optimizer/optimizer.h
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | #pragma once | ||
| 2 | |||
| 3 | #include <string> | ||
| 4 | #include <vector> | ||
| 5 | #include <unordered_map> | ||
| 6 | #include <cmath> | ||
| 7 | #include <memory> | ||
| 8 | #include <stdexcept> | ||
| 9 | #include "sparse_tensor.h" | ||
| 10 | #include "ps/base/base_client.h" | ||
| 11 | #include "ps/base/parameters.h" | ||
| 12 | |||
| 13 | using ::ParameterCompressReader; | ||
| 14 | using recstore::EmbeddingTableConfig; | ||
| 15 | |||
| 16 | class Optimizer { | ||
| 17 | protected: | ||
| 18 | std::unordered_map<std::string, SparseTensor*> tensor_map_; | ||
| 19 | |||
| 20 | public: | ||
| 21 | 40 | virtual ~Optimizer() { | |
| 22 |
2/2✓ Branch 5 taken 44 times.
✓ Branch 6 taken 40 times.
|
84 | for (auto& pair : tensor_map_) { |
| 23 |
1/2✓ Branch 0 taken 44 times.
✗ Branch 1 not taken.
|
44 | delete pair.second; |
| 24 | } | ||
| 25 | 40 | } | |
| 26 | |||
| 27 | virtual void Init(const std::vector<std::string> table_name, | ||
| 28 | const EmbeddingTableConfig& config, | ||
| 29 | BaseKV* base_kv) = 0; | ||
| 30 | |||
| 31 | virtual void Update(std::string table, | ||
| 32 | const ParameterCompressReader* reader, | ||
| 33 | unsigned tid) = 0; | ||
| 34 | virtual void UpdateFlat( | ||
| 35 | std::string table, | ||
| 36 | const base::ConstArray<uint64_t>& keys, | ||
| 37 | const float* grads, | ||
| 38 | int64_t num_rows, | ||
| 39 | int64_t embedding_dim, | ||
| 40 | unsigned tid) = 0; | ||
| 41 | }; | ||
| 42 | |||
| 43 | class SGD : public Optimizer { | ||
| 44 | private: | ||
| 45 | float learning_rate_; | ||
| 46 | |||
| 47 | public: | ||
| 48 | 30 | explicit SGD(float lr = 0.01) : learning_rate_(lr) {} | |
| 49 | |||
| 50 | void Init(const std::vector<std::string> table_name, | ||
| 51 | const EmbeddingTableConfig& config, | ||
| 52 | BaseKV* base_kv) override; | ||
| 53 | void Update(std::string table, | ||
| 54 | const ParameterCompressReader* reader, | ||
| 55 | unsigned tid) override; | ||
| 56 | void UpdateFlat(std::string table, | ||
| 57 | const base::ConstArray<uint64_t>& keys, | ||
| 58 | const float* grads, | ||
| 59 | int64_t num_rows, | ||
| 60 | int64_t embedding_dim, | ||
| 61 | unsigned tid) override; | ||
| 62 | }; | ||
| 63 | |||
| 64 | class AdaGrad : public Optimizer { | ||
| 65 | private: | ||
| 66 | float learning_rate_; | ||
| 67 | float epsilon_; | ||
| 68 | |||
| 69 | public: | ||
| 70 | explicit AdaGrad(float lr = 0.01, float epsilon = 1e-10) | ||
| 71 | : learning_rate_(lr), epsilon_(epsilon) {} | ||
| 72 | |||
| 73 | void Init(const std::vector<std::string> table_name, | ||
| 74 | const EmbeddingTableConfig& config, | ||
| 75 | BaseKV* base_kv) override; | ||
| 76 | void Update(std::string table, | ||
| 77 | const ParameterCompressReader* reader, | ||
| 78 | unsigned tid) override; | ||
| 79 | void UpdateFlat(std::string table, | ||
| 80 | const base::ConstArray<uint64_t>& keys, | ||
| 81 | const float* grads, | ||
| 82 | int64_t num_rows, | ||
| 83 | int64_t embedding_dim, | ||
| 84 | unsigned tid) override; | ||
| 85 | }; | ||
| 86 | |||
| 87 | class RowWiseAdaGrad : public Optimizer { | ||
| 88 | private: | ||
| 89 | float learning_rate_; | ||
| 90 | float epsilon_; | ||
| 91 | |||
| 92 | public: | ||
| 93 | 4 | explicit RowWiseAdaGrad(float lr = 0.01, float epsilon = 1e-10) | |
| 94 | 4 | : learning_rate_(lr), epsilon_(epsilon) {} | |
| 95 | |||
| 96 | void Init(const std::vector<std::string> table_name, | ||
| 97 | const EmbeddingTableConfig& config, | ||
| 98 | BaseKV* base_kv) override; | ||
| 99 | void Update(std::string table, | ||
| 100 | const ParameterCompressReader* reader, | ||
| 101 | unsigned tid) override; | ||
| 102 | void UpdateFlat(std::string table, | ||
| 103 | const base::ConstArray<uint64_t>& keys, | ||
| 104 | const float* grads, | ||
| 105 | int64_t num_rows, | ||
| 106 | int64_t embedding_dim, | ||
| 107 | unsigned tid) override; | ||
| 108 | }; | ||
| 109 | |||
| 110 | // Sparse AdamW keeps first/second moments and a persisted step counter in | ||
| 111 | // RecStore. Updates are applied to rows present in the submitted sparse | ||
| 112 | // gradient (the same sparse visibility contract as the existing optimizers). | ||
| 113 | class AdamW : public Optimizer { | ||
| 114 | private: | ||
| 115 | float learning_rate_; | ||
| 116 | float beta1_; | ||
| 117 | float beta2_; | ||
| 118 | float epsilon_; | ||
| 119 | float weight_decay_; | ||
| 120 | |||
| 121 | void UpdateRows(const std::string& table, | ||
| 122 | const uint64_t* keys, | ||
| 123 | const float* grads, | ||
| 124 | int64_t num_rows, | ||
| 125 | int64_t embedding_dim, | ||
| 126 | unsigned tid); | ||
| 127 | |||
| 128 | public: | ||
| 129 | 6 | explicit AdamW(float lr = 0.001, | |
| 130 | float beta1 = 0.9, | ||
| 131 | float beta2 = 0.98, | ||
| 132 | float epsilon = 1e-8, | ||
| 133 | float weight_decay = 0.0) | ||
| 134 | 12 | : learning_rate_(lr), | |
| 135 | 6 | beta1_(beta1), | |
| 136 | 6 | beta2_(beta2), | |
| 137 | 6 | epsilon_(epsilon), | |
| 138 | 6 | weight_decay_(weight_decay) {} | |
| 139 | |||
| 140 | void Init(const std::vector<std::string> table_name, | ||
| 141 | const EmbeddingTableConfig& config, | ||
| 142 | BaseKV* base_kv) override; | ||
| 143 | void Update(std::string table, | ||
| 144 | const ParameterCompressReader* reader, | ||
| 145 | unsigned tid) override; | ||
| 146 | void UpdateFlat(std::string table, | ||
| 147 | const base::ConstArray<uint64_t>& keys, | ||
| 148 | const float* grads, | ||
| 149 | int64_t num_rows, | ||
| 150 | int64_t embedding_dim, | ||
| 151 | unsigned tid) override; | ||
| 152 | }; | ||
| 153 | |||
| 154 | std::unique_ptr<Optimizer> CreateOptimizer(const json& config); | ||
| 155 |