• Hardware Platform : dGPU
• DeepStream Version: 7.0
• TensorRT Version: 8.6.1.6
**• NVIDIA GPU Driver Version: 555.58.02 **
• Issue Type: Question
Currently my pipeline look like this:
pgie -> nvdspreprocess -> sgie
This is my custom preprocessing code for nvdspreprocess. I have visualized my image and everything works, and the cuda copy doesnt give me any error.
NvDsPreProcessStatus
CustomTensorPreparation(CustomCtx *ctx, NvDsPreProcessBatch *batch, NvDsPreProcessCustomBuf *&buf,
CustomTensorParams &tensorParam, NvDsPreProcessAcquirer *acquirer)
{
NvDsPreProcessStatus status = NVDSPREPROCESS_TENSOR_NOT_READY;
/** acquire a buffer from tensor pool */
buf = acquirer->acquire();
// size_t buffer_offset = 0; // Offset within the buffer
float * pDst = (float*)buf->memory_ptr;
int units = batch->units.size();
for(int i = 0; i < units; i++)
{
printf("Preprocessing");
//unsigned char* raw_data = (unsigned char*)batch->units[i].converted_frame_ptr;
NvBufSurfaceParams *surf_params = batch->units[i].roi_meta.converted_buffer;
switch (surf_params->colorFormat) {
case NVBUF_COLOR_FORMAT_RGBA:
std::cout << "Color Format: RGBA" << std::endl;
break;
case NVBUF_COLOR_FORMAT_RGB:
std::cout << "Color Format: RGB" << std::endl;
break;
// Add cases for other color formats as needed
default:
std::cout << "Color Format: Unknown" << std::endl;
break;
}
// dim3 blockDim(16, 16);
// dim3 gridDim((width + blockDim.x - 1) / blockDim.x,
// (height + blockDim.y - 1) / blockDim.y);
// NvDsPreProcessNormalizeInPlaceKernel<<<gridDim, blockDim>>>(
// surf_params->dataPtr, width, height, pitch);
NvDsRoiMeta roi_meta = batch->units[i].roi_meta;
NvDsObjectMeta *obj_meta = (NvDsObjectMeta *)batch->units[i].roi_meta.object_meta;
printf("Object ID: %d\n", obj_meta->unique_component_id);
float x_scale = roi_meta.scale_ratio_x;
float y_scale = roi_meta.scale_ratio_y;
float left = obj_meta->rect_params.left;
float top = obj_meta->rect_params.top;
int width = obj_meta->rect_params.width;
int height = obj_meta->rect_params.height;
printf("Tensor buffer size: %d",te)
printf("retrieved array:%ld",obj_meta->misc_obj_info[0]);
printf("left: %f\n", left);
printf("top: %f\n", top);
printf("width: %d\n", width);
printf("height: %d\n", height);
printf("Landmark height: %d\n", obj_meta->mask_params.height);
printf("Landmark width: %d\n", obj_meta->mask_params.width);
printf("Landmark size: %d\n", obj_meta->mask_params.size);
if (obj_meta->mask_params.data != nullptr && obj_meta->mask_params.size > 0) {
float* mask_data = obj_meta->mask_params.data;
unsigned int mask_size = obj_meta->mask_params.size / sizeof(float); // Number of float elements
for (unsigned int i = 0; i < mask_size; ++i) {
float value = mask_data[i];
printf("Landmark value: %f\n",value);
// Process the mask value as needed
}
cv::Point2f leftEye(mask_data[0]* x_scale,mask_data[1] * y_scale);
cv::Point2f rightEye(mask_data[2] * x_scale,mask_data[3]* y_scale);
cv::Mat frame(surf_params->height, surf_params->width,CV_8UC4 ,surf_params->dataPtr, surf_params->pitch);
cv::Mat image_data(surf_params->height, surf_params->width,CV_8UC4 );
frame.copyTo(image_data);
cv::Mat outptut_image(surf_params->height, surf_params->width,CV_8UC3 );
cv::cvtColor(image_data,outptut_image, cv::COLOR_RGBA2BGR);
cv::Mat alignedFace = alignFace(outptut_image, leftEye, rightEye);
cv::imwrite("output.jpg", alignedFace);
cv::Mat floatAlignedFace;
alignedFace.convertTo(floatAlignedFace, CV_32FC3, 1.0 / 255.0);
printf("Start copying \n");
void *outPtr = (void*)((float*)pDst + i * 3
* 160 * 160
* sizeof(float));
size_t image_size =3 * 160 * 160 * sizeof(float);
printf("Start copying \n");
// Log the pointer and size for verification
printf("outPtr: %p\n", outPtr);
printf("image_size: %zu bytes\n", image_size);
printf("floatAlignedFace.data: %p\n", floatAlignedFace.data);
cudaMemcpy(outPtr, floatAlignedFace.data, image_size, cudaMemcpyHostToDevice);
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
printf("CUDA memcpy error: %s\n", cudaGetErrorString(err));
} else {
printf("CUDA memcpy successful.\n");
}
}
}
tensorParam.params.network_input_shape[0] = (int)batch->units.size();
status = NVDSPREPROCESS_SUCCESS;
return status;
}
Here is my config:
face_embedding_preprocess.txt (2.7 KB)
face_embedding_config.txt (822 Bytes)
I also have setted input-tensor-meta to 1.
sgie_face_embedding = make_element("nvinfer", "face_embedding")
sgie_face_embedding.set_property('config-file-path', "Face/face_embedding_config.txt")
sgie_face_embedding.set_property("input-tensor-meta", 1)
sgie_face_embedding.set_property("output-tensor-meta", 1)
There is no prediction from sgie (no tensor attached to user meta in sgie’s src_pad). Note that my pipeline still works without nvdspreprocess. I wonder if there is anything wrong with my config or my nvdspreprocess code.
