• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2021 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_LITE_INCLUDE_TRAIN_TRAIN_CFG_H_
17 #define MINDSPORE_LITE_INCLUDE_TRAIN_TRAIN_CFG_H_
18 #include <string>
19 #include <vector>
20 
21 namespace mindspore {
22 namespace lite {
23 
24 /// \brief MixPrecisionCfg defined for holding mix precision training configuration.
25 class MixPrecisionCfg {
26  public:
MixPrecisionCfg()27   MixPrecisionCfg() {
28     this->dynamic_loss_scale_ = false;
29     this->loss_scale_ = 128.0f;
30     this->keep_batchnorm_fp32_ = true;
31     this->num_of_not_nan_iter_th_ = 1000;
32   }
MixPrecisionCfg(const MixPrecisionCfg & rhs)33   MixPrecisionCfg(const MixPrecisionCfg &rhs) {
34     this->dynamic_loss_scale_ = rhs.dynamic_loss_scale_;
35     this->loss_scale_ = rhs.loss_scale_;
36     this->keep_batchnorm_fp32_ = rhs.keep_batchnorm_fp32_;
37     this->num_of_not_nan_iter_th_ = rhs.num_of_not_nan_iter_th_;
38   }
39   MixPrecisionCfg &operator=(MixPrecisionCfg const &rhs) = default;
40   bool dynamic_loss_scale_ = false; /**< Enable\disable dynamic loss scale during mix precision training */
41   float loss_scale_;                /**< Initial loss scale factor  */
42   bool keep_batchnorm_fp32_ = true; /**< Keep batch norm in FP32 while training */
43   uint32_t num_of_not_nan_iter_th_; /**< a threshold for modifying loss scale when dynamic loss scale is enabled */
44 };
45 
46 /// \brief TrainCfg defined for holding train configuration.
47 class TrainCfg {
48  public:
TrainCfg()49   TrainCfg() {
50     this->loss_name_.emplace_back("_loss_fn");
51     this->loss_name_.emplace_back("SigmoidCrossEntropy");
52   }
TrainCfg(const TrainCfg & rhs)53   TrainCfg(const TrainCfg &rhs) {
54     this->loss_name_ = rhs.loss_name_;
55     this->mix_precision_cfg_ = rhs.mix_precision_cfg_;
56     this->accumulate_gradients_ = rhs.accumulate_gradients_;
57   }
58   TrainCfg &operator=(const TrainCfg &rhs) = default;
59   std::vector<std::string> loss_name_ = {"loss_fct"}; /**< Set part of the name that identify a loss kernel */
60   MixPrecisionCfg mix_precision_cfg_;                 /**< Mix precision configuration */
61   bool accumulate_gradients_ = false; /**< If true gardents are accmulated and can be read by GetGradients */
62 };
63 
64 }  // namespace lite
65 }  // namespace mindspore
66 #endif  // MINDSPORE_LITE_INCLUDE_TRAIN_TRAIN_CFG_H_
67