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 20 namespace mindspore { 21 namespace lite { 22 23 /// \brief MixPrecisionCfg defined for holding mix precision training configuration. 24 class MixPrecisionCfg { 25 public: MixPrecisionCfg()26 MixPrecisionCfg() { 27 this->dynamic_loss_scale_ = false; 28 this->loss_scale_ = 128.0f; 29 this->keep_batchnorm_fp32_ = true; 30 this->num_of_not_nan_iter_th_ = 1000; 31 this->is_raw_mix_precision_ = false; 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 this->is_raw_mix_precision_ = rhs.is_raw_mix_precision_; 39 } 40 MixPrecisionCfg &operator=(MixPrecisionCfg const &rhs) { 41 this->dynamic_loss_scale_ = rhs.dynamic_loss_scale_; 42 this->loss_scale_ = rhs.loss_scale_; 43 this->keep_batchnorm_fp32_ = rhs.keep_batchnorm_fp32_; 44 this->num_of_not_nan_iter_th_ = rhs.num_of_not_nan_iter_th_; 45 this->is_raw_mix_precision_ = rhs.is_raw_mix_precision_; 46 return *this; 47 } 48 bool dynamic_loss_scale_ = false; /**< Enable\disable dynamic loss scale during mix precision training */ 49 float loss_scale_; /**< Initial loss scale factor */ 50 bool keep_batchnorm_fp32_ = true; /**< Keep batch norm in FP32 while training */ 51 uint32_t num_of_not_nan_iter_th_; /**< a threshold for modifying loss scale when dynamic loss scale is enabled */ 52 bool is_raw_mix_precision_ = false; /**< Is mix precision model export from mindspore */ 53 }; 54 55 /// \brief TrainCfg defined for holding train configuration. 56 class TrainCfg { 57 public: TrainCfg()58 TrainCfg() { this->loss_name_ = "_loss_fn"; } TrainCfg(const TrainCfg & rhs)59 TrainCfg(const TrainCfg &rhs) { 60 this->loss_name_ = rhs.loss_name_; 61 this->mix_precision_cfg_ = rhs.mix_precision_cfg_; 62 this->accumulate_gradients_ = rhs.accumulate_gradients_; 63 } 64 TrainCfg &operator=(const TrainCfg &rhs) { 65 this->loss_name_ = rhs.loss_name_; 66 this->mix_precision_cfg_ = rhs.mix_precision_cfg_; 67 this->accumulate_gradients_ = rhs.accumulate_gradients_; 68 return *this; 69 } 70 std::string loss_name_; /**< Set part of the name that identify a loss kernel */ 71 MixPrecisionCfg mix_precision_cfg_; /**< Mix precision configuration */ 72 bool accumulate_gradients_ = false; /**< If true gardents are accmulated and can be read by GetGradients */ 73 }; 74 75 } // namespace lite 76 } // namespace mindspore 77 #endif // MINDSPORE_LITE_INCLUDE_TRAIN_TRAIN_CFG_H_ 78