fix bug in temp tensor strategy

include/sita/types.h

@@ -11,12 +11,19 @@
 #include "sita/proto/sita.h"
 namespace sita{
 
+const int PRE_TEMP_TENSOR_NUM = 64;
+template <typename Dtype>
+struct TempTensor{
+ int key;
+ Tensor<Dtype> * tensor;
+};
 
 template <typename Dtype>
 struct OperatorParam{
  std::string type;
  std::map<std::string, Tensor<Dtype> > params;
  std::map<std::string, FillerParameter> fillers;
+ std::map<std::string, bool> is_inited;
 };
 
 };

include/sita/workspace.h

@@ -43,37 +43,30 @@ template <typename Dtype>
 class GlobalWorkSpace : public WorkSpace{
 
 public:
- GlobalWorkSpace():WorkSpace(){};
+ GlobalWorkSpace():WorkSpace(){
+ _temp_tensor_control.resize(PRE_TEMP_TENSOR_NUM);
+ _temp_tensor.resize(PRE_TEMP_TENSOR_NUM);
+ _temp_tensor_num = 0;
+ };
  ~GlobalWorkSpace(){};
 
  //temp tensor
- std::pair<int, Tensor<Dtype> * > fetch_temp_tensor();
-
- void release_temp_tensor(int released_id);
-
- float temp_tensor_memory_size();
+ TempTensor<Dtype> new_tensor();
+ void free_tensor(TempTensor<Dtype> t);
+ void temp_tensor_memory();
 
  //flow tensor
  void init_input(std::string name);
-
  void init_output(std::string name);
-
  Tensor<Dtype>* fetch_input(std::string name);
-
  Tensor<Dtype>* fetch_output(std::string name);
-
  std::string flow_tensor_list();
-
 
  //params
- void init_param(std::string op_name, std::string op_type, std::string param_name, std::vector<int> shape, Filler);
-
+ void init_param(std::string op_name, std::string op_type, std::string param_name, std::vector<int> shape, FillerParameter filler);
  Tensor<Dtype>* fetch_param(std::string op_name, std::string param_name);
-
  std::string param_list();
 
-
-
  //grap
  inline void build_graph(Graph * graph){
  _graph = graph;
@@ -92,12 +85,13 @@ class GlobalWorkSpace : public WorkSpace{
  // temp_tensor bool true: using false:released
  std::vector<std::pair<Tensor<Dtype> *, bool> > _temp_tensor_control;
  std::vector<Tensor<Dtype> > _temp_tensor;
+ int _temp_tensor_num;
  //graph
  Graph * _graph;
  std::vector<boost::shared_ptr<Operator<Dtype> > > _ops;
  // input/output name
  std::map<std::string, Tensor<Dtype> > _flow_tensor;
- // params name  type weight/bias name weight/bias
+ // params name type weight/bias name weight/bias
  std::map<std::string, OperatorParam<Dtype> > _params;
  DISABLE_COPY_AND_ASSIGN(GlobalWorkSpace);
 };

main.cpp

@@ -25,8 +25,14 @@ int main(int argc, char** argv) {
 
  int k = 0;
 
- while(k != 40) {
+ while(k != 10) {
  k++;
+ sita::TempTensor<float> t = gws.new_tensor();
+ sita::Tensor<float> * a = t.tensor;
+ a->reshape(9,4,5,6);
+ // gws.free_tensor(t);
+
+
  sita::MnistBatch<float> * batch = mnistdp.fetch_batch();
  // LOG(INFO)<<batch->label()->cpu_data()[0];
 
@@ -43,11 +49,11 @@ int main(int argc, char** argv) {
  // std::cout<<std::endl;
  }
  cv::imwrite("vis/" + std::to_string(k)+"_"+ std::to_string(b)+"__"+std::to_string(int(batch->label()->cpu_data()[b]))+ ".jpg", cv_img_original);
- // LOG(INFO) << gws.temp_tensor_memory_size();
+
  }
  gws.train();
  }
-
+ gws.temp_tensor_memory();
 
  return 0;
 }

src/sita/workspace.cpp

@@ -26,48 +26,58 @@ void WorkSpace::set_device(int gpu_id){
 }
 
 template <typename Dtype>
-std::pair<int, Tensor<Dtype> * > GlobalWorkSpace<Dtype>::fetch_temp_tensor() {
- if (_temp_tensor.size() == 0) {
+TempTensor<Dtype> GlobalWorkSpace<Dtype>::new_tensor() {
+ TempTensor<Dtype> t;
+ if (_temp_tensor_num == 0) {
  Tensor<Dtype> temp_tensor;
- _temp_tensor.push_back(temp_tensor);
- _temp_tensor_control.push_back(std::make_pair(&(_temp_tensor[0]), true));
- return std::make_pair(0, &(_temp_tensor[0]));
+ _temp_tensor[0] = temp_tensor;
+ _temp_tensor_control[0] = std::make_pair(&(_temp_tensor[0]), true);
+ _temp_tensor_num++;
+ t.key = 0;
+ t.tensor = &(_temp_tensor[0]);
+ return t;
  } else {
  int released_id = -1;
- for (int i = 0; i < _temp_tensor.size(); i++) {
+ for (int i = 0; i < _temp_tensor_num; i++) {
  if (_temp_tensor_control[i].second == false) {
  released_id = i;
  }
  }
  if (released_id == -1) {
  Tensor<Dtype> temp_tensor;
- _temp_tensor.push_back(temp_tensor);
- _temp_tensor_control.push_back(std::make_pair(&(_temp_tensor[_temp_tensor.size() - 1]), true));
- return std::make_pair(0, &(_temp_tensor[_temp_tensor.size()-1]));
-
+ _temp_tensor[_temp_tensor_num] = temp_tensor;
+ _temp_tensor_control[_temp_tensor_num]= std::make_pair(&(_temp_tensor[_temp_tensor_num]), true);
+ t.key = 0;
+ t.tensor = &(_temp_tensor[_temp_tensor_num]);
+ _temp_tensor_num++;
+ return t;
  } else {
  _temp_tensor_control[released_id].second = true;
- return std::make_pair(released_id, &_temp_tensor[released_id]);
+ t.key = released_id;
+ t.tensor = &_temp_tensor[released_id];
+ return t;
  }
  }
 }
 
 template <typename Dtype>
-void GlobalWorkSpace<Dtype>::release_temp_tensor(int released_id) {
- _temp_tensor_control[released_id].second = false;
+void GlobalWorkSpace<Dtype>::free_tensor(TempTensor<Dtype> t) {
+ _temp_tensor_control[t.key].second = false;
 }
 
 template <typename Dtype>
-float GlobalWorkSpace<Dtype>::temp_tensor_memory_size(){
- int memory_size = 0;
- for(int i = 0; i < _temp_tensor.size(); i++){
+void GlobalWorkSpace<Dtype>::temp_tensor_memory(){
+ float memory_size = 0;
+ for(int i = 0; i < _temp_tensor_num; i++){
  memory_size += (_temp_tensor[i].count() * sizeof(Dtype));
  }
- return memory_size/(1024 * 1024 * 8);
+ LOG(INFO) << "the fact of temp tensor being used: "<<"[number:" << _temp_tensor_num << " " << "memory size: " <<
+ std::to_string(memory_size/(1024 * 8))<<" KB].";
+
+ return;
 }
 
 // flow tensor
-//std::map<std::string, Tensor<Dtype> > _flow_tensor;
 template <typename Dtype>
 void GlobalWorkSpace<Dtype>::init_input(std::string name){
  bool has_flow_tensor = false;
@@ -115,7 +125,7 @@ Tensor<Dtype>* GlobalWorkSpace<Dtype>::fetch_output(std::string name){
  return &(it->second);
  }
  }
- LOG(FATAL) << "no this onput in flow tensors, do you have init it?" << flow_tensor_list();
+ LOG(FATAL) << "no this output in flow tensors, do you have init it?" << flow_tensor_list();
 }
 
 template <typename Dtype>
@@ -129,7 +139,7 @@ std::string GlobalWorkSpace<Dtype>::flow_tensor_list(){
 }
 
 template <typename Dtype>
-void GlobalWorkSpace<Dtype>::init_param(std::string op_name, std::string op_type, std::string param_name, std::vector<int> shape, Filler filler){
+void GlobalWorkSpace<Dtype>::init_param(std::string op_name, std::string op_type, std::string param_name, std::vector<int> shape, FillerParameter filler){
  bool has_param = false;
  bool has_op_name = false;
  for(auto i = _params.begin(); i != _params.end(); i++){
@@ -153,6 +163,7 @@ void GlobalWorkSpace<Dtype>::init_param(std::string op_name, std::string op_type
  Tensor<Dtype> t(shape);
  _params[op_name].params[param_name] = t;
  _params[op_name].fillers[param_name] = filler;
+ _params[op_name].is_inited[param_name] = false;
 
  }
 }