隨著網路越來約複雜，訓練難度越來越大，有條件的可以採用GPU進行學習。本文介紹如何在GPU環境下使用TensorFlow.NET。
TensorFlow.NET使用GPU非常的簡單，程式碼不用做任何修改，更換一個依賴庫即可，程式是否能執行成功主要看環境是否安裝正確，這篇文章重點介紹的也就是環境的安裝了。
CUDA和cuDNN的安裝都比較容易，重點是要裝對版本。
1、確認安裝版本
首先電腦得有一塊NVIDIA的顯示卡！
在桌面右鍵選擇NVIDIA控制皮膚，在程式左下角點選 ：系統資訊。
首先要確認顯示卡CUDA的版本，我的是10.1，實際安裝時要根據自己的環境來。
根據顯示卡CUDA版本，需要安裝對應的CUDA Toolkit和cuDNN的包，同時也要引用對應版本的TensorFlow_gpu執行庫。
這是TensorFlow官方網站提供的一個對應關係。
2、安裝CUDA Toolkit
安裝包下載地址：CUDA Toolkit 11.5 Update 1 Downloads | NVIDIA Developer
由於不可描述的原因，網站開啟很慢，需要有一定的耐心。
我下載的程式為：cuda_10.1.243_win10_network.exe，要保證前兩個版本號和本機顯示卡的CUDA版本一致，最後一位影響不大。
只要選對版本，安裝就沒有問題了，基本一路下一步就可以了。
3、安裝cuDNN
在NVIDIA網站搜尋cuDNN即可找到下載地址，同樣要注意版本。
我下載的檔案為：cudnn-10.1-windows10-x64-v8.0.5.39.zip
從檔名可以看出，該軟體版本為8.0，對應的CUDA版本為10.1，這點和TensonFlow網站上說的不完全一致，但實測可以使用。
注意cuDNN軟體不需要安裝，只要解壓到指定目錄即可。
假設CUDA安裝目錄為：C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\V10.1
解壓後的檔案包括：bin、inclue、lib三個資料夾，和CUDA安裝目錄下的三個同名資料夾合併即可。
【我懷疑隨便解壓到什麼位置，只要把路徑加入到path裡即可，但我沒有測試過。】
4、更換依賴庫
本系列第一篇就介紹過一個TensorFlow,NET專案的依賴庫如下：
如果需要採用GPU學習，更改其中的一個庫如下即可。
需要注意執行庫的版本。我的GPU最高支援2.3
如果此時你的執行環境沒有具備，程式執行時會彈出警告，並自動切換到CPU進行計算。
5、除錯
如果環境沒有安裝成功，程式執行會列印下列資訊：
2021-08-29 09:46:26.631756: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found 2021-08-29 09:46:26.633225: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine. 2021-08-29 09:46:27.323608: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags. 2021-08-29 09:46:27.993386: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found 2021-08-29 09:46:27.994493: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cublas64_11.dll'; dlerror: cublas64_11.dll not found 2021-08-29 09:46:27.997410: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cufft64_10.dll'; dlerror: cufft64_10.dll not found 2021-08-29 09:46:27.998498: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'curand64_10.dll'; dlerror: curand64_10.dll not found 2021-08-29 09:46:27.999403: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cusolver64_11.dll'; dlerror: cusolver64_11.dll not found 2021-08-29 09:46:28.000402: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cusparse64_11.dll'; dlerror: cusparse64_11.dll not found 2021-08-29 09:46:28.001807: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudnn64_8.dll'; dlerror: cudnn64_8.dll not found 2021-08-29 09:46:28.001917: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1835] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform. Skipping registering GPU devices...
關鍵字：Could not load dynamic library
如果環境安裝成功，則顯示如下：
2021-08-30 15:10:25.201050: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudart64_101.dll 2021-08-30 15:10:25.471127: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN)to use the following CPU instructions in performance-critical operations: AVX2 To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags. 2021-08-30 15:10:25.481201: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x1fa714fedc0 initialized for platform Host (this does not guarantee that XLA will be used). Devices: 2021-08-30 15:10:25.481288: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version 2021-08-30 15:10:25.485142: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library nvcuda.dll 2021-08-30 15:10:26.377044: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 0 with properties: pciBusID: 0000:01:00.0 name: GeForce GTX 1060 computeCapability: 6.1 coreClock: 1.6705GHz coreCount: 10 deviceMemorySize: 6.00GiB deviceMemoryBandwidth: 178.99GiB/s 2021-08-30 15:10:26.377230: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudart64_101.dll 2021-08-30 15:10:26.385642: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll 2021-08-30 15:10:26.390730: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cufft64_10.dll 2021-08-30 15:10:26.392476: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library curand64_10.dll 2021-08-30 15:10:26.402477: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cusparse64_10.dll 2021-08-30 15:10:26.413454: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudnn64_7.dll 2021-08-30 15:10:26.414923: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0 2021-08-30 15:10:27.287767: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1257] Device interconnect StreamExecutor with strength 1 edge matrix: 2021-08-30 15:10:27.287889: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1263] 0 2021-08-30 15:10:27.289201: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1276] 0: N 2021-08-30 15:10:27.290348: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1402] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4826 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1) 2021-08-30 15:10:27.297029: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x1fa3d8c6890 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices: 2021-08-30 15:10:27.297143: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): GeForce GTX 1060, Compute Capability 6.1 2021-08-30 15:10:27.890649: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll
關鍵字：Successfully opened dynamic library
【相關資源】