Pytorch C++ RuntimeError：设备类型为 cuda 的预期对象，但在调用 _th_index_select 时获得了参数 #1 'self' 的设备类型 cpu

Question

我在 Ubuntu 18.04LTS PyTorch C++（1.5.1，CUDA 10.1）上通过预训练的 wordvector (glove.300d) 使用 torch::Embedding 模块计算单词相似度。我相信我已经将所有我能做的事情都转移到了 GPU 上，但是当我执行它时，它仍然显示（问题末尾的完整错误日志）：

Expected object of device type cuda but got device type cpu for
  argument #1 'self' in call to _th_index_select
  (checked_dense_tensor_unwrap at /pytorch/aten/src/ATen/Utils.h:72)

我在ma​​in.cpp中检查了我的模型初始化方法，只做初始化就可以了。

SimilarityModel simiModel(args, 400000, 300);
simiModel.to(device);

//model forward
torch::Tensor data = ids.index({Slice(i*batch_size, (i+1)*batch_size), Slice()}).to(torch::kInt64).to(device);        //take a batch
tie(score, indice) = simiModel.forward(data);   //forward and transfer score, indice to cpu for further calculation

这就是我在 Similarity.h 中定义 SimilarityModel 的方式：

class SimilarityModel : public torch::nn::Module {
    public:
        int64_t topk;       // num of top words;
        Dictionary dict;
        int64_t vocab_size;
        int64_t embedding_dim;
        torch::nn::Embedding embedding{nullptr};
        vector<vector<float> > vec_embed;

        SimilarityModel(unordered_map<string, string> args, int64_t vocab_size, int64_t embed_dim);
        tuple<torch::Tensor, torch::Tensor> forward(torch::Tensor x);
};

同时我在Similarity.cpp的SimilarityModel函数中做了嵌入初始化：

SimilarityModel::SimilarityModel(unordered_map<string, string> args, int64_t vocab_size, int64_t embed_dim)
        :embedding(vocab_size, embed_dim) {      //Embedding initialize
    
    this->topk = stoi(args["topk"]);
    vector<vector<float> > pre_embed;
    tie(pre_embed, dict) = loadwordvec(args);       //load pretrained wordvec from txt file

    this->vocab_size = int64_t(dict.size());
    this->embedding_dim = int64_t(pre_embed[0].size());
    this->vec_embed = pre_embed;
    this->dict = dict;

    vector<float> temp_embed;
    for(const auto& i : pre_embed)      //faltten to 1-d
        for(const auto& j : i)
            temp_embed.push_back(j);
    torch::Tensor data = torch::from_blob(temp_embed.data(), {this->vocab_size, this->embedding_dim}, torch::TensorOptions().dtype(torch::kFloat32)).clone();   //vector to tensor    
    register_module("embedding", embedding);      
    this->embedding = embedding.from_pretrained(data, torch::nn::EmbeddingFromPretrainedOptions().freeze(true));
}

Similarity.cpp中的

和forward函数：

tuple<torch::Tensor, torch::Tensor> SimilarityModel::forward(torch::Tensor x) {     

    auto cuda_available = torch::cuda::is_available();      //copy to gpu
    torch::Device device(cuda_available ? torch::kCUDA : torch::kCPU);
    
    torch::Tensor wordvec;
    wordvec = this->embedding->forward(x).to(device);      //python:embedding(x)
    torch::Tensor similarity_score = wordvec.matmul(this->embedding->weight.transpose(0, 1)).to(device);
    torch::Tensor score, indice;
    tie(score, indice) = similarity_score.topk(this->topk, -1, true, true);        //Tensor.topk(int64_t k, int64_t dim, bool largest = true, bool sorted = true)

    score = score.to(device);
    indice = indice.to(device);
    score.slice(1, 1, score.size(1));       //Tensor.slice(int64_t dim, int64_t start, int64_t end, int64_t step)
    indice.slice(1, 1, indice.size(1));
    return {score.cpu(), indice.cpu()};   //transfer to cpu for further calculation
}

至于forward()中的中间变量也已经放到GPU上了。但是，我完全不知道 CPU 中还剩下哪一个，而且错误日志也没有多大帮助。我已经尝试过Expected object of device type cuda but got device type cpu for argument #1 'self' in call to _th_index_select 中的方法来做SimilarityModel().to(device)，但这不起作用。我仍然很难阅读此错误日志，并且想要一些有关如何调试此类问题的说明。

错误日志：

terminate called after throwing an instance of 'c10::Error'
  what():  Expected object of device type cuda but got device type cpu for argument #1 'self' in call to _th_index_select (checked_dense_tensor_unwrap at /pytorch/aten/src/ATen/Utils.h:72)
frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x46 (0x7fb566a27536 in /home/switchsyj/Downloads/libtorch/lib/libc10.so)
frame #1: <unknown function> + 0x101a80b (0x7fb520fa380b in /home/switchsyj/Downloads/libtorch/lib/libtorch_cuda.so)
frame #2: <unknown function> + 0x105009c (0x7fb520fd909c in /home/switchsyj/Downloads/libtorch/lib/libtorch_cuda.so)
frame #3: <unknown function> + 0xf9d76b (0x7fb520f2676b in /home/switchsyj/Downloads/libtorch/lib/libtorch_cuda.so)
frame #4: <unknown function> + 0x10c44e3 (0x7fb558d224e3 in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #5: at::native::embedding(at::Tensor const&, at::Tensor const&, long, bool, bool) + 0x2e2 (0x7fb558870712 in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #6: <unknown function> + 0x114ef9d (0x7fb558dacf9d in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #7: <unknown function> + 0x1187b4d (0x7fb558de5b4d in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #8: <unknown function> + 0x2bfe42f (0x7fb55a85c42f in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #9: <unknown function> + 0x1187b4d (0x7fb558de5b4d in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #10: <unknown function> + 0x32b63a9 (0x7fb55af143a9 in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #11: torch::nn::EmbeddingImpl::forward(at::Tensor const&) + 0x71 (0x7fb55af127b1 in /home/switchsyj/Downloads/libtorch/lib/libtorch_cpu.so)
frame #12: SimilarityModel::forward(at::Tensor) + 0xa9 (0x55c96b8e5793 in ./demo)
frame #13: main + 0xaba (0x55c96b8bfe5c in ./demo)
frame #14: __libc_start_main + 0xe7 (0x7fb51edf5b97 in /lib/x86_64-linux-gnu/libc.so.6)
frame #15: _start + 0x2a (0x55c96b8bd74a in ./demo)

Aborted (core dumped)

Answer 1

根据错误消息，当您运行 SimilarityModel::forward() 时，以下两个 Tensors 之一不在 GPU 中：

• this->embedding->weight
• x

鉴于错误指向argument #1，我会说weight 是CPU 上的错误。

这是index.select的电话：

Tensor embedding(const Tensor & weight, const Tensor & indices,
                 int64_t padding_idx, bool scale_grad_by_freq, bool sparse) {
  auto indices_arg = TensorArg(indices, "indices", 1);
  checkScalarType("embedding", indices_arg, kLong);

  // TODO: use tensor.index() after improving perf
  if (indices.dim() == 1) {
    return weight.index_select(0, indices);
  }

  auto size = indices.sizes().vec();
  for (auto d : weight.sizes().slice(1)) {
    size.push_back(d);
  }
  return weight.index_select(0, indices.reshape(-1)).view(size);
}

首先，尝试将权重直接移动到 GPU。如果它有效，则意味着当您致电TORCH_MODULE(SimilarityModel) 并将模型移动到设备时，它也应该有效。请记住，在这种情况下，您必须将名称更改为 SimilarityModelImpl (Name+Impl)。否则，它将无法正常工作。