?這幾天想搞明白tensorflow的架構(gòu)慧邮，從網(wǎng)上搜集了一些資料铅匹。質(zhì)量參差不齊栅受，我把覺得質(zhì)量不錯(cuò)的附錄在這里株灸。
Tensorflow源碼解析2 -- 前后端連接的橋梁 - Session
?session->Run以DirectSession中的實(shí)現(xiàn)。

Status DirectSession::Run(const RunOptions& run_options,
                          const NamedTensorList& inputs,
                          const std::vector<string>& output_names,
                          const std::vector<string>& target_nodes,
                          std::vector<Tensor>* outputs,
                          RunMetadata* run_metadata,
                          const thread::ThreadPoolOptions& threadpool_options) {
  for (const auto& it : inputs) {
    input_tensor_names.push_back(it.first);
    input_size += it.second.AllocatedBytes();
  }
  TF_RETURN_IF_ERROR(GetOrCreateExecutors(input_tensor_names, output_names,
                                          target_nodes, &executors_and_keys,
                                          &run_state_args));
}

GetOrCreateExecutors的分析可以參考下面的博客：
TensorFlow 拆包（三）：Graph 和 Node

Status DirectSession::GetOrCreateExecutors(
    gtl::ArraySlice<string> inputs, gtl::ArraySlice<string> outputs,
    gtl::ArraySlice<string> target_nodes, ExecutorsAndKeys** executors_and_keys,
    RunStateArgs* run_state_args) {
  CallableOptions callable_options;
  callable_options.mutable_feed()->Reserve(inputs_sorted.size());
  for (const string& input : inputs_sorted) {
    callable_options.add_feed(input);
  }
  callable_options.mutable_fetch()->Reserve(outputs_sorted.size());
  for (const string& output : outputs_sorted) {
    callable_options.add_fetch(output);
  }

  TF_RETURN_IF_ERROR(
      CreateExecutors(callable_options, &ek, &func_info, run_state_args));    
}
Status DirectSession::CreateExecutors(
    const CallableOptions& callable_options,
    std::unique_ptr<ExecutorsAndKeys>* out_executors_and_keys,
    std::unique_ptr<FunctionInfo>* out_func_info,
    RunStateArgs* run_state_args) {
  std::unordered_map<string, std::unique_ptr<Graph>> graphs;
  TF_RETURN_IF_ERROR(CreateGraphs(
      options, &graphs, &func_info->flib_def, run_state_args, &ek->input_types,
      &ek->output_types, &ek->collective_graph_key));        
}
Status DirectSession::CreateGraphs(
    const BuildGraphOptions& subgraph_options,
    std::unordered_map<string, std::unique_ptr<Graph>>* outputs,
    std::unique_ptr<FunctionLibraryDefinition>* flib_def,
    RunStateArgs* run_state_args, DataTypeVector* input_types,
    DataTypeVector* output_types, int64* collective_graph_key) {
  std::unordered_map<string, GraphDef> partitions;
  TF_RETURN_IF_ERROR(Partition(popts, &client_graph->graph, &partitions));        
}

Status Partition(const PartitionOptions& opts, Graph* g,
                 std::unordered_map<string, GraphDef>* partitions) { }

?Executor的分析拉鹃，可以參考：
TensorFlow Executor解析

?其他一些值得一讀的博客：
TensorFlow的自動(dòng)求導(dǎo)具體是在哪部分代碼里實(shí)現(xiàn)的？
動(dòng)手實(shí)現(xiàn)TensorFlow--反向傳播Backpropagation
實(shí)現(xiàn)屬于自己的TensorFlow(三) - 反向傳播與梯度下降實(shí)現(xiàn)
Tensorflow compute_gradirnts和apply_gradients原理淺析
tensorflow optimizer源碼閱讀筆記
TensorFlow優(yōu)化器淺析 反向傳播圖
TensorFlow中的Placement啟發(fā)式算法模塊——Placer