apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-07 01:22:59 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[CVCUDA] PaddleDetection preprocessor support CV-CUDA (#1493)

Browse Source 
* ppdet preproc use manager

* pad_to_size chw opencv

* pad_to_size chw flycv

* fix pad_to_size flycv

* add warning message

* cvcuda convert cubic to linear, padToSize cvcuda

* stridedpad cvcuda

* fix flycv include

* fix flycv include

* fix flycv build

* cast cvcuda

* fix pybind

* fix normalize permute cuda

* base processor move funcs to cc

* Update pad_to_size.cc
This commit is contained in:
Wang Xinyu
2023-03-10 12:43:57 +08:00
committed by GitHub
parent 9ee2118e1f
commit cb7c8a07d4
23 changed files with 537 additions and 239 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
fastdeploy/vision/detection/ppdet/preprocessor.cc
View File

@@ -129,13 +129,13 @@ bool PaddleDetPreprocessor::BuildPreprocessPipelineFromConfig() {
return true;
}
bool PaddleDetPreprocessor::Run(std::vector<FDMat>* images,
std::vector<FDTensor>* outputs) {
bool PaddleDetPreprocessor::Apply(FDMatBatch* image_batch,
std::vector<FDTensor>* outputs) {
if (!initialized_) {
FDERROR << "The preprocessor is not initialized." << std::endl;
return false;
}
if (images->empty()) {
if (image_batch->mats->empty()) {
FDERROR << "The size of input images should be greater than 0."
<< std::endl;
return false;
@@ -146,7 +146,7 @@ bool PaddleDetPreprocessor::Run(std::vector<FDMat>* images,
// So preprocessor will output the 3 FDTensors, and how to use `im_shape`
// is decided by the model itself
outputs->resize(3);
int batch = static_cast<int>(images->size());
int batch = static_cast<int>(image_batch->mats->size());
// Allocate memory for scale_factor
(*outputs)[1].Resize({batch, 2}, FDDataType::FP32);
// Allocate memory for im_shape
@@ -158,63 +158,51 @@ bool PaddleDetPreprocessor::Run(std::vector<FDMat>* images,
auto* scale_factor_ptr =
reinterpret_cast<float*>((*outputs)[1].MutableData());
auto* im_shape_ptr = reinterpret_cast<float*>((*outputs)[2].MutableData());
for (size_t i = 0; i < images->size(); ++i) {
int origin_w = (*images)[i].Width();
int origin_h = (*images)[i].Height();
for (size_t i = 0; i < image_batch->mats->size(); ++i) {
FDMat* mat = &(image_batch->mats->at(i));
int origin_w = mat->Width();
int origin_h = mat->Height();
scale_factor_ptr[2 * i] = 1.0;
scale_factor_ptr[2 * i + 1] = 1.0;
for (size_t j = 0; j < processors_.size(); ++j) {
if (!(*(processors_[j].get()))(&((*images)[i]))) {
if (!(*(processors_[j].get()))(mat)) {
FDERROR << "Failed to processs image:" << i << " in "
<< processors_[i]->Name() << "." << std::endl;
<< processors_[j]->Name() << "." << std::endl;
return false;
}
if (processors_[j]->Name().find("Resize") != std::string::npos) {
scale_factor_ptr[2 * i] = (*images)[i].Height() * 1.0 / origin_h;
scale_factor_ptr[2 * i + 1] = (*images)[i].Width() * 1.0 / origin_w;
scale_factor_ptr[2 * i] = mat->Height() * 1.0 / origin_h;
scale_factor_ptr[2 * i + 1] = mat->Width() * 1.0 / origin_w;
}
}
if ((*images)[i].Height() > max_hw[0]) {
max_hw[0] = (*images)[i].Height();
if (mat->Height() > max_hw[0]) {
max_hw[0] = mat->Height();
}
if ((*images)[i].Width() > max_hw[1]) {
max_hw[1] = (*images)[i].Width();
if (mat->Width() > max_hw[1]) {
max_hw[1] = mat->Width();
}
im_shape_ptr[2 * i] = max_hw[0];
im_shape_ptr[2 * i + 1] = max_hw[1];
}
// Concat all the preprocessed data to a batch tensor
std::vector<FDTensor> im_tensors(images->size());
for (size_t i = 0; i < images->size(); ++i) {
if ((*images)[i].Height() < max_hw[0] || (*images)[i].Width() < max_hw[1]) {
// if the size of image less than max_hw, pad to max_hw
FDTensor tensor;
(*images)[i].ShareWithTensor(&tensor);
function::Pad(tensor, &(im_tensors[i]),
{0, 0, max_hw[0] - (*images)[i].Height(),
max_hw[1] - (*images)[i].Width()},
0);
} else {
// No need pad
(*images)[i].ShareWithTensor(&(im_tensors[i]));
// if the size of image less than max_hw, pad to max_hw
for (size_t i = 0; i < image_batch->mats->size(); ++i) {
FDMat* mat = &(image_batch->mats->at(i));
if (mat->Height() < max_hw[0] || mat->Width() < max_hw[1]) {
pad_op_->SetWidthHeight(max_hw[1], max_hw[0]);
(*pad_op_)(mat);
}
// Reshape to 1xCxHxW
im_tensors[i].ExpandDim(0);
}
if (im_tensors.size() == 1) {
// If there's only 1 input, no need to concat
// skip memory copy
(*outputs)[0] = std::move(im_tensors[0]);
} else {
// Else concat the im tensor for each input image
// compose a batched input tensor
function::Concat(im_tensors, &((*outputs)[0]), 0);
}
// Get the NCHW tensor
FDTensor* tensor = image_batch->Tensor();
(*outputs)[0].SetExternalData(tensor->Shape(), tensor->Dtype(),
tensor->Data(), tensor->device,
tensor->device_id);
return true;
}
void PaddleDetPreprocessor::DisableNormalize() {
this->disable_normalize_ = true;
// the DisableNormalize function will be invalid if the configuration file is
@@ -224,6 +212,7 @@ void PaddleDetPreprocessor::DisableNormalize() {
<< std::endl;
}
}
void PaddleDetPreprocessor::DisablePermute() {
this->disable_permute_ = true;
// the DisablePermute function will be invalid if the configuration file is