深度神经网络模型生成DSL - DNNGraph使用教程-优快云博客

本文链接：https://blog.youkuaiyun.com/gitblog_00229/article/details/146993715

深度神经网络模型生成DSL - DNNGraph使用教程

dnngraph A DSL for deep neural networks, supporting Caffe and Torch 项目地址: https://gitcode.com/gh_mirrors/dn/dnngraph

1. 项目介绍

DNNGraph是一个用Haskell编写的深度神经网络模型生成领域特定语言（DSL）。它提供了用于指定模型、优化模型性能以及生成不同平台代码的工具。DNNGraph使用lens库进行优雅、可组合的构造，并使用fgl图库来定义网络布局。目前，DNNGraph支持生成Caffe和Torch平台的代码，并提供了CLI工具来导出、可视化和理解模型。

2. 项目快速启动

首先，确保安装了Python 2和Protocol Buffers的protoc。以下是安装和启动DNNGraph的基本步骤：

# 安装hprotoc
cabal install hprotoc

# 生成协议缓冲区代码
./lens_proto.sh

# 安装DNNGraph
cabal install

以下是一个生成AlexNet模型的Haskell脚本示例：

import Control.Lens
import Control.Monad
import NN.DSL
import NN.Examples.ImageNet

alexTrain = train & cropSize' 227 & batchSize' 256 & mirror' True
alexTest = test & cropSize' 227 & batchSize' 50 & mirror' False
alexLrn = lrn & localSize' 5 & alphaLRN' 0.0001 & betaLRN' 0.75
alexConv = conv & param' alexMult & weightFillerC' (gaussian 0.01) & biasFillerC' zero
alexIP n = ip n & param' alexMult & weightFillerIP' (gaussian 0.005) & biasFillerIP' (constant 0.1)
alexPool = maxPool & sizeP' 3
alexMult = [def & lrMult' 1 & decayMult' 1, weights def & lrMult' 2 & decayMult' 0]

alexNet = do
    (input', representation) <- sequential [
        conv1, relu, alexLrn, alexPool & strideP' 3,
        conv2, relu, alexLrn, alexPool & strideP' 2,
        conv3, relu,
        conv4, relu,
        conv5, relu, alexPool & strideP' 2,
        alexIP 4096, relu, dropout 0.5,
        alexIP 4096, relu, dropout 0.5,
        alexIP 1000 & weightFillerIP' (gaussian 0.01) & biasFillerIP' zero]
    forM_ [alexTrain, alexTest] $ attach (To input')
    forM_ [accuracy 1, accuracy 5, softmax] $ attach (From representation)

3. 应用案例和最佳实践

DNNGraph的最佳实践包括定义清晰的模型结构、利用内置优化传递来提升性能，以及生成特定平台的代码。以下是一个定义GoogLeNet模型的例子：

import Control.Lens
import Control.Monad
import Data.Sequence (singleton)
import Data.Word
import NN
import NN.Examples.ImageNet

googleTrain = train & mirror' True & batchSize' 32 & cropSize' 224
googleTest = test & mirror' False & batchSize' 50 & cropSize' 224
googleMult = [def & lrMult' 1 & decayMult' 1, weights def & lrMult' 2 & decayMult' 0]

googleConv = conv & param' googleMult & biasFillerC' (constant 0.2)
googleLRN = lrn & localSize' 5 & alphaLRN' 0.0001 & betaLRN' 0.75
googlePool = maxPool & sizeP' 3 & strideP' 2

-- 定义Inception模块的函数
inception :: Node -> Inception -> NetBuilder Node
inception input (Inception n1 n3r n3 n5r n5 p) = do
    let columns = [
            [googleConv & numOutputC' n1 & kernelSizeC' 1 & weightFillerC' (xavier 0.03), relu],
            [googleConv & numOutputC' n3r & kernelSizeC' 1 & weightFillerC' (xavier 0.09), relu, googleConv & numOutputC' n3 & kernelSizeC' 3 & weightFillerC' (xavier 0.03) & padC' 1, relu],
            [googleConv & numOutputC' n5r & kernelSizeC' 1 & weightFillerC' (xavier 0.2), relu, googleConv & numOutputC' n5 & kernelSizeC' 5 & weightFillerC' (xavier 0.03) & padC' 2, relu],
            [maxPool & sizeP' 3 & strideP' 3 & padP' 1, googleConv & numOutputC' p & kernelSizeC' 1 & weightFillerC' (xavier 0.1), relu]]
        concat'' <- layer' concat'
    forM_ (zip columns [input, input, input, input]) $ \(bottoms, tops) -> do
        mapM_ (flip (>-) bottoms) tops
        tops >>= mapM_ (>- concat'')
    return concat''

-- 定义GoogLeNet网络
googLeNet :: NetBuilder ()
googLeNet = do
    (input, initial) <- sequential [conv1, relu, googlePool, googleLRN, conv2, relu, googleLRN, googlePool]
    top <- foldM insertRow initial [
        I $ Inception 64 96 128 16 32 32,
        I $ Inception 128 128 192 32 96 64,
        MaxPool,
        I $ Inception 192 96 208 16 48 64,
        Classifier,
        I $ Inception 150 112 224 24 64 64,
        I $ Inception 128 128 256 24 64 64,
        I $ Inception 112 144 288 32 64 64,
        Classifier,
        I $ Inception 256 160 320 32 128 128,
        MaxPool,
        I $ Inception 256 160 320 32 128 128
    ]