Col_lin
/
mindspore2022
forked from huawei/mindspore2022
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

!18328 [MS][LITE][TOD] replace internal API to CamelCase 

Merge pull request !18328 from yonibaehr/export_yoni
This commit is contained in:
i-robot 2021-06-17 09:11:08 +08:00 committed by Gitee
parent ec74f46684 40cb67b26e
commit beb66e483c
32 changed files with 88 additions and 74 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
mindspore/lite/src/inner_kernel.h
View File

@ -183,9 +183,9 @@ class InnerKernel : public Kernel {
virtual bool IsEval() const { return !this->train_mode_; }
virtual void set_trainable(bool trainable = true) { this->trainable_ = trainable; }
virtual void SetTrainable(bool trainable = true) { this->trainable_ = trainable; }
virtual bool is_trainable() const { return this->trainable_; }
virtual bool IsTrainable() const { return this->trainable_; }
TypeId registry_data_type(void) { return registry_data_type_; }

8
mindspore/lite/src/lite_kernel.h
View File

@ -192,17 +192,17 @@ class LiteKernel {
return false;
}
virtual void set_trainable(bool trainable = true) {
virtual void SetTrainable(bool trainable = true) {
MS_ASSERT(kernel_ != nullptr);
if (desc_.provider == kBuiltin) {
std::static_pointer_cast<InnerKernel>(kernel_)->set_trainable(trainable);
std::static_pointer_cast<InnerKernel>(kernel_)->SetTrainable(trainable);
}
}
virtual bool is_trainable() const {
virtual bool IsTrainable() const {
MS_ASSERT(kernel_ != nullptr);
if (desc_.provider == kBuiltin) {
return std::static_pointer_cast<InnerKernel>(kernel_)->is_trainable();
return std::static_pointer_cast<InnerKernel>(kernel_)->IsTrainable();
}
return false;
}

2
mindspore/lite/src/runtime/kernel/arm/base/convolution_base.h
View File

@ -61,8 +61,8 @@ class ConvolutionBaseCPUKernel : public InnerKernel {
void FreeAlignedData(void **ptr);
protected:
bool IsRepack() { return is_repack_; }
std::unordered_map<uintptr_t, void *> addr_map;
bool is_repack() { return is_repack_; }
void *bias_data_ = nullptr;
const InnerContext *ctx_ = nullptr;
ConvParameter *conv_param_ = nullptr;

4
mindspore/lite/src/runtime/kernel/arm/fp16/biasadd_fp16.cc
View File

@ -53,7 +53,7 @@ int BiasAddCPUFp16Kernel::Run() {
}
}
if (op_parameter_->is_train_session_) {
if ((is_trainable() && (IsTrain() || is_repack())) || (bias_data_type_ == kNumberTypeFloat16)) {
if ((IsTrainable() && (IsTrain() || IsRepack())) || (bias_data_type_ == kNumberTypeFloat16)) {
PackWeight();
is_repack_ = false;
}
@ -133,7 +133,7 @@ void BiasAddCPUFp16Kernel::PackWeight() {
int BiasAddCPUFp16Kernel::Eval() {
InnerKernel::Eval();
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return RET_OK;

2
mindspore/lite/src/runtime/kernel/arm/fp16/biasadd_fp16.h
View File

@ -38,7 +38,7 @@ class BiasAddCPUFp16Kernel : public InnerKernel {
private:
int GetBiasData();
void PackWeight();
bool is_repack() { return is_repack_; }
bool IsRepack() { return is_repack_; }
ArithmeticParameter *bias_param_ = nullptr;
float16_t *bias_data_ = nullptr;
lite::Tensor *bias_tensor_ = nullptr;

8
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_1x1_fp16.cc
View File

@ -97,7 +97,7 @@ int Convolution1x1FP16CPUKernel::InitWeightBias() {
return RET_ERROR;
}
}
void *bias_origin_tmp = is_trainable() ? in_tensors_.at(2)->data_c() : origin_bias_;
void *bias_origin_tmp = IsTrainable() ? in_tensors_.at(2)->data_c() : origin_bias_;
memcpy(bias_data_, bias_origin_tmp, output_channel * sizeof(float16_t));
memset(reinterpret_cast<char *>(bias_data_) + bias_size, 0, size - bias_size);
}
@ -111,7 +111,7 @@ int Convolution1x1FP16CPUKernel::InitWeightBias() {
return RET_ERROR;
}
}
void *weight_origin_tmp = is_trainable() ? weight_tensor->data_c() : origin_weight_;
void *weight_origin_tmp = IsTrainable() ? weight_tensor->data_c() : origin_weight_;
memset(reinterpret_cast<char *>(weight_ptr_) + down_size, 0, size - down_size);
ColMajor2Row8MajorFp16(weight_origin_tmp, weight_ptr_, input_channel, output_channel, true);
@ -240,7 +240,7 @@ int Convolution1x1FP16CPUKernel::Run() {
return RET_MEMORY_FAILED;
}
if (is_trainable() && (IsTrain() || is_repack())) {
if (IsTrainable() && (IsTrain() || IsRepack())) {
auto ret = InitWeightBias();
if (ret != 0) {
MS_LOG(ERROR) << "Convolution 1x1 fp16 repack weight failure";
@ -283,7 +283,7 @@ int Convolution1x1FP16CPUKernel::Run() {
}
int Convolution1x1FP16CPUKernel::Eval() {
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return InnerKernel::Eval();

6
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_delegate_fp16.h
View File

@ -55,9 +55,9 @@ class ConvolutionDelegateFP16CPUKernel : public InnerKernel {
InnerKernel::Train();
return fp16_conv_kernel_->Train();
}
void set_trainable(bool trainable) override {
InnerKernel::set_trainable(trainable);
return fp16_conv_kernel_->set_trainable(trainable);
void SetTrainable(bool trainable) override {
InnerKernel::SetTrainable(trainable);
return fp16_conv_kernel_->SetTrainable(trainable);
}
void set_in_tensor(lite::Tensor *in_tensor, int index) override {

4
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_fp16.cc
View File

@ -109,7 +109,7 @@ static int ConvDwFp16Run(void *cdata, int task_id, float lhs_scale, float rhs_sc
}
int ConvolutionDepthwiseFp16CPUKernel::Run() {
if (is_trainable() && (IsTrain() || is_repack())) {
if (IsTrainable() && (IsTrain() || IsRepack())) {
auto ret = InitWeightBias();
if (ret != 0) {
MS_LOG(ERROR) << "Convolution depthwise fp16 repack weight failure";
@ -125,7 +125,7 @@ int ConvolutionDepthwiseFp16CPUKernel::Run() {
}
int ConvolutionDepthwiseFp16CPUKernel::Eval() {
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return InnerKernel::Eval();

4
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_depthwise_slidewindow_fp16.cc
View File

@ -161,7 +161,7 @@ int ConvolutionDepthwiseSWFp16CPUKernel::Run() {
packed_output_ = output_ptr;
}
if (is_trainable() && (IsTrain() || is_repack())) {
if (IsTrainable() && (IsTrain() || IsRepack())) {
ret = InitWeightBias();
if (ret != 0) {
MS_LOG(ERROR) << "Convolution depthwise fp16 repack weight failure";
@ -192,7 +192,7 @@ void ConvolutionDepthwiseSWFp16CPUKernel::FreePackedInputOutput() {
}
int ConvolutionDepthwiseSWFp16CPUKernel::Eval() {
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return InnerKernel::Eval();

8
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc
View File

@ -46,7 +46,7 @@ int ConvolutionFP16CPUKernel::InitWeightBias() {
}
}
memset(packed_weight_, 0, pack_weight_size * sizeof(float16_t));
void *weight_origin_tmp = is_trainable() ? filter_tensor->data_c() : origin_weight_;
void *weight_origin_tmp = IsTrainable() ? filter_tensor->data_c() : origin_weight_;
RowMajor2Col8MajorFp16(weight_origin_tmp, packed_weight_, out_channel, in_channel * kernel_plane, false);
// init bias
@ -60,7 +60,7 @@ int ConvolutionFP16CPUKernel::InitWeightBias() {
memset(bias_data_, 0, oc8 * sizeof(float16_t));
if (in_tensors_.size() == kInputSize2) {
auto bias_tensor = in_tensors_.at(kBiasIndex);
void *bias_origin_tmp = is_trainable() ? bias_tensor->data_c() : origin_bias_;
void *bias_origin_tmp = IsTrainable() ? bias_tensor->data_c() : origin_bias_;
memcpy(bias_data_, bias_origin_tmp, out_channel * sizeof(float16_t));
}
return RET_OK;
@ -152,7 +152,7 @@ int ConvolutionFP16CPUKernel::Run() {
return RET_ERROR;
}
if (is_trainable() && (IsTrain() || is_repack())) {
if (IsTrainable() && (IsTrain() || IsRepack())) {
ret = InitWeightBias();
if (ret != 0) {
MS_LOG(ERROR) << "Convolution 1x1 fp16 repack weight failure";
@ -170,7 +170,7 @@ int ConvolutionFP16CPUKernel::Run() {
}
int ConvolutionFP16CPUKernel::Eval() {
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return InnerKernel::Eval();

8
mindspore/lite/src/runtime/kernel/arm/fp16/convolution_winograd_fp16.cc
View File

@ -66,7 +66,7 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
MS_LOG(ERROR) << "get matrix g from CookToomFilter failed.";
return ret;
}
void *weight_origin_tmp = is_trainable() ? weight_tensor->data_c() : origin_weight_;
void *weight_origin_tmp = IsTrainable() ? weight_tensor->data_c() : origin_weight_;
ret = WinogradFilterTransformFp16(reinterpret_cast<float16_t *>(weight_origin_tmp), matrix_g, matrix_gt, col_tile_);
if (ret != RET_OK) {
MS_LOG(ERROR) << "winograd filter transform failed.";
@ -84,7 +84,7 @@ int ConvolutionWinogradFP16CPUKernel::InitWeightBias() {
memset(bias_data_, 0, oc_block_num * col_tile_ * sizeof(float16_t));
if (in_tensors_.size() == kInputSize2) {
auto bias_tensor = in_tensors_.at(kBiasIndex);
void *bias_origin_tmp = is_trainable() ? bias_tensor->data_c() : origin_bias_;
void *bias_origin_tmp = IsTrainable() ? bias_tensor->data_c() : origin_bias_;
memcpy(bias_data_, bias_origin_tmp, out_channel * sizeof(float16_t));
}
return RET_OK;
@ -229,7 +229,7 @@ int ConvolutionWinogradFP16CPUKernel::Run() {
FreeTmpBuffer();
return RET_ERROR;
}
if (is_trainable() && (IsTrain() || is_repack())) {
if (IsTrainable() && (IsTrain() || IsRepack())) {
ret = InitWeightBias();
if (ret != 0) {
MS_LOG(ERROR) << "ConvolutionWinogradFP16 repack weight failure";
@ -246,7 +246,7 @@ int ConvolutionWinogradFP16CPUKernel::Run() {
}
int ConvolutionWinogradFP16CPUKernel::Eval() {
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return InnerKernel::Eval();

4
mindspore/lite/src/runtime/kernel/arm/fp16/fused_batchnorm_fp16.cc
View File

@ -88,7 +88,7 @@ int FusedBatchnormFp16CPUKernel::DoExecute(int task_id) {
Float32ToFloat16(reinterpret_cast<float *>(variance->data_c()), reinterpret_cast<float16_t *>(variance_fp16),
variance->ElementsNum());
if (IsTrain() && is_trainable() && in_tensors_.size() >= 5) {
if (IsTrain() && IsTrainable() && in_tensors_.size() >= 5) {
CalcMeanVar(reinterpret_cast<float16_t *>(input_fp16), reinterpret_cast<float16_t *>(scale_fp16),
reinterpret_cast<float16_t *>(offset_fp16), reinterpret_cast<float16_t *>(mean_fp16),
reinterpret_cast<float16_t *>(variance_fp16));
@ -108,7 +108,7 @@ int FusedBatchnormFp16CPUKernel::DoExecute(int task_id) {
return RET_OK;
}
if (IsTrain() && is_trainable() && in_tensors_.size() >= 5) {
if (IsTrain() && IsTrainable() && in_tensors_.size() >= 5) {
CalcMeanVar(
static_cast<float16_t *>(in_tensors_.at(0)->data_c()), static_cast<float16_t *>(in_tensors_.at(1)->data_c()),
static_cast<float16_t *>(in_tensors_.at(2)->data_c()), static_cast<float16_t *>(in_tensors_.at(3)->data_c()),

4
mindspore/lite/src/runtime/kernel/arm/fp16/matmul_base_fp16.cc
View File

@ -270,13 +270,13 @@ int MatmulBaseFP16CPUKernel::RunImpl(int task_id) {
int MatmulBaseFP16CPUKernel::Run() {
auto c_ptr = reinterpret_cast<float16_t *>(out_tensors_.at(0)->data_c());
if ((params_->a_const_ == false) || is_repack()) {
if ((params_->a_const_ == false) || IsRepack()) {
if (RET_OK != InitBufferA()) {
return RET_ERROR;
}
InitMatrixA(in_tensors_.at(0)->data_c());
}
if ((params_->b_const_ == false) || is_repack()) {
if ((params_->b_const_ == false) || IsRepack()) {
if (RET_OK != InitBufferB()) {
FreeResizeBufA();
return RET_ERROR;

2
mindspore/lite/src/runtime/kernel/arm/fp16/matmul_base_fp16.h
View File

@ -42,7 +42,7 @@ class MatmulBaseFP16CPUKernel : public InnerKernel {
protected:
void InitParameter();
bool is_repack() { return is_repack_; }
bool IsRepack() { return is_repack_; }
bool is_repack_ = false;
private:

4
mindspore/lite/src/runtime/kernel/arm/fp16/matmul_fp16.cc
View File

@ -86,7 +86,7 @@ int MatmulFP16CPUKernel::ReSize() {
}
int MatmulFP16CPUKernel::Run() {
if (is_trainable() && (IsTrain())) {
if (IsTrainable() && (IsTrain())) {
is_repack_ = true;
}
auto ret = MatmulBaseFP16CPUKernel::Run();
@ -99,7 +99,7 @@ int MatmulFP16CPUKernel::Run() {
int MatmulFP16CPUKernel::Eval() {
InnerKernel::Eval();
if (is_trainable()) {
if (IsTrainable()) {
is_repack_ = true;
}
return RET_OK;

3
mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic_fp32.cc
View File

@ -76,6 +76,9 @@ int ArithmeticCPUKernel::CheckDataType() {
<< in0_dataType << " input 1 dataType: " << in1_dataType;
return RET_ERROR;
}
if (op_parameter_->is_train_session_) {
data_type_len_ = lite::DataTypeSize(in_tensors_.at(0)->data_type());
}
return RET_OK;
}

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_1x1_fp32.cc
View File

@ -242,7 +242,7 @@ int Convolution1x1CPUKernel::Run() {
MS_LOG(ERROR) << "Conv1x1 Malloc pack_input_ error!";
return RET_MEMORY_FAILED;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -294,7 +294,7 @@ int Convolution1x1CPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

6
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_delegate_fp32.h
View File

@ -88,9 +88,9 @@ class ConvolutionDelegateCPUKernel : public InnerKernel {
InnerKernel::Train();
return conv_kernel_->Train();
}
void set_trainable(bool trainable) override {
InnerKernel::set_trainable(trainable);
return conv_kernel_->set_trainable(trainable);
void SetTrainable(bool trainable) override {
InnerKernel::SetTrainable(trainable);
return conv_kernel_->SetTrainable(trainable);
}
protected:

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise_3x3_fp32.cc
View File

@ -119,7 +119,7 @@ int ConvolutionDepthwise3x3CPUKernel::Run() {
return RET_MEMORY_FAILED;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
if (InitWeightBias() != RET_OK) {
ctx_->allocator->Free(buffer_);
MS_LOG(ERROR) << "Convolution depthwise 3x3 run InitWeightBias failed.";
@ -148,7 +148,7 @@ int ConvolutionDepthwise3x3CPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
if (InitWeightBias() != RET_OK) {
MS_LOG(ERROR) << "Convolution depthwise 3x3 fp32 Eval:InitWeightBias failed.";
return RET_ERROR;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise_fp32.cc
View File

@ -105,7 +105,7 @@ int ConvDwRun(void *cdata, int task_id, float lhs_scale, float rhs_scale) {
}
int ConvolutionDepthwiseCPUKernel::Run() {
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -139,7 +139,7 @@ int ConvolutionDepthwiseCPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise_indirect_fp32.cc
View File

@ -194,7 +194,7 @@ int ConvolutionDepthwiseIndirectCPUKernel::Run() {
packed_input_ = input_ptr;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -233,7 +233,7 @@ int ConvolutionDepthwiseIndirectCPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise_slidewindow_fp32.cc
View File

@ -150,7 +150,7 @@ int ConvolutionDepthwiseSWCPUKernel::Run() {
return RET_ERROR;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -207,7 +207,7 @@ int ConvolutionDepthwiseSWCPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_depthwise_slidewindow_x86_fp32.cc
View File

@ -138,7 +138,7 @@ int ConvolutionDepthwiseSWCPUKernelX86::Run() {
return RET_ERROR;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -198,7 +198,7 @@ int ConvolutionDepthwiseSWCPUKernelX86::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_fp32.cc
View File

@ -147,7 +147,7 @@ int ConvolutionCPUKernel::Run() {
FreeTmpBuffer();
return RET_ERROR;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
PackWeight();
}
@ -180,7 +180,7 @@ void ConvolutionCPUKernel::PackWeight() {
int ConvolutionCPUKernel::Eval() {
InnerKernel::Eval();
if (is_trainable()) {
if (IsTrainable()) {
PackWeight();
}
return RET_OK;

4
mindspore/lite/src/runtime/kernel/arm/fp32/convolution_winograd_fp32.cc
View File

@ -216,7 +216,7 @@ int ConvolutionWinogradCPUKernel::Run() {
FreeTmpBuffer();
return RET_ERROR;
}
if (IsTrain() && is_trainable()) {
if (IsTrain() && IsTrainable()) {
ret = InitWeightBias();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Init weight bias failed.";
@ -239,7 +239,7 @@ int ConvolutionWinogradCPUKernel::Eval() {
MS_LOG(ERROR) << "eval failed!";
return ret;
}
if (is_trainable()) {
if (IsTrainable()) {
ret = InitWeightBias();
if (ret != RET_OK) {
MS_LOG(ERROR) << "Init weight bias failed.";

2
mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc
View File

@ -66,7 +66,7 @@ int FusedBatchnormCPUKernel::InitConstTensor() {
int FusedBatchnormCPUKernel::Run() {
auto param = reinterpret_cast<BatchNormParameter *>(op_parameter_);
if (IsTrain() && is_trainable() && in_tensors_.size() >= 5) {
if (IsTrain() && IsTrainable() && in_tensors_.size() >= 5) {
float *in = static_cast<float *>(in_tensors_[0]->MutableData());
float *scale = static_cast<float *>(in_tensors_[1]->MutableData());
float *offset = static_cast<float *>(in_tensors_[2]->MutableData());

3
mindspore/lite/src/scheduler.cc
View File

@ -445,6 +445,7 @@ int Scheduler::FindCpuKernel(const std::vector<Tensor *> &in_tensors, const std:
if (ret == RET_OK) {
MS_LOG(DEBUG) << "Get TypeId(" << kernel_data_type << ") op success: " << PrimitiveCurVersionTypeName(op_type);
if (is_train_session_) {
(*kernel)->Init();
RestoreTensorData(&restored_origin_tensors);
} else {
FreeRestoreTensors(&restored_origin_tensors);
@ -721,7 +722,7 @@ int Scheduler::ScheduleSubGraphToKernels(size_t subgraph_index, std::vector<kern
kernel = SchedulePartialToKernel(node);
} else { // kernel
kernel = ScheduleNodeToKernel(node, prefer_data_type);
if (kernel != nullptr) {
if ((kernel != nullptr) && (!is_train_session_)) {
ret = kernel->Init();
}
}

18
mindspore/lite/src/train/train_session.cc
View File

@ -61,6 +61,10 @@ TrainSession::TrainSession() {
int TrainSession::Init(const Context *context, const TrainCfg *train_cfg) {
if (train_cfg != nullptr) {
if (train_cfg->mix_precision_cfg_.loss_scale_ <= 0) {
MS_LOG(ERROR) << "illegal loss scale configuration";
return RET_NULL_PTR;
}
cfg_ = *train_cfg;
}
return lite::LiteSession::Init(context);
@ -115,6 +119,9 @@ void TrainSession::FreeWorkSpace() {
int TrainSession::InitCallBack() {
sched_mix_precision_callback_ = [&](const Model::Node *node) {
if (!context_->IsCpuFloat16Enabled()) {
return false;
}
auto node_type = GetPrimitiveType(node->primitive_);
if (node_type == schema::PrimitiveType_Cast) {
return false;
@ -548,7 +555,7 @@ void TrainSession::CompileOptimizedKernels() {
if (!IsOptimizer(kernel)) {
for (auto it : kernel->in_tensors()) {
if (std::find(out_tensor.begin(), out_tensor.end(), it) != out_tensor.end()) {
kernel->set_trainable(true);
kernel->SetTrainable(true);
break;
}
}
@ -610,7 +617,7 @@ int TrainSession::AdminSetupVirtualBatch(int virtual_batch_multiplier, float lr,
}
}
if (IsBN(kernel) && kernel->is_trainable()) {
if (IsBN(kernel) && kernel->IsTrainable()) {
auto batchnorm = static_cast<kernel::BatchnormCPUKernel *>(kernel->kernel());
auto ret = RET_OK;
if (mod == kernel::OptimizerKernel::WeightUpdateMode::VIRTUAL_BATCH) {
@ -711,13 +718,6 @@ int TrainSession::Export(const std::string &file_name, ModelType model_type, Qua
session::LiteSession *session::LiteSession::CreateTrainSession(const std::string &fn, const lite::Context *context,
bool train_mode, const lite::TrainCfg *cfg) {
if (cfg != nullptr) {
// test legal configuration
if (cfg->mix_precision_cfg_.loss_scale_ <= 0) {
MS_LOG(ERROR) << "illegal loss scale configuration";
return nullptr;
}
}
auto session = std::make_unique<lite::TrainSession>();
if (session == nullptr) {
MS_LOG(ERROR) << "create session failed";

10
mindspore/lite/test/config/models_ms_train.cfg
View File

@ -23,15 +23,15 @@ effnet_tune weight_quant 7
googlenet weight_quant 10
densenet weight_quant 11
shufflenetv2 weight_quant 3
#mini_alexnet fp16 6
#nin fp16 8.0
mini_alexnet fp16 6
nin fp16 8.0
#lenet fp16 2
mobilenetv1 fp16 2
mobilenetv2 fp16 2
#mobilenetv3 fp16 7
mobilenetv3 fp16 7
effnet fp16 2
#effnet_tune fp16 5.0
effnet_tune fp16 5.0
resnet fp16 2
#googlenet fp16 10
googlenet fp16 10
#xception fp16 20.0
# LAST

12
mindspore/lite/test/st/scripts/run_benchmark_arm.sh
View File

@ -1060,12 +1060,17 @@ function Run_arm64() {
model_prefix=${line_array[0]}
model_name=${line_array[0]}'_train'
accuracy_limit=0.5
enable_fp16="false"
if [[ $model_name == \#* ]]; then
continue
fi
if [[ "${line_array[1]}" == "weight_quant" ]]; then
model_name=${line_array[0]}'_train_quant'
accuracy_limit=${line_array[2]}
elif [[ "${line_array[1]}" == "fp16" ]]; then
enable_fp16="true"
suffix_print="_fp16"
accuracy_limit=${line_array[2]}
fi
export_file="${tmp_dir}/${model_name}_tod"
inference_file="${tmp_dir}/${model_name}_infer"
@ -1086,6 +1091,7 @@ function Run_arm64() {
--expectedDataFile=${tmp_dir}/${model_prefix}_output \
--numThreads=${threads} \
--accuracyThreshold=${accuracy_limit} \
--enableFp16=${enable_fp16} \
--inferenceFile=${inference_file} \
--exportFile=${export_file}
ENDM
@ -1095,9 +1101,9 @@ ENDM
adb -s ${device_id} shell < adb_run_cmd.txt >> ${run_arm64_fp32_log_file}
# TODO: change to arm_type
if [ $? = 0 ]; then
run_result='arm64_train: '${model_name}' pass'; echo ${run_result} >> ${run_benchmark_train_result_file}
run_result='arm64_train: '${model_name}''${suffix_print}' pass'; echo ${run_result} >> ${run_benchmark_train_result_file}
else
run_result='arm64_train: '${model_name}' failed'; echo ${run_result} >> ${run_benchmark_train_result_file};
run_result='arm64_train: '${model_name}''${suffix_print}' failed'; echo ${run_result} >> ${run_benchmark_train_result_file};
fail=1
fi
done < ${models_ms_train_config}
@ -1183,6 +1189,8 @@ function Run_arm32() {
if [[ "${line_array[1]}" == "weight_quant" ]]; then
model_name=${line_array[0]}'_train_quant'
accuracy_limit=${line_array[2]}
elif [[ "${line_array[1]}" != "" ]]; then
continue
fi
export_file="${tmp_dir}/${model_name}_tod"
inference_file="${tmp_dir}/${model_name}_infer"

2
mindspore/lite/test/st/scripts/run_benchmark_x86.sh
View File

@ -737,6 +737,8 @@ function Run_x86() {
if [[ "${line_array[1]}" == "weight_quant" ]]; then
model_name=${line_array[0]}'_train_quant'
accuracy_limit=${line_array[2]}
elif [[ "${line_array[1]}" != "" ]]; then
continue
fi
export_file="${ms_train_models_path}/${model_name}_tod"
inference_file="${ms_train_models_path}/${model_name}_infer"

2
mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc
View File

@ -194,7 +194,7 @@ TEST_F(TestBNGradFp32, BNTtrainFp32) {
EXPECT_EQ(0, ret);
kernel_obj->Train();
kernel_obj->set_trainable(true);
kernel_obj->SetTrainable(true);
kernel_obj->Run();
std::cout << "================save_mean==============================\n";