thebears/deepstream_yolo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Big update

Browse Source
This commit is contained in:
Marcos Luciano
2023-05-19 03:05:43 -03:00
parent 68f762d5bd
commit 07feae9509
86 changed files with 1523 additions and 5223 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
nvdsinfer_custom_impl_Yolo/yoloForward.cu
View File

@@ -4,13 +4,13 @@
*/
#include <stdint.h>
#include <stdio.h>
inline __device__ float sigmoidGPU(const float& x) { return 1.0f / (1.0f + __expf(-x)); }
__global__ void gpuYoloLayer(const float* input, int* num_detections, float* detection_boxes, float* detection_scores,
int* detection_classes, const float scoreThreshold, const uint netWidth, const uint netHeight, const uint gridSizeX,
const uint gridSizeY, const uint numOutputClasses, const uint numBBoxes, const float scaleXY, const float* anchors,
const int* mask)
__global__ void gpuYoloLayer(const float* input, float* output, int* count, const uint netWidth, const uint netHeight,
const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses, const uint numBBoxes, const float scaleXY,
const float* anchors, const int* mask)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
uint y_id = blockIdx.y * blockDim.y + threadIdx.y;
@@ -24,18 +24,13 @@ __global__ void gpuYoloLayer(const float* input, int* num_detections, float* det
const float objectness = sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
if (objectness < scoreThreshold)
return;
int count = (int)atomicAdd(num_detections, 1);
const float alpha = scaleXY;
const float beta = -0.5 * (scaleXY - 1);
float x = (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) * alpha + beta + x_id)
float xc = (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) * alpha + beta + x_id)
* netWidth / gridSizeX;
float y = (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) * alpha + beta + y_id)
float yc = (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) * alpha + beta + y_id)
* netHeight / gridSizeY;
float w = __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]) * anchors[mask[z_id] * 2];
@@ -53,23 +48,26 @@ __global__ void gpuYoloLayer(const float* input, int* num_detections, float* det
}
}
detection_boxes[count * 4 + 0] = x - 0.5 * w;
detection_boxes[count * 4 + 1] = y - 0.5 * h;
detection_boxes[count * 4 + 2] = x + 0.5 * w;
detection_boxes[count * 4 + 3] = y + 0.5 * h;
detection_scores[count] = objectness * maxProb;
detection_classes[count] = maxIndex;
int _count = (int)atomicAdd(count, 1);
output[_count * 7 + 0] = xc;
output[_count * 7 + 1] = yc;
output[_count * 7 + 2] = w;
output[_count * 7 + 3] = h;
output[_count * 7 + 4] = maxProb;
output[_count * 7 + 5] = maxIndex;
output[_count * 7 + 6] = objectness;
}
cudaError_t cudaYoloLayer(const void* input, void* num_detections, void* detection_boxes, void* detection_scores,
void* detection_classes, const uint& batchSize, uint64_t& inputSize, uint64_t& outputSize, const float& scoreThreshold,
const uint& netWidth, const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses,
const uint& numBBoxes, const float& scaleXY, const void* anchors, const void* mask, cudaStream_t stream);
cudaError_t cudaYoloLayer(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize,
uint64_t& outputSize, const uint& netWidth, const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes, const float& scaleXY, const void* anchors, const void* mask,
cudaStream_t stream);
cudaError_t cudaYoloLayer(const void* input, void* num_detections, void* detection_boxes, void* detection_scores,
void* detection_classes, const uint& batchSize, uint64_t& inputSize, uint64_t& outputSize, const float& scoreThreshold,
const uint& netWidth, const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses,
const uint& numBBoxes, const float& scaleXY, const void* anchors, const void* mask, cudaStream_t stream)
cudaError_t cudaYoloLayer(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize,
uint64_t& outputSize, const uint& netWidth, const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes, const float& scaleXY, const void* anchors, const void* mask,
cudaStream_t stream)
{
dim3 threads_per_block(16, 16, 4);
dim3 number_of_blocks((gridSizeX / threads_per_block.x) + 1, (gridSizeY / threads_per_block.y) + 1,
@@ -77,12 +75,11 @@ cudaError_t cudaYoloLayer(const void* input, void* num_detections, void* detecti
for (unsigned int batch = 0; batch < batchSize; ++batch) {
gpuYoloLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*>(input) + (batch * inputSize), reinterpret_cast<int*>(num_detections) + (batch),
reinterpret_cast<float*>(detection_boxes) + (batch * 4 * outputSize),
reinterpret_cast<float*>(detection_scores) + (batch * outputSize),
reinterpret_cast<int*>(detection_classes) + (batch * outputSize), scoreThreshold, netWidth, netHeight, gridSizeX,
gridSizeY, numOutputClasses, numBBoxes, scaleXY, reinterpret_cast<const float*>(anchors),
reinterpret_cast<const int*>(mask));
reinterpret_cast<const float*> (input) + (batch * inputSize),
reinterpret_cast<float*> (output) + (batch * 7 * outputSize),
reinterpret_cast<int*> (count) + (batch),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY,
reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask));
}
return cudaGetLastError();
}