You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

323 lines
14 KiB
C++

/* Copyright 2021 iwatake2222
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
/*** Include ***/
/* for general */
#include <cstdint>
#include <cstdlib>
#include <cmath>
#include <cstring>
#include <string>
#include <vector>
#include <array>
#include <algorithm>
#include <chrono>
/* for MNN */
#include <MNN/ImageProcess.hpp>
#include <MNN/Interpreter.hpp>
#include <MNN/AutoTime.hpp>
/* for My modules */
#include "inference_helper_log.h"
#include "inference_helper_mnn.h"
/*** Macro ***/
#define TAG "InferenceHelperMnn"
#define PRINT(...) INFERENCE_HELPER_LOG_PRINT(TAG, __VA_ARGS__)
#define PRINT_E(...) INFERENCE_HELPER_LOG_PRINT_E(TAG, __VA_ARGS__)
/*** Function ***/
InferenceHelperMnn::InferenceHelperMnn()
{
num_threads_ = 1;
}
InferenceHelperMnn::~InferenceHelperMnn()
{
}
int32_t InferenceHelperMnn::SetNumThreads(const int32_t num_threads)
{
num_threads_ = num_threads;
return kRetOk;
}
int32_t InferenceHelperMnn::SetCustomOps(const std::vector<std::pair<const char*, const void*>>& custom_ops)
{
PRINT("[WARNING] This method is not supported\n");
return kRetOk;
}
int32_t InferenceHelperMnn::ParameterInitialization(std::vector<InputTensorInfo>& input_tensor_info_list, std::vector<OutputTensorInfo>& output_tensor_info_list) {
/* Check tensor info fits the info from model */
for (auto& input_tensor_info : input_tensor_info_list) {
auto input_tensor = net_->getSessionInput(session_, input_tensor_info.name.c_str());
if (input_tensor == nullptr) {
PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
return kRetErr;
}
if ((input_tensor->getType().code == halide_type_float) && (input_tensor_info.tensor_type == TensorInfo::kTensorTypeFp32)) {
/* OK */
} else if ((input_tensor->getType().code == halide_type_uint) && (input_tensor_info.tensor_type == TensorInfo::kTensorTypeUint8)) {
/* OK */
} else {
PRINT_E("Incorrect input tensor type (%d, %d)\n", input_tensor->getType().code, input_tensor_info.tensor_type);
return kRetErr;
}
if ((input_tensor->channel() != -1) && (input_tensor->height() != -1) && (input_tensor->width() != -1)) {
if (input_tensor_info.GetChannel() != -1) {
if ((input_tensor->channel() == input_tensor_info.GetChannel()) && (input_tensor->height() == input_tensor_info.GetHeight()) && (input_tensor->width() == input_tensor_info.GetWidth())) {
/* OK */
} else {
PRINT_E("W: %d != %d\n", input_tensor->width() , input_tensor_info.GetWidth());
PRINT_E("H: %d != %d\n", input_tensor->height() , input_tensor_info.GetHeight());
PRINT_E("C: %d != %d\n", input_tensor->channel() , input_tensor_info.GetChannel());
PRINT_E("Incorrect input tensor size\n");
return kRetErr;
}
} else {
PRINT("Input tensor size is set from the model\n");
input_tensor_info.tensor_dims.clear();
for (int32_t dim = 0; dim < input_tensor->dimensions(); dim++) {
input_tensor_info.tensor_dims.push_back(input_tensor->length(dim));
}
}
} else {
if (input_tensor_info.GetChannel() != -1) {
PRINT("Input tensor size is resized\n");
/* In case the input size is not fixed */
net_->resizeTensor(input_tensor, { 1, input_tensor_info.GetChannel(), input_tensor_info.GetHeight(), input_tensor_info.GetWidth() });
net_->resizeSession(session_);
} else {
PRINT_E("Model input size is not set\n");
return kRetErr;
}
}
}
for (const auto& output_tensor_info : output_tensor_info_list) {
auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
if (output_tensor == nullptr) {
PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
return kRetErr;
}
/* Output size is set when run inference later */
}
/* Convert normalize parameter to speed up */
for (auto& input_tensor_info : input_tensor_info_list) {
ConvertNormalizeParameters(input_tensor_info);
}
/* Check if tensor info is set */
for (const auto& input_tensor_info : input_tensor_info_list) {
for (const auto& dim : input_tensor_info.tensor_dims) {
if (dim <= 0) {
PRINT_E("Invalid tensor size\n");
return kRetErr;
}
}
}
//for (const auto& output_tensor_info : output_tensor_info_list) {
// for (const auto& dim : output_tensor_info.tensor_dims) {
// if (dim <= 0) {
// PRINT_E("Invalid tensor size\n");
// return kRetErr;
// }
// }
//}
return kRetOk;
}
int32_t InferenceHelperMnn::Initialize(char* model_buffer, int model_size, std::vector<InputTensorInfo>& input_tensor_info_list, std::vector<OutputTensorInfo>& output_tensor_info_list) {
/*** Create network ***/
net_.reset(MNN::Interpreter::createFromBuffer(model_buffer, model_size));
if (!net_) {
PRINT_E("Failed to load model model buffer\n");
return kRetErr;
}
MNN::ScheduleConfig scheduleConfig;
scheduleConfig.type = MNN_FORWARD_CPU;
scheduleConfig.numThread = num_threads_; // it seems, setting 1 has better performance on Android
// MNN::BackendConfig bnconfig;
// bnconfig.power = MNN::BackendConfig::Power_High;
// bnconfig.precision = MNN::BackendConfig::Precision_Low;
// scheduleConfig.backendConfig = &bnconfig;
session_ = net_->createSession(scheduleConfig);
if (!session_) {
PRINT_E("Failed to create session\n");
return kRetErr;
}
return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
}
int32_t InferenceHelperMnn::Initialize(const std::string& model_filename, std::vector<InputTensorInfo>& input_tensor_info_list, std::vector<OutputTensorInfo>& output_tensor_info_list)
{
/*** Create network ***/
net_.reset(MNN::Interpreter::createFromFile(model_filename.c_str()));
if (!net_) {
PRINT_E("Failed to load model file (%s)\n", model_filename.c_str());
return kRetErr;
}
MNN::ScheduleConfig scheduleConfig;
#ifdef IOS_CPU_FORWARD
PRINT("ios use cpu backend.");
scheduleConfig.type = MNN_FORWARD_CPU;
#else
scheduleConfig.type = MNN_FORWARD_AUTO;
#endif
scheduleConfig.numThread = num_threads_; // it seems, setting 1 has better performance on Android
// MNN::BackendConfig bnconfig;
// bnconfig.power = MNN::BackendConfig::Power_High;
// bnconfig.precision = MNN::BackendConfig::Precision_Low;
// scheduleConfig.backendConfig = &bnconfig;
session_ = net_->createSession(scheduleConfig);
if (!session_) {
PRINT_E("Failed to create session\n");
return kRetErr;
}
return ParameterInitialization(input_tensor_info_list, output_tensor_info_list);
};
int32_t InferenceHelperMnn::Finalize(void)
{
net_->releaseSession(session_);
net_->releaseModel();
net_.reset();
out_mat_list_.clear();
return kRetOk;
}
int32_t InferenceHelperMnn::PreProcess(const std::vector<InputTensorInfo>& input_tensor_info_list)
{
for (const auto& input_tensor_info : input_tensor_info_list) {
auto input_tensor = net_->getSessionInput(session_, input_tensor_info.name.c_str());
if (input_tensor == nullptr) {
PRINT_E("Invalid input name (%s)\n", input_tensor_info.name.c_str());
return kRetErr;
}
if (input_tensor_info.data_type == InputTensorInfo::kDataTypeImage) {
/* Crop */
if ((input_tensor_info.image_info.width != input_tensor_info.image_info.crop_width) || (input_tensor_info.image_info.height != input_tensor_info.image_info.crop_height)) {
PRINT_E("Crop is not supported\n");
return kRetErr;
}
MNN::CV::ImageProcess::Config image_processconfig;
/* Convert color type */
if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 3)) {
image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
if (input_tensor_info.image_info.swap_color) {
image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::RGB : MNN::CV::BGR;
} else {
image_processconfig.destFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
}
} else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 1)) {
image_processconfig.sourceFormat = MNN::CV::GRAY;
image_processconfig.destFormat = MNN::CV::GRAY;
} else if ((input_tensor_info.image_info.channel == 3) && (input_tensor_info.GetChannel() == 1)) {
image_processconfig.sourceFormat = (input_tensor_info.image_info.is_bgr) ? MNN::CV::BGR : MNN::CV::RGB;
image_processconfig.destFormat = MNN::CV::GRAY;
} else if ((input_tensor_info.image_info.channel == 1) && (input_tensor_info.GetChannel() == 3)) {
image_processconfig.sourceFormat = MNN::CV::GRAY;
image_processconfig.destFormat = MNN::CV::BGR;
} else {
PRINT_E("Unsupported color conversion (%d, %d)\n", input_tensor_info.image_info.channel, input_tensor_info.GetChannel());
return kRetErr;
}
/* Normalize image */
std::memcpy(image_processconfig.mean, input_tensor_info.normalize.mean, sizeof(image_processconfig.mean));
std::memcpy(image_processconfig.normal, input_tensor_info.normalize.norm, sizeof(image_processconfig.normal));
/* Resize image */
image_processconfig.filterType = MNN::CV::BILINEAR;
MNN::CV::Matrix trans;
trans.setScale(static_cast<float>(input_tensor_info.image_info.crop_width) / input_tensor_info.GetWidth(), static_cast<float>(input_tensor_info.image_info.crop_height) / input_tensor_info.GetHeight());
/* Do pre-process */
std::shared_ptr<MNN::CV::ImageProcess> pretreat(MNN::CV::ImageProcess::create(image_processconfig));
pretreat->setMatrix(trans);
pretreat->convert(static_cast<uint8_t*>(input_tensor_info.data), input_tensor_info.image_info.crop_width, input_tensor_info.image_info.crop_height, 0, input_tensor);
} else if ( (input_tensor_info.data_type == InputTensorInfo::kDataTypeBlobNhwc) || (input_tensor_info.data_type == InputTensorInfo::kDataTypeBlobNchw) ) {
std::unique_ptr<MNN::Tensor> tensor;
if (input_tensor_info.data_type == InputTensorInfo::kDataTypeBlobNhwc) {
tensor.reset(new MNN::Tensor(input_tensor, MNN::Tensor::TENSORFLOW));
} else {
tensor.reset(new MNN::Tensor(input_tensor, MNN::Tensor::CAFFE));
}
if (tensor->getType().code == halide_type_float) {
for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
tensor->host<float>()[i] = static_cast<float*>(input_tensor_info.data)[i];
}
} else {
for (int32_t i = 0; i < input_tensor_info.GetWidth() * input_tensor_info.GetHeight() * input_tensor_info.GetChannel(); i++) {
tensor->host<uint8_t>()[i] = static_cast<uint8_t*>(input_tensor_info.data)[i];
}
}
input_tensor->copyFromHostTensor(tensor.get());
} else {
PRINT_E("Unsupported data type (%d)\n", input_tensor_info.data_type);
return kRetErr;
}
}
return kRetOk;
}
int32_t InferenceHelperMnn::Process(std::vector<OutputTensorInfo>& output_tensor_info_list)
{
net_->runSession(session_);
out_mat_list_.clear();
for (auto& output_tensor_info : output_tensor_info_list) {
auto output_tensor = net_->getSessionOutput(session_, output_tensor_info.name.c_str());
if (output_tensor == nullptr) {
PRINT_E("Invalid output name (%s)\n", output_tensor_info.name.c_str());
return kRetErr;
}
auto dimType = output_tensor->getDimensionType();
std::unique_ptr<MNN::Tensor> outputUser(new MNN::Tensor(output_tensor, dimType));
output_tensor->copyToHostTensor(outputUser.get());
auto type = outputUser->getType();
if (type.code == halide_type_float) {
output_tensor_info.tensor_type = TensorInfo::kTensorTypeFp32;
output_tensor_info.data = outputUser->host<float>();
} else if (type.code == halide_type_uint && type.bytes() == 1) {
output_tensor_info.tensor_type = TensorInfo::kTensorTypeUint8;
output_tensor_info.data = outputUser->host<uint8_t>();
} else {
PRINT_E("Unexpected data type\n");
return kRetErr;
}
output_tensor_info.tensor_dims.clear();
for (int32_t dim = 0; dim < outputUser->dimensions(); dim++) {
output_tensor_info.tensor_dims.push_back(outputUser->length(dim));
}
out_mat_list_.push_back(std::move(outputUser)); // store data in member variable so that data keep exist
}
return kRetOk;
}