operand0, Out <: INDArrayLayer](operand0: Operand0, poolSize: (Int, Int))(
    implicit deepLearning: DeepLearning.Aux[Operand0, INDArray, INDArray],
    layerImplicits: ImplicitApply.Aux[indArrayPartialApplyRawForward.Rest, Out]): Out = {
    INDArrayLayer.unary(operand0) { data0: INDArray =>
      val shape0 = data0.shape
      val kernelAndStrideSize: Array[Int] = toArray(poolSize)
      val preMaxPool: INDArray =
        Convolution
          .im2col(data0, kernelAndStrideSize, kernelAndStrideSize, Array(0, 0))
          .permute(0, 1, 4, 5, 2, 3)
      val preShape: Seq[Int] = preMaxPool.shape().toSeq
      val lastDimensionSize: Int = preShape.takeRight(2).product
      val reshapedPreMaxPool: INDArray = preMaxPool
        .reshape(preShape.take(preShape.length - 2) :+ lastDimensionSize: _*)
      val outputData = reshapedPreMaxPool.max(4)
      val delta0 = { outputDelta: INDArray =>
        val a = reshapedPreMaxPool
        val upStreamDup = a.dup()
        val rows = ArrayUtil.prod(a.length())

        val isMax: INDArray = Nd4j.getExecutioner
          .execAndReturn(new IsMax(upStreamDup, 4))
          .reshape(preShape.take(preShape.length - 2) :+ poolSize._2 :+ poolSize._1: _*)
          .permute(0, 1, 2, 4, 3, 5)
          .reshape('c', rows, 1)

        val outputDelta1d = {
          outputDelta
            .repeat(-1, poolSize._1)
            .permute(1, 0, 3, 2)
            .repeat(-1, poolSize._2)
            .permute(1, 0, 3, 2)
            .reshape('c', shape0.product, 1)
        }

        isMax
          .muliColumnVector(outputDelta1d)
          .reshape(shape0: _*)
      }
      (outputData, delta0)
    }
  }

}

LICENSE

MIT License

Copyright (c) 2017 ThoughtWorks Inc.

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.