Encoder-Decoder-Parser/dynet-2.0/exec.h at master · LeonCrashCode/Encoder-Decoder-Parser · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
#ifndef DYNET_EXEC_H
#define DYNET_EXEC_H

#include "dynet/dynet.h"

namespace dynet {

class ExecutionEngine {
 public:
  virtual ~ExecutionEngine();
  virtual void invalidate() = 0;
  virtual void invalidate(unsigned) = 0;
  virtual const Tensor& forward() = 0;
  virtual const Tensor& forward(VariableIndex i) = 0;
  virtual std::vector<const Tensor*> forward(std::vector<VariableIndex> is);  // forward on multiple nodes
  virtual const Tensor& incremental_forward() = 0;  // if you want to add nodes and evaluate just the new parts
  virtual const Tensor& incremental_forward(VariableIndex i) = 0;
  virtual const Tensor& get_value(VariableIndex i) = 0;
  virtual const Tensor& get_gradient(VariableIndex i) = 0;
  virtual void backward(bool full = false) = 0;
  virtual void backward(VariableIndex i, bool full = false) = 0;
 protected:
  explicit ExecutionEngine(const ComputationGraph& cg) : cg(cg), backward_computed(0) {}
  const ComputationGraph& cg;
  VariableIndex backward_computed;
};

class SimpleExecutionEngine : public ExecutionEngine {
 public:
  explicit SimpleExecutionEngine(const ComputationGraph& cg) : ExecutionEngine(cg), num_nodes_evaluated(0) {}
  void invalidate() override;
  void invalidate(unsigned i) override;
  const Tensor& forward() override;
  const Tensor& forward(VariableIndex i) override;
  const Tensor& incremental_forward() override;  // if you want to add nodes and evaluate just the new parts
  const Tensor& incremental_forward(VariableIndex i) override;
  const Tensor& get_value(VariableIndex i) override;
  const Tensor& get_gradient(VariableIndex i) override;
  void backward(bool full = false) override;
  void backward(VariableIndex i, bool full = false) override;
 private:
  std::vector<Tensor> nfxs;
  std::vector<Tensor> ndEdfs;
  VariableIndex num_nodes_evaluated;
};

struct BatchInfo {
public:
  BatchInfo() : pseudo_node(nullptr) { }
  Tensor nfx;             // The forward tensor, may be null if singleton batch
  Node* pseudo_node;      // The pseudo node used for calculation, also may be null if not needed
  std::vector<VariableIndex> ids; // IDs of the batch components
  std::vector<int> concat;       // 0=no need to concat, 1=need to concat, 2=need to concat + already contiguous in space
  std::vector<const Tensor*> arg_nfxs; // Concatenated arguments
};

class BatchedExecutionEngine : public ExecutionEngine {
 public:
  explicit BatchedExecutionEngine(const ComputationGraph& cg) : ExecutionEngine(cg), num_nodes_evaluated(0), num_batches_evaluated(0) { }
  ~BatchedExecutionEngine() { garbage_collect(); }
  void invalidate() override;
  void invalidate(unsigned i) override;
  const Tensor& forward() override;
  const Tensor& forward(VariableIndex i) override;
  const Tensor& incremental_forward() override;  // if you want to add nodes and evaluate just the new parts
  const Tensor& incremental_forward(VariableIndex i) override;
  const Tensor& get_value(VariableIndex i) override;
  const Tensor& get_gradient(VariableIndex i) override;
  void backward(bool full = false) override;
  void backward(VariableIndex i, bool full = false) override;
  void garbage_collect();
 private:
  const Tensor& incremental_forward_no_update(VariableIndex i, int autobatch_strategy);
  void combine_tensors(std::vector<VariableIndex> batch_ids, int aid, Tensor &tout);
  void accumulate_tensors(const Tensor& my_ndEdf, std::vector<VariableIndex> batch_ids, int aid);
  const Tensor& get_nfx(VariableIndex i);
  std::vector<Tensor> nfx_cache;
  std::vector<Tensor> ndEdfs;
  VariableIndex num_nodes_evaluated, num_batches_evaluated;
  // Information about the batched computation graph
  std::vector<VariableIndex> node2batch; // length: number of nodes
  std::vector<size_t> node2offset, node2size; // length: number of nodes
  std::vector<BatchInfo> batches; // length: number of batches
  SigMap sigmap;

};

} // namespace dynet

#endif