• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2022-2023 Huawei Technologies Co., Ltd
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 #ifndef MINDSPORE_INCLUDE_API_CFG_H
17 #define MINDSPORE_INCLUDE_API_CFG_H
18 
19 #include <cstddef>
20 #include <string>
21 #include <vector>
22 #include <memory>
23 #include "include/api/data_type.h"
24 #include "include/api/dual_abi_helper.h"
25 #include "include/api/types.h"
26 
27 namespace mindspore {
28 constexpr int iter_th = 1000;
29 class MS_API MixPrecisionCfg {
30  public:
MixPrecisionCfg()31   MixPrecisionCfg() {
32     this->dynamic_loss_scale_ = false;
33     this->loss_scale_ = 128.0f;
34     this->keep_batchnorm_fp32_ = true;
35     this->num_of_not_nan_iter_th_ = iter_th;
36   }
MixPrecisionCfg(const MixPrecisionCfg & rhs)37   MixPrecisionCfg(const MixPrecisionCfg &rhs) {
38     this->dynamic_loss_scale_ = rhs.dynamic_loss_scale_;
39     this->loss_scale_ = rhs.loss_scale_;
40     this->keep_batchnorm_fp32_ = rhs.keep_batchnorm_fp32_;
41     this->num_of_not_nan_iter_th_ = rhs.num_of_not_nan_iter_th_;
42   }
43   ~MixPrecisionCfg() = default;
44 
45   bool dynamic_loss_scale_ = false;   /**< Enable/disable dynamic loss scale during mix precision training */
46   float loss_scale_;                  /**< Initial loss scale factor  */
47   bool keep_batchnorm_fp32_ = true;   /**< Keep batch norm in FP32 while training */
48   uint32_t num_of_not_nan_iter_th_;   /**< a threshold for modifying loss scale when dynamic loss scale is enabled */
49   bool is_raw_mix_precision_ = false; /**< Is mix precision model export from mindspore  */
50 };
51 
52 class MS_API TrainCfg {
53  public:
54   TrainCfg() = default;
TrainCfg(const TrainCfg & rhs)55   TrainCfg(const TrainCfg &rhs) {
56     this->loss_name_ = rhs.loss_name_;
57     this->mix_precision_cfg_ = rhs.mix_precision_cfg_;
58     this->accumulate_gradients_ = rhs.accumulate_gradients_;
59   }
60   ~TrainCfg() = default;
61 
62   /// \brief obtain part of the name that identify a loss kernel.
63   ///
64   /// \return loss_name.
65   inline std::vector<std::string> GetLossName() const;
66   /// \brief Set part of the name that identify a loss kernel.
67   ///
68   /// \param[in] loss_name define part of the name that identify a loss kernel.
69   inline void SetLossName(const std::vector<std::string> &loss_name);
70 
71   OptimizationLevel optimization_level_ = kO0;
72   MixPrecisionCfg mix_precision_cfg_; /**< Mix precision configuration */
73   bool accumulate_gradients_ = false;
74 
75  private:
76   std::vector<std::vector<char>> loss_name_ = VectorStringToChar({"loss_fct", "_loss_fn", "SigmoidCrossEntropy"});
77 };
78 
GetLossName()79 std::vector<std::string> TrainCfg::GetLossName() const { return VectorCharToString(loss_name_); }
SetLossName(const std::vector<std::string> & loss_name)80 void TrainCfg::SetLossName(const std::vector<std::string> &loss_name) { loss_name_ = VectorStringToChar(loss_name); }
81 }  // namespace mindspore
82 #endif  // MINDSPORE_INCLUDE_API_CFG_H
83