LeNet5, AlexNet, VGGNet 的 MLJFlux 实现，有Nvida GPU的同学来试一下

nesteiner · 2022 年10 月 29 日 11:39

介绍

下面给出几个网络的实现和对比，有 Nvidia GPU 的同学试一下 usegpu=true ，我看看你们训练要多久
另外 Flux 的版本是 Flux#master

LeNet5

Tensorflow

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(filters=6, kernel_size=(5, 5), activation='sigmoid'),
    tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2),
    tf.keras.layers.Conv2D(filters=16, kernel_size=(5, 5), activation='sigmoid'),
    tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(120, activation='sigmoid'),
    tf.keras.layers.Dense(84, activation='sigmoid'),
    tf.keras.layers.Dense(10, activation='softmax')
])

MLJFlux

using Flux, MLJFlux
import StableRNGs: StableRNG


struct LeNet5 
    batchsize::Int
end

function MLJFlux.build(lenet5::LeNet5, rng, nin, nout, nchannels)
    return @autosize (nin..., nchannels, lenet5.batchsize) Chain(
      Conv((5, 5), _ => 6, relu),
      MaxPool((2, 2)),
      Conv((5, 5), _ => 16, relu),
      MaxPool((2, 2)),
      Flux.flatten,
      Dense(_ => 120, relu),
      Dense(120 => 84, relu),
      Dense(84 => nout)
    )
end

function buildLeNet5(;batchsize::Int, epochs::Int, lambda::Float32, alpha::Float32, usegpu = false)
    return ImageClassifier(
	builder = LeNet5(batchsize),
	batch_size = batchsize,
	epochs = epochs,
	rng = StableRNG(1234),
	lambda = lambda,
	alpha = alpha,
	acceleration = usegpu ? CUDALibs() : CPU1()
    )
end

AlexNet

Tensorflow

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(filters=96, kernel_size=(3, 3)),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=2),

    tf.keras.layers.Conv2D(filters=256, kernel_size=(3, 3)),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=2),

    tf.keras.layers.Conv2D(filters=384, kernel_size=(3, 3), padding='same', activation='relu'),
    tf.keras.layers.Conv2D(filters=384, kernel_size=(3, 3), padding='same', activation='relu'),
    tf.keras.layers.Conv2D(filters=256, kernel_size=(3, 3), padding='same', activation='relu'),
    tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=2),

    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(2048, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(2048, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(10, activation='softmax')
])

MLJFlux

using Flux, MLJFlux
import StableRNGs: StableRNG
import MLJFlux

struct AlexNet
    batchsize::Int
end

function MLJFlux.build(alexnet::AlexNet, rng, nin, nout, nchannels)
    return @autosize (nin..., nchannels, alexnet.batchsize) Chain(
	Conv((3, 3), nchannels => 96),
	BatchNorm(_, relu),
	MaxPool((3, 3), stride = 2),

	Conv((3, 3), _ => 256),
	BatchNorm(_, relu),
	MaxPool((3, 3), stride = 2),

	Conv((3, 3), _ => 384, relu, pad = SamePad()),
	Conv((3, 3), _ => 384, relu, pad = SamePad()),
	Conv((3, 3), _ => 256, relu, pad = SamePad()),
	MaxPool((3, 3), stride = 2),

	flatten,
	Dense(_ => 2048, relu),
	Dropout(0.5),
	Dense(_ => 2048, relu),
	Dropout(0.5),
	Dense(_ => nout)
    )
end


function buildAlexNet(;batchsize::Int, epochs::Int, lambda::Float32, alpha::Float32, usegpu = false)
    return ImageClassifier(
	builder = AlexNet(batchsize),
	batch_size = batchsize,
	epochs = epochs,
	rng = StableRNG(1234),
	lambda = lambda,
	alpha = alpha,
	acceleration = usegpu ? CUDALibs() : CPU1()
    )
end

VGGNet

Tensorflow

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(filters=64, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=64, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2, padding='same'),
    tf.keras.layers.Dropout(0.2),

    tf.keras.layers.Conv2D(filters=128, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=128, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2, padding='same'),
    tf.keras.layers.Dropout(0.2),

    tf.keras.layers.Conv2D(filters=256, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=256, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=256, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size=(2, 2), strides=2, padding='same'),
    tf.keras.layers.Dropout(0.2),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same'),
    tf.keras.layers.Dropout(0.2),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),

    tf.keras.layers.Conv2D(filters=512, kernel_size=(3, 3), padding='same'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same'),
    tf.keras.layers.Dropout(0.2),

    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512, activation = 'relu'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(512, activation = 'relu'),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation = 'softmax')
])

MLJFlux

using Flux, MLJFlux

struct VGGNet
    batchsize::Int
end

function MLJFlux.build(vggnet::VGGNet, rng, nin, nout, nchannels)
    return @autosize (nin..., nchannels, vggnet.batchsize) Chain(
	Conv((3, 3), nchannels => 64, pad = SamePad()),
	BatchNorm(64, relu),

	Conv((3, 3), _ => 64, pad = SamePad()),
	BatchNorm(_, relu),
	MaxPool((2, 2), stride = 2, pad = SamePad()),
	Dropout(0.2),

	Conv((3, 3), _ => 128, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 128, pad = SamePad()),
	BatchNorm(_, relu),
	MaxPool((2, 2), stride = 2, pad = SamePad()),
	Dropout(0.2),

	Conv((3, 3), _ => 256, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 256, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 256, pad = SamePad()),
	BatchNorm(_, relu),
	MaxPool((2, 2), stride = 2, pad = SamePad()),
	Dropout(0.2),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),
	MaxPool((2, 2), stride = 2, pad = SamePad()),
	Dropout(0.2),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),

	Conv((3, 3), _ => 512, pad = SamePad()),
	BatchNorm(_, relu),
	MaxPool((2, 2), stride = 2, pad = SamePad()),
	Dropout(0.2),

	flatten,

	Dense(_ => 512, relu),
	Dropout(0.2),
	Dense(_ => 512, relu),
	Dropout(0.2),
	Dense(_ => nout)
    )
end

function buildVGGNet(;batchsize::Int, epochs::Int, lambda::Float32, alpha::Float32, usegpu = false)
    return ImageClassifier(
	builder = VGGNet(batchsize),
	batch_size = batchsize,
	epochs = epochs,
	rng = StableRNG(1234),
	lambda = lambda,
	alpha = alpha,
	acceleration = usegpu ? CUDALibs() : CPU1()
    )
end

预测代码示例

数据下载

https://www.kaggle.com/competitions/digit-recognizer

下载完数据后，解压放在 data/digits-recognizer 目录下即可

2022-10-29_19-37-18_screenshot

准备阶段

using MLJFlux, Flux, StableRNGs, MLJ, CSV
using Flux: onehotbatch
using DataFrames: DataFrame

# load data
function transformDataType!(dataframe::DataFrame)
  coerce!(dataframe, Count => Continuous)
  if in(:label, names(dataframe))
    coerce!(dataframe, :label => Multiclass)
  end

  return dataframe
end


function loaddata(path::AbstractString)
  origindata = CSV.read(path, DataFrame)
  transformDataType!(origindata)
  y, X = unpack(origindata, colname -> colname == :label, colname -> true)
  images = reshape(transpose(Matrix(X)) ./ 255.0, (28, 28, :))

  labels = coerce(y, Multiclass)
  images = coerce(images, GrayImage)

  return labels, images
end

function loadtestdata(path::AbstractString)
  testdata = CSV.read(path, DataFrame)
  transformDataType!(testdata)
  images = reshape(transpose(Matrix(testdata)) ./ 255.0, (28, 28, :))
  images = coerce(images, GrayImage)
  return images
end

预测

function makepredict(pathtrain::AbstractString, pathtest::AbstractString, pathsubmission::AbstractString)
  rng = StableRNG(1234)
  y, X = loaddata(pathtrain)
  # model = buildLeNet5(batchsize = 32, epochs = 50, lambda = 10.0f0, alpha = 0.5f0, usegpu = true)
  # model = buildVGGNet(batchsize = 32, epochs = 50, lambda = 10.0f0, alpha = 0.5f0, usegpu = true)
  model = buildAlexNet(batchsize = 32, epochs = 1, lambda = 10.0f0, alpha = 0.5f0, usegpu = false)
  mach = machine(model, X, y)
  fit!(mach; verbosity = 2)

  testdata = loadtestdata(pathtest)
  output = map(x -> convert(Int, x), mode.(predict(mach, testdata)))

  outputdataframe = DataFrame()
  outputdataframe[!, :ImageId] = 1:length(output);
  outputdataframe[!, :Label] = output
  CSV.write(pathsubmission, outputdataframe)
end

makepredict("data/digits-recognizer/train.csv", 
	    "data/digits-recognizer/test.csv", 
	    "data/digits-recognizer/submission.csv")