Microsoft CNTK, Cognitive Toolkit 微軟深度學習工具
1 like387 views
Microsoft Cognitive Toolkit (CNTK) is an open-source deep learning framework that supports various programming languages and algorithms for artificial intelligence applications, such as DNNs, CNNs, and RNNs. The document outlines installation instructions, configurations for training and testing on the MNIST dataset, and provides commands for building and executing neural network models. It also includes details on how to evaluate model performance using various metrics and handle different layers in computational networks.
![建立設定檔：MAKEMODEL.CNTK
CNTK.EXE CONFIGFILE=MAKEMODEL.CNTK MAKEMODE=FALSE
Train =
[
action="train"
BrainScriptNetworkBuilder =
[
# define network here
]
]
Test =
[
# training commands here
]
WriteProbs =
[
# output commands here
]
DumpWeights =
[
# output commands here
# 加權和偏差定義預測模型
]](https://image.slidesharecdn.com/cntk-170516050520/85/Microsoft-CNTK-Cognitive-Toolkit-13-320.jpg)
Microsoft CNTK, Cognitive Toolkit 微軟深度學習工具
- 2. THE MICROSOFT COGNITIVE TOOLKIT Microsoft Research 64bits Linux & Windows Open Source Docker images Docker, Jupyter Support Python, C++, Brain Script, C#, Network definition language (NDL) and model editing language (MEL) DNN, CNN, RNN, LSTM, DSSM, Log-linear model
- 3. ACCESSIBILITY：HTTP://BLOGS.NVIDIA.COM/BLOG/2016/01/12/ACCELERATING-AI-ARTIFICIAL-INTELLIGENCE-GPUS/
- 4. ARCHITECTURE CNICNBuilderCN Description Use Build ILearnerIDataReaderFeatures & Labels Load Get data IExecutionEngine CPU/GPU Task-specific reader SGD, AdaGrad, etc. Evaluate Compute Gradient
- 5. MICROSOFT DEVELOPED ARTIFICIAL INTELLIGENCE Vision • Computer Vision APIs • Face APIs • Emotion APIs • Video APIs Speech • Speech APIs • Speaker Recognition APIs • Custom Recognition Intelligent Service Language • Spell Check APIs • Language Understanding Intelligent Service • Web Language Model APIs
- 6. LUIS, LANGUAGE UNDERSTANDING INTELLIGENT SERVICE BOT FRAMEWORK
- 7. 安裝CNTK 從 GitHub 下載並解壓縮檔案 github.com/Microsoft/CNTK/releases CPU、GPU、GPU – SGD (Stochastic Gradient Descent) Command Prompt cntkScriptsinstallwindowsinstall.bat cntkScriptscntkpy35.bat 檢查 Environment Path增加路徑：cntk/cntk/cntk.exe 路徑 VS2015 Runtime MSMPI：C++多執行緒支援SDK Anaconda3 C:localAnaconda3-4.1.1-Windows-x86_64 A CNTK-PY35 environment C:localAnaconda3-4.1.1-Windows-x86_64envs The CNTK Python module CNTK-PY35 environment
- 8. 資料檔案MNIST 解壓縮資料夾/Examples/Image/DataSets/MNIST python install_mnist.py 資料來源 http://yann.lecun.com/exdb/mnist/ 資料集合 Feature：1,000 Train-28x28_cntk_text.txt：60,000 Test-28x28_cntk_text.txt：10,000
- 9. 測試MNIST 解壓縮資料夾cntkExamplesImageGettingStartedREADME.md cntk configFile=01_OneHidden.cntk cntk configFile=02_OneConv.cntk cntk configFile=03_OneConvDropout.cntk cntk configFile=04_OneConvBN.cntk cntk configFile=05_OneConvRegr.cntk mpiexec -n 2 cntk configFile=06_OneConvRegrMultiNode.cntk parallelTrain=True parallelizationMethod=DataParallelSGD cntk configFile=07_Deconvolution_BS.cntk
- 10. 測試MNIST Epocch 15 Minibatch 500 Perplexity errs (*10000) ce (*10000) 01_OneHidden ReLU 1.06159522 1.720% 0.05977271 02_OneConv #sudo-image 16 x 24 x 24 ConvolutionalLayer {16,(5:5),pad = true} #feature map 16 x 12 x 12 MaxPoolingLayer {(2:2),stride=(2:2)(5:5)} # HiddenLayer 64 DenseLayer {64} 1.03299019 1.020% 0.03245770 03_OneConvDropout #feature map 16 x 12 x 12 MaxPoolingLayer：Dropout 1.03015284 0.980% 0.02970718 04_OneConvBN BatchNormalizationLayer (on CPU not yet implemented.)
- 11. 測試MNIST Epocch 15 Minibatch 500 rmse(*10000) totalSamplesSeen learningRatePerSample 05_OneConvRegr 02_OneConv RMSE 0.05052825 06_OneConvRegr MultiNode parallelizationMetho d 0.05116006 x 900000 0.00050000002 07_Deconvolution (Minibatch 10) Deconvolutio Unpooling 0.22284064
- 12. 建立設定檔：MAKEMODEL.CNTK CNTK.EXE CONFIGFILE=MAKEMODEL.CNTK MAKEMODE=FALSE 執行參數 modelPath = "ModelSimpleNet.snn“ 模型存放檔案名稱 deviceId = -1 CPU (-1) GPU (0) dimension = 2 labelDimension = 3 precision = "float“ float double Modularized computational networks execution engine learning algorithms model description data readers
- 13. 建立設定檔：MAKEMODEL.CNTK CNTK.EXE CONFIGFILE=MAKEMODEL.CNTK MAKEMODE=FALSE Train = [ action="train" BrainScriptNetworkBuilder = [ # define network here ] ] Test = [ # training commands here ] WriteProbs = [ # output commands here ] DumpWeights = [ # output commands here # 加權和偏差定義預測模型 ]
- 14. TRAIN模組：SGD參數 epochSize = 0 要使用多少訓練資料；0 = 使用所有資料 minibatchSize = 1 MinibatchSize 變數會指定每次訓練處理多少資料 learningRatesPerSample = 0.04 maxEpochs = 500 變數指定的訓練執行多少反覆項目 momentumPerMB = 0.90 處理多少資料時作紀錄
- 15. TEST模組 ce = CrossEntropyWithSoftmax (labels, myNet) 使用 Cross Entropy 計算訓練資料中的實際輸出值 err = ErrorPrediction (labels, myNet) 預測模型的精確度（百分比）、Cross Entropy、Perplexity 計算結果輸出和實際輸出之間的誤差 pn = Softmax (myNet) 輸出結果，機率值，正規化值總和為 1.0
- 16. WRITEPROBS模組 此模組是選擇性的但很有幫助，因為它可讓您對測試資料所做預測的其他資訊 WriteProbs 模組等同於測試模組，但不包括三個變更 動作參數設為 「 寫入 」 而不是"test“ 移除 randomize 參數 (因為 false 是預設值) 加入 outputPath 參數 WriteProbs 模組執行時，它會寫入指定的檔案來測試資料的確切的輸出值 .pn：這是在訓練模組的輸出節點使用的變數名稱
- 17. DUMP WEIGHTS模組 在執行時，此模組會儲存已訓練的模型加權和偏差值檔案 檔案會儲存在 modelPath 參數指定的路徑 .snn 是 使用二進位格式儲存的模型檔 maxEpochs 參數設定為 500 數字編號，最後 500 個檔案 .ckp是二進位的檢查點檔案 訓練複雜的類神經網路可以需要幾小時甚至幾天
- 22. THANK YOU!