xxq250
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

!23089 fix config parse bug && full quant bug 

Merge pull request !23089 from yeyunpeng2020/quant
This commit is contained in:
i-robot 2021-09-08 12:26:57 +00:00 committed by Gitee
parent b5ba7b1a83 a274d915c3
commit a0829c11e5
10 changed files with 33 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
mindspore/lite/micro/coder/opcoders/nnacl/int8/sigmoid_int8_coder.cc
View File

@ -25,11 +25,10 @@
namespace mindspore::lite::micro::nnacl {
constexpr auto kInt8Range = 256;
void CalculateTableList(int8_t *table, const float input_scale, const int32_t input_zp) {
void CalculateTableList(int8_t *table, const float input_scale, const int32_t input_zp, const float output_scale,
const int32_t output_zp) {
constexpr int32_t min_value = std::numeric_limits<int8_t>::min();
constexpr int32_t max_value = std::numeric_limits<int8_t>::max();
constexpr float output_scale = 1.0f / kInt8Range;
constexpr int32_t output_zp = std::numeric_limits<int8_t>::min();
for (int i = min_value; i < max_value; ++i) {
const float real_input_value = input_scale * (i - input_zp);
@ -49,12 +48,7 @@ int SigmodInt8Coder::Prepare(CoderContext *const context) {
const int32_t input_zp = input_tensor_->quant_params().at(0).zeroPoint;
const float output_scale = output_tensor_->quant_params().at(0).scale;
const int32_t output_zp = output_tensor_->quant_params().at(0).zeroPoint;
if (output_scale != (1.0f / kInt8Range) || output_zp != static_cast<int32_t>(std::numeric_limits<int8_t>::min())) {
MS_LOG(ERROR) << "Output scale is : " << output_scale << ", should be 1/kInt8Range. Output zp is : " << output_zp
<< ", should be -128.";
return RET_ERROR;
}
CalculateTableList(table_list_, input_scale, input_zp);
CalculateTableList(table_list_, input_scale, input_zp, output_scale, output_zp);
return RET_OK;
}

7
mindspore/lite/src/runtime/kernel/arm/int8/reduce_int8.cc
View File

@ -560,14 +560,19 @@ int ReduceInt8CPUKernel::CallReduceUnit(int task_id) {
MS_LOG(ERROR) << "Input data of reduce int8 operator is null.";
return RET_NULL_PTR;
}
if (!is_last_axis_) {
if (dst_data_ == nullptr) {
MS_LOG(ERROR) << "Output data of reduce int8 operator is null.";
return RET_NULL_PTR;
}
if (!is_last_axis_) {
return reducer_(outer_size_, inner_size_, axis_size_, src_data_, dst_data_, &quant_arg_, task_id,
op_parameter_->thread_num_);
} else {
if (last_dst_data_ == nullptr) {
MS_LOG(ERROR) << "Output data of reduce int8 operator is null.";
return RET_NULL_PTR;
}
return last_reducer_(outer_size_, inner_size_, axis_size_, src_data_, last_dst_data_, &quant_arg_, task_id,
op_parameter_->thread_num_);
}

7
mindspore/lite/src/runtime/kernel/arm/int8/sigmoid_int8.cc
View File

@ -30,11 +30,10 @@ using mindspore::lite::RET_OK;
using mindspore::schema::ActivationType_SIGMOID;
namespace mindspore::kernel {
void CalculateTableList(int8_t *table, const float input_scale, const int32_t input_zp) {
void CalculateTableList(int8_t *table, const float input_scale, const int32_t input_zp, const float output_scale,
const int32_t output_zp) {
int32_t min_value = std::numeric_limits<int8_t>::min();
int32_t max_value = std::numeric_limits<int8_t>::max();
const float output_scale = 1.0f / 256;
const int32_t output_zp = -128;
for (int i = min_value; i < max_value; ++i) {
const float real_input_value = input_scale * (i - input_zp);
@ -58,7 +57,7 @@ int SigmoidInt8CPUKernel::Init() {
<< ", should be -128.";
return RET_ERROR;
}
CalculateTableList(table_list_, input_scale, input_zp);
CalculateTableList(table_list_, input_scale, input_zp, output_scale, output_zp);
return RET_OK;
}

2
mindspore/lite/test/config/quant/ml_face_mnet_image_posttraining.config
View File

@ -18,7 +18,7 @@ normalize_std=[1, 1, 1]
# Image resize
resize_width=640
resize_height=640
# Resize method supports LINEAR or NEARST or CUBIC
# Resize method supports LINEAR or NEAREST or CUBIC
resize_method=LINEAR
[full_quant_param]

3
mindspore/lite/tools/common/parse_config_utils.cc
View File

@ -78,6 +78,9 @@ int SplitLineToMap(std::ifstream *ifs, std::map<std::string, std::map<std::strin
return RET_ERROR;
}
if (raw_line.empty()) {
continue;
}
if (raw_line.at(0) == '[') {
section = raw_line.substr(1, raw_line.size() - 2);
continue;

7
mindspore/lite/tools/converter/config_parser/quant_param_parser.cc
View File

@ -22,6 +22,7 @@ namespace mindspore {
namespace lite {
namespace {
constexpr int kQuantBitNumInt16 = 16;
constexpr int kQuantBitNumInt8 = 8;
} // namespace
int QuantParamParser::ParseCommonQuant(const CommonQuantString &common_quant_string,
quant::CommonQuantParam *common_quant) {
@ -39,12 +40,12 @@ int QuantParamParser::ParseCommonQuant(const CommonQuantString &common_quant_str
}
if (common_quant->quant_type == schema::QuantType_WeightQuant) {
if (common_quant->bit_num < 0 || common_quant->bit_num > kQuantBitNumInt16) {
MS_LOG(ERROR) << "INPUT ILLEGAL: bit_num should be greater than zero and less than 16 currently.";
MS_LOG(ERROR) << "INPUT ILLEGAL: bit_num should be [0,16].";
return RET_INPUT_PARAM_INVALID;
}
} else if (common_quant->quant_type == schema::QuantType_PostTraining) {
if (common_quant->bit_num <= 0 || common_quant->bit_num > kQuantBitNumInt16) {
MS_LOG(ERROR) << "INPUT ILLEGAL: bit_num should be greater or equal to zero and less than 16 currently.";
if (common_quant->bit_num <= 0 || common_quant->bit_num > kQuantBitNumInt8) {
MS_LOG(ERROR) << "INPUT ILLEGAL: bit_num should be [1,8].";
return RET_INPUT_PARAM_INVALID;
}
}

2
mindspore/lite/tools/converter/preprocess/opencv_utils.cc
View File

@ -51,7 +51,7 @@ cv::InterpolationFlags ConvertResizeMethod(const std::string &method) {
} else if (method == "CUBIC") {
return cv::INTER_CUBIC;
} else {
MS_LOG(ERROR) << "Unsupported resize method:" << method;
MS_LOG(ERROR) << "INPUT ILLEGAL: resize_method must be NEAREST|LINEAR|CUBIC.";
return cv::INTER_MAX;
}
}

2
mindspore/lite/tools/converter/quantizer/config/full_quant.cfg
View File

@ -29,7 +29,7 @@ normalize_std=[127.5, 127.5, 127.5]
# Image resize
resize_width=224
resize_height=224
# Resize method supports LINEAR or NEARST or CUBIC
# Resize method supports LINEAR or NEAREST or CUBIC
resize_method=LINEAR
# Image center crop
center_crop_width=224

11
mindspore/lite/tools/converter/quantizer/quantize_util.cc
View File

@ -23,6 +23,7 @@
#include <memory>
#include <vector>
#include <set>
#include <functional>
#include "include/version.h"
#include "ops/concat.h"
#include "ops/crop.h"
@ -339,13 +340,13 @@ static bool SearchLowerBound(const std::vector<float> &data, const size_t &index
return false;
}
if (fabs(max_tmp - *min_tmp) <= 0.0f || fabs(length - *min_idx) <= 0.0f) {
MS_LOG(ERROR) << "divisor cannot be 0";
MS_LOG(INFO) << "divisor cannot be 0";
return false;
}
float range_ratio = (data.at(index) - *min_tmp) / (max_tmp - *min_tmp);
float index_ratio = static_cast<float>(index - *min_idx) / (length - *min_idx);
if (fabs(index_ratio) <= 0.0f) {
MS_LOG(ERROR) << "divisor cannot be 0";
MS_LOG(INFO) << "divisor cannot be 0";
return false;
}
if (index_ratio > 0 && range_ratio / index_ratio > ratio) {
@ -362,13 +363,13 @@ static bool SearchUpperBound(const std::vector<float> &data, const size_t &index
return false;
}
if (fabs(*max_tmp - min_tmp) <= 0.0f || fabs(length - *max_idx) <= 0.0f) {
MS_LOG(ERROR) << "divisor cannot be 0";
MS_LOG(INFO) << "divisor cannot be 0";
return false;
}
float range_ratio = (*max_tmp - data.at(index)) / (*max_tmp - min_tmp);
float index_ratio = static_cast<float>(index - *max_idx) / (length - *max_idx);
if (fabs(index_ratio) <= 0.0f) {
MS_LOG(ERROR) << "divisor cannot be 0";
MS_LOG(INFO) << "divisor cannot be 0";
return false;
}
if (index_ratio > 0 && range_ratio / index_ratio > ratio) {
@ -384,8 +385,10 @@ static float CalPercentile(const std::vector<float> &data, const int &outlier_pe
int index = std::ceil(val);
float result;
if (index - val > 0) {
MS_ASSERT(index - 1 >= 0);
result = data.at(index - 1);
} else {
MS_ASSERT(index - 1 >= 0);
result = (data.at(index - 1) + data.at(index)) / 2;
}
return result;

2
mindspore/lite/tools/converter/quantizer/quantize_util.h
View File

@ -57,7 +57,7 @@ enum WeightQuantType {
constexpr size_t kUint8Quantization = 8;
constexpr size_t kMaxBit = 8;
constexpr size_t kMaxNum1024 = 1024;
constexpr size_t kPercentBase = 100;
constexpr float kPercentBase = 100.0;
constexpr size_t kMillisecondsBase = 10;
constexpr size_t kWightIndex = 1;
constexpr double kScaleThreashold = 1e-38;