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

DeepStream 7.1 + Fixes + New model output format

Browse Source
This commit is contained in:
Marcos Luciano
2024-11-07 11:25:17 -03:00
parent bca9e59d07
commit b451b036b2
75 changed files with 2383 additions and 1113 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
nvdsinfer_custom_impl_Yolo/yoloForward.cu
View File

@@ -7,8 +7,8 @@
inline __device__ float sigmoidGPU(const float& x) { return 1.0f / (1.0f + __expf(-x)); }
__global__ void gpuYoloLayer(const float* input, float* boxes, float* scores, float* classes, const uint netWidth,
const uint netHeight, const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses, const uint numBBoxes,
__global__ void gpuYoloLayer(const float* input, float* output, const uint netWidth, const uint netHeight,
const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses, const uint numBBoxes,
const uint64_t lastInputSize, const float scaleXY, const float* anchors, const int* mask)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
@@ -50,22 +50,22 @@ __global__ void gpuYoloLayer(const float* input, float* boxes, float* scores, fl
int count = numGridCells * z_id + bbindex + lastInputSize;
boxes[count * 4 + 0] = xc;
boxes[count * 4 + 1] = yc;
boxes[count * 4 + 2] = w;
boxes[count * 4 + 3] = h;
scores[count] = maxProb * objectness;
classes[count] = (float) maxIndex;
output[count * 6 + 0] = xc - w * 0.5;
output[count * 6 + 1] = yc - h * 0.5;
output[count * 6 + 2] = xc + w * 0.5;
output[count * 6 + 3] = yc + h * 0.5;
output[count * 6 + 4] = maxProb * objectness;
output[count * 6 + 5] = (float) maxIndex;
}
cudaError_t cudaYoloLayer(const void* input, void* boxes, void* scores, void* classes, const uint& batchSize,
const uint64_t& inputSize, const uint64_t& outputSize, const uint64_t& lastInputSize, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses, const uint& numBBoxes,
cudaError_t cudaYoloLayer(const void* input, void* output, const uint& batchSize, const uint64_t& inputSize,
const uint64_t& outputSize, const uint64_t& lastInputSize, 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* boxes, void* scores, void* classes, const uint& batchSize,
const uint64_t& inputSize, const uint64_t& outputSize, const uint64_t& lastInputSize, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses, const uint& numBBoxes,
cudaError_t cudaYoloLayer(const void* input, void* output, const uint& batchSize, const uint64_t& inputSize,
const uint64_t& outputSize, const uint64_t& lastInputSize, 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);
@@ -75,9 +75,7 @@ cudaError_t cudaYoloLayer(const void* input, void* boxes, void* scores, void* cl
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<float*> (boxes) + (batch * 4 * outputSize),
reinterpret_cast<float*> (scores) + (batch * 1 * outputSize),
reinterpret_cast<float*> (classes) + (batch * 1 * outputSize),
reinterpret_cast<float*> (output) + (batch * 6 * outputSize),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, lastInputSize, scaleXY,
reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask));
}