在 C++ 版本的 Tensorflow 上使用多个 GPU

Use multiple gpus on the C++ version of tensorflow

本文关键字：GPU C++ 版本 Tensorflow      更新时间：2023-10-16

首先解释一下我的操作环境：

win10x64  
cuda9.1  and cudnn7  
gtx1080Ti x2  
i7-6850k

我使用 c++ 版本的 tensorflow 编写了一个程序，该程序读取 pb 文件，然后输入图像进行预测。我的目标是在一个线程中使用张量流或一个线程一个 GPU 时可以调用所有 GPU。

首先使用 windows 下的 python 调用 TensorFlow slim 训练，然后使用 freeze_graph.py 将保存的模型文件转换为冻结文件。

但是，我发现在使用session->Run((函数时只调用了一个gpu。无论是创建多个线程还是一个线程，我都使用以下方法来调用多个 gpu：

tensorflow::graph::SetDefaultDevice("0", &graphdef);

或

GraphDef graphdef; //Graph Definition for current model
Status status_load = ReadBinaryProto(Env::Default(), model_path, &graphdef); //read graph from pb_file
if (!status_load.ok()) {
        std::cout << " ERROR: Loading model failed...n"
        << model_path
        << std::endl;
    std::cout << status_load.ToString() << "n";
    system("pause");
    return;
}
tensorflow::SessionOptions options;
tensorflow::ConfigProto &config = options.config;
config.set_log_device_placement(true);
config.mutable_gpu_options()->set_allow_growth(true);
//config.mutable_gpu_options()->set_allocator_type(std::string("BFC"));
//config.mutable_gpu_options()->set_visible_device_list("");//this no error,but still can only call one gpu
//config.mutable_gpu_options()->set_visible_device_list("0");//error!
config.mutable_gpu_options()->set_visible_device_list("0,1");//error!
config.mutable_gpu_options()->set_per_process_gpu_memory_fraction(1);
Session* session;
Status status = NewSession(SessionOptions(options), &session);
Status status_create = session->Create(graphdef);

以上两种方法都失败了，错误提示相同：

2018-08-08 09:25:55.953495: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcoreplatformcpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-08-08 09:25:56.541237: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:1404] Found device 0 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582
pciBusID: 0000:06:00.0
totalMemory: 11.00GiB freeMemory: 9.02GiB
2018-08-08 09:25:56.708385: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:1404] Found device 1 with properties:
name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.582
pciBusID: 0000:0b:00.0
totalMemory: 11.00GiB freeMemory: 9.02GiB
2018-08-08 09:25:56.731390: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:1483] Adding visible gpu devices: 0, 1
2018-08-08 09:26:04.117910: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:964] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-08-08 09:26:04.131670: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:970]      0 1
2018-08-08 09:26:04.142367: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:983] 0:   N N
2018-08-08 09:26:04.152745: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_device.cc:983] 1:   N N
2018-08-08 09:26:04.173833: E D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimegpugpu_process_state.cc:105] Invalid allocator type: 0,1
2018-08-08 09:26:04.189278: E D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimedirect_session.cc:158] Internal: Failed to get memory allocator for TF GPU 0 with 11811160064 bytes of memory.
ERROR: Creating Session failed...
Internal: Failed to create session.
Press any key to continue......

根据提示，我切换到"/gpu/：0"和"/device：GPU：0"作为 GPU 的 ID。但提示解决失败，如下所示：

2018-08-08 09:31:07.052736: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcoreplatformcpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-08-08 09:31:07.643228: E D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimedirect_session.cc:158] Invalid argument: Could not parse entry in 'visible_device_list': '/device:GPU:0'. visible_device_list = /device:GPU:0
ERROR: Creating Session failed...
Internal: Failed to create session.

或

2018-08-08 09:32:28.753232: I D:MyProjecttensorflow-1.10.0-rc1tensorflowcoreplatformcpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
2018-08-08 09:32:29.082282: E D:MyProjecttensorflow-1.10.0-rc1tensorflowcorecommon_runtimedirect_session.cc:158] Invalid argument: Could not parse entry in 'visible_device_list': '/gpu:0'. visible_device_list = /gpu:0
ERROR: Creating Session failed...
Internal: Failed to create session.

然后我在/github/tensorflow 的问题中发现了同样的错误。我根据他们的方法尝试了以下方法：

遵循这些计划 #5379
1.{tf_root}\tensorflow\tf_version_script.lds
修改此文件，添加">protobuf;">
失败！
2. 添加相应的库。 

tf_core_gpu_kernelss.lib
training_ops_gen_cc.lib
transform_graph.lib
tf_protos_cc.lib
user_ops_gen_cc.lib

失败！

但是如果我使用以下方法：

config.mutable_gpu_options()->set_visible_device_list("")

或

tensorflow::graph::SetDefaultDevice("", &graphdef)

这可以通过并运行，但仍然只调用一个 gpu！

我在这个问题#18861中发现了同样的错误，但我在下面没有找到C++的解决方案，所以我怀疑这是我的张量流问题，我重新编译了 1.9.0 和最新的 1.10.0-rc1。但得到同样的错误

有人可以帮我解决这个问题吗？ └(^o^)┘
我真的很感激！
谢谢你重播我！

我可能已经找到了解决方案，但今天的测试没有达到我的要求。

tensorflow::SessionOptions options;
tensorflow::ConfigProto &config = options.config;
auto* device_count = config.mutable_device_count();
/*device_count->insert({ "CPU", 1 });*/
//device_count->insert({ "GPU", 1 });//1 represents one gpu, not the "/gpu:0"
device_count->insert({ "GPU", 2 });//2 represents two gpu, it is "/gpu:0" and "/gpu:1"
Session* session;
Status status = NewSession(options, &session);//creat new Session
std::vector<DeviceAttributes> response;
session->ListDevices(&response);
//print the device list
for (int temIndex = 0; temIndex < response.size(); ++temIndex) {
    auto temValue= response[temIndex];
    std::cout << "ListDevices(): " << temIndex << "  " << response[temIndex].name() << std::endl;
}

使用此方法与以下方法相同：

options.config.mutable_gpu_options()->set_visible_device_list("");

仍然无法明确定义要使用的GPU，并且仍然将所有计算都放在一个GPU上，我认为这可能是我的方法仍然存在问题。

但我觉得我会找到一个解决方案......

使用CUDA_VISIBLE_DEVICES为不同的进程设置特定的设备，这就是我的解决方案

相关文章：