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

Update Benchmarks + Add YOLOv7-u6 + Fixes

Browse Source
This commit is contained in:
Marcos Luciano
2023-05-21 02:12:09 -03:00
parent af20c2f72c
commit 79d22283c1
13 changed files with 176 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
README.md
View File

@@ -5,6 +5,8 @@ NVIDIA DeepStream SDK 6.2 / 6.1.1 / 6.1 / 6.0.1 / 6.0 configuration for YOLO mod
------------------------------------- -------------------------------------
### **Big update on DeepStream-Yolo** ### **Big update on DeepStream-Yolo**
------------------------------------- -------------------------------------
### Important: please generate the ONNX model and the TensorRT engine again with the updated files
-------------------------------------
### Future updates ### Future updates
@@ -149,7 +151,7 @@ sample = 1920x1080 video
- Eval - Eval
``` ```
nms-iou-threshold = 0.6 (Darknet) / 0.65 (YOLOv5, YOLOv6, YOLOv7, YOLOR and YOLOX) / 0.7 (Paddle, YOLO-NAS and YOLOv8) nms-iou-threshold = 0.6 (Darknet) / 0.65 (YOLOv5, YOLOv6, YOLOv7, YOLOR and YOLOX) / 0.7 (Paddle, YOLO-NAS, YOLOv8 and YOLOv7-u6)
pre-cluster-threshold = 0.001 pre-cluster-threshold = 0.001
topk = 300 topk = 300
``` ```
@@ -164,40 +166,49 @@ topk = 300
#### Results #### Results
**NOTE**: * = PyTorch **NOTE**: * = PyTorch.
**NOTE**: ** = The YOLOv4 is trained with the trainvalno5k set, so the mAP is high on val2017 test **NOTE**: ** = The YOLOv4 is trained with the trainvalno5k set, so the mAP is high on val2017 test.
**NOTE**: The p3.2xlarge instance (AWS) seems to max out at 625-635 FPS on DeepStream even using lighter models **NOTE**: The p3.2xlarge instance (AWS) seems to max out at 625-635 FPS on DeepStream even using lighter models.
| DeepStream | Precision | Resolution | IoU=0.5:0.95 | IoU=0.5 | IoU=0.75 | FPS<br />(without display) | | DeepStream | Precision | Resolution | IoU=0.5:0.95 | IoU=0.5 | IoU=0.75 | FPS<br />(without display) |
|:----------------:|:---------:|:----------:|:------------:|:-------:|:--------:|:--------------------------:| |:------------------:|:---------:|:----------:|:------------:|:-------:|:--------:|:--------------------------:|
| YOLO-NAS L | FP16 | 640 | 0.484 | 0.658 | 0.532 | 235.27 | | YOLO-NAS L | FP16 | 640 | 0.484 | 0.658 | 0.532 | 235.27 |
| YOLO-NAS M | FP16 | 640 | 0.480 | 0.651 | 0.524 | 287.39 | | YOLO-NAS M | FP16 | 640 | 0.480 | 0.651 | 0.524 | 287.39 |
| YOLO-NAS S | FP16 | 640 | 0.442 | 0.614 | 0.485 | 478.52 | | YOLO-NAS S | FP16 | 640 | 0.442 | 0.614 | 0.485 | 478.52 |
| PP-YOLOE+_x | FP16 | 640 | 0. | 0. | 0. | | | PP-YOLOE+_x | FP16 | 640 | 0.528 | 0.705 | 0.579 | 121.17 |
| PP-YOLOE+_l | FP16 | 640 | 0. | 0. | 0. | | | PP-YOLOE+_l | FP16 | 640 | 0.511 | 0.686 | 0.557 | 191.82 |
| PP-YOLOE+_m | FP16 | 640 | 0. | 0. | 0. | | | PP-YOLOE+_m | FP16 | 640 | 0.483 | 0.658 | 0.528 | 264.39 |
| PP-YOLOE+_s | FP16 | 640 | 0.424 | 0.594 | 0.464 | 476.13 | | PP-YOLOE+_s | FP16 | 640 | 0.424 | 0.594 | 0.464 | 476.13 |
| PP-YOLOE-s (400) | FP16 | 640 | 0.423 | 0.589 | 0.463 | 461.23 | | PP-YOLOE-s (400) | FP16 | 640 | 0.423 | 0.589 | 0.463 | 461.23 |
| YOLOX-x | FP16 | 640 | 0.447 | 0.616 | 0.483 | 125.40 | | YOLOX-x | FP16 | 640 | 0.447 | 0.616 | 0.483 | 125.40 |
| YOLOX-l | FP16 | 640 | 0.430 | 0.598 | 0.466 | 193.10 | | YOLOX-l | FP16 | 640 | 0.430 | 0.598 | 0.466 | 193.10 |
| YOLOX-m | FP16 | 640 | 0.397 | 0.566 | 0.431 | 298.61 | | YOLOX-m | FP16 | 640 | 0.397 | 0.566 | 0.431 | 298.61 |
| YOLOX-s | FP16 | 640 | 0.335 | 0.502 | 0.365 | 522.05 | | YOLOX-s | FP16 | 640 | 0.335 | 0.502 | 0.365 | 522.05 |
| YOLOX-s legacy | FP16 | 640 | 0.375 | 0.569 | 0.407 | 518.52 | | YOLOX-s legacy | FP16 | 640 | 0.375 | 0.569 | 0.407 | 518.52 |
| YOLOX-Darknet | FP16 | 640 | 0.414 | 0.595 | 0.453 | 212.88 | | YOLOX-Darknet | FP16 | 640 | 0.414 | 0.595 | 0.453 | 212.88 |
| YOLOX-Tiny | FP16 | 640 | 0.274 | 0.427 | 0.292 | 633.95 | | YOLOX-Tiny | FP16 | 640 | 0.274 | 0.427 | 0.292 | 633.95 |
| YOLOX-Nano | FP16 | 640 | 0.212 | 0.342 | 0.222 | 633.04 | | YOLOX-Nano | FP16 | 640 | 0.212 | 0.342 | 0.222 | 633.04 |
| YOLOv8x | FP16 | 640 | 0.499 | 0.669 | 0.545 | 130.49 | | YOLOv8x | FP16 | 640 | 0.499 | 0.669 | 0.545 | 130.49 |
| YOLOv8l | FP16 | 640 | 0.491 | 0.660 | 0.535 | 180.75 | | YOLOv8l | FP16 | 640 | 0.491 | 0.660 | 0.535 | 180.75 |
| YOLOv8m | FP16 | 640 | 0.468 | 0.637 | 0.510 | 278.08 | | YOLOv8m | FP16 | 640 | 0.468 | 0.637 | 0.510 | 278.08 |
| YOLOv8s | FP16 | 640 | 0.415 | 0.578 | 0.453 | 493.45 | | YOLOv8s | FP16 | 640 | 0.415 | 0.578 | 0.453 | 493.45 |
| YOLOv8n | FP16 | 640 | 0.343 | 0.492 | 0.373 | 627.43 | | YOLOv8n | FP16 | 640 | 0.343 | 0.492 | 0.373 | 627.43 |
| YOLOv7 | FP16 | 640 | 0. | 0. | 0. | | | YOLOv7-u6 | FP16 | 640 | 0.484 | 0.652 | 0.530 | 193.54 |
| YOLOv6s 3.0 | FP16 | 640 | 0. | 0. | 0. | | | YOLOv7x* | FP16 | 640 | 0.496 | 0.679 | 0.536 | 155.07 |
| YOLOv5s 7.0 | FP16 | 640 | 0. | 0. | 0. | | | YOLOv7* | FP16 | 640 | 0.476 | 0.660 | 0.518 | 226.01 |
| YOLOv4 | FP16 | 640 | 0. | 0. | 0. | | | YOLOv7-Tiny Leaky* | FP16 | 640 | 0.345 | 0.516 | 0.372 | 626.23 |
| YOLOv3 | FP16 | 640 | 0. | 0. | 0. | | | YOLOv7-Tiny Leaky* | FP16 | 416 | 0.328 | 0.493 | 0.349 | 633.90 |
| YOLOv6-L 4.0 | FP16 | 640 | 0.490 | 0.671 | 0.535 | 178.41 |
| YOLOv6-M 4.0 | FP16 | 640 | 0.460 | 0.635 | 0.502 | 293.39 |
| YOLOv6-S 4.0 | FP16 | 640 | 0.416 | 0.585 | 0.453 | 513.90 |
| YOLOv6-N 4.0 | FP16 | 640 | 0.349 | 0.503 | 0.378 | 633.37 |
| YOLOv5x 7.0 | FP16 | 640 | 0.471 | 0.652 | 0.513 | 149.93 |
| YOLOv5l 7.0 | FP16 | 640 | 0.455 | 0.637 | 0.497 | 235.55 |
| YOLOv5m 7.0 | FP16 | 640 | 0.421 | 0.604 | 0.459 | 351.69 |
| YOLOv5s 7.0 | FP16 | 640 | 0.344 | 0.529 | 0.372 | 618.13 |
| YOLOv5n 7.0 | FP16 | 640 | 0.247 | 0.414 | 0.257 | 629.66 |
## ##

10
docs/YOLOv6.md
View File

@@ -1,5 +1,7 @@
# YOLOv6 usage # YOLOv6 usage
**NOTE**: You need to change the branch of the YOLOv6 repo according to the version of the model you want to convert.
**NOTE**: The yaml file is not required. **NOTE**: The yaml file is not required.
* [Convert model](#convert-model) * [Convert model](#convert-model)
@@ -29,17 +31,17 @@ Copy the `export_yoloV6.py` file from `DeepStream-Yolo/utils` directory to the `
#### 3. Download the model #### 3. Download the model
Download the `pt` file from [YOLOv6](https://github.com/meituan/YOLOv6/releases/) releases (example for YOLOv6-S 3.0) Download the `pt` file from [YOLOv6](https://github.com/meituan/YOLOv6/releases/) releases (example for YOLOv6-S 4.0)
``` ```
wget https://github.com/meituan/YOLOv6/releases/download/0.3.0/yolov6s.pt wget https://github.com/meituan/YOLOv6/releases/download/0.4.0/yolov6s.pt
``` ```
**NOTE**: You can use your custom model. **NOTE**: You can use your custom model.
#### 4. Convert model #### 4. Convert model
Generate the ONNX model file (example for YOLOv6-S 3.0) Generate the ONNX model file (example for YOLOv6-S 4.0)
``` ```
python3 export_yoloV6.py -w yolov6s.pt --simplify python3 export_yoloV6.py -w yolov6s.pt --simplify
@@ -122,7 +124,7 @@ Open the `DeepStream-Yolo` folder and compile the lib
### Edit the config_infer_primary_yoloV6 file ### Edit the config_infer_primary_yoloV6 file
Edit the `config_infer_primary_yoloV6.txt` file according to your model (example for YOLOv6-S 3.0 with 80 classes) Edit the `config_infer_primary_yoloV6.txt` file according to your model (example for YOLOv6-S 4.0 with 80 classes)
``` ```
[property] [property]

34
nvdsinfer_custom_impl_Yolo/nvdsparsebbox_Yolo.cpp
View File

@@ -73,22 +73,22 @@ addBBoxProposal(const float bx1, const float by1, const float bx2, const float b
} }
static std::vector<NvDsInferParseObjectInfo> static std::vector<NvDsInferParseObjectInfo>
decodeTensorYolo(const float* detection, const uint& outputSize, const uint& count, const uint& netW, const uint& netH, decodeTensorYolo(const float* detection, const uint& outputSize, const uint& netW, const uint& netH,
const std::vector<float>& preclusterThreshold) const std::vector<float>& preclusterThreshold)
{ {
std::vector<NvDsInferParseObjectInfo> binfo; std::vector<NvDsInferParseObjectInfo> binfo;
for (uint b = 0; b < outputSize; ++b) { for (uint b = 0; b < outputSize; ++b) {
float maxProb = count == 6 ? detection[b * count + 4] : detection[b * count + 4] * detection[b * count + 6]; float maxProb = detection[b * 6 + 4];
int maxIndex = (int) detection[b * count + 5]; int maxIndex = (int) detection[b * 6 + 5];
if (maxProb < preclusterThreshold[maxIndex]) if (maxProb < preclusterThreshold[maxIndex])
continue; continue;
float bxc = detection[b * count + 0]; float bxc = detection[b * 6 + 0];
float byc = detection[b * count + 1]; float byc = detection[b * 6 + 1];
float bw = detection[b * count + 2]; float bw = detection[b * 6 + 2];
float bh = detection[b * count + 3]; float bh = detection[b * 6 + 3];
float bx1 = bxc - bw / 2; float bx1 = bxc - bw / 2;
float by1 = byc - bh / 2; float by1 = byc - bh / 2;
@@ -102,22 +102,22 @@ decodeTensorYolo(const float* detection, const uint& outputSize, const uint& cou
} }
static std::vector<NvDsInferParseObjectInfo> static std::vector<NvDsInferParseObjectInfo>
decodeTensorYoloE(const float* detection, const uint& outputSize, const uint& count, const uint& netW, const uint& netH, decodeTensorYoloE(const float* detection, const uint& outputSize, const uint& netW, const uint& netH,
const std::vector<float>& preclusterThreshold) const std::vector<float>& preclusterThreshold)
{ {
std::vector<NvDsInferParseObjectInfo> binfo; std::vector<NvDsInferParseObjectInfo> binfo;
for (uint b = 0; b < outputSize; ++b) { for (uint b = 0; b < outputSize; ++b) {
float maxProb = count == 6 ? detection[b * count + 4] : detection[b * count + 4] * detection[b * count + 6]; float maxProb = detection[b * 6 + 4];
int maxIndex = (int) detection[b * count + 5]; int maxIndex = (int) detection[b * 6 + 5];
if (maxProb < preclusterThreshold[maxIndex]) if (maxProb < preclusterThreshold[maxIndex])
continue; continue;
float bx1 = detection[b * count + 0]; float bx1 = detection[b * 6 + 0];
float by1 = detection[b * count + 1]; float by1 = detection[b * 6 + 1];
float bx2 = detection[b * count + 2]; float bx2 = detection[b * 6 + 2];
float by2 = detection[b * count + 3]; float by2 = detection[b * 6 + 3];
addBBoxProposal(bx1, by1, bx2, by2, netW, netH, maxIndex, maxProb, binfo); addBBoxProposal(bx1, by1, bx2, by2, netW, netH, maxIndex, maxProb, binfo);
} }
@@ -139,9 +139,8 @@ NvDsInferParseCustomYolo(std::vector<NvDsInferLayerInfo> const& outputLayersInfo
const NvDsInferLayerInfo& layer = outputLayersInfo[0]; const NvDsInferLayerInfo& layer = outputLayersInfo[0];
const uint outputSize = layer.inferDims.d[0]; const uint outputSize = layer.inferDims.d[0];
const uint count = layer.inferDims.d[1];
std::vector<NvDsInferParseObjectInfo> outObjs = decodeTensorYolo((const float*) (layer.buffer), outputSize, count, std::vector<NvDsInferParseObjectInfo> outObjs = decodeTensorYolo((const float*) (layer.buffer), outputSize,
networkInfo.width, networkInfo.height, detectionParams.perClassPreclusterThreshold); networkInfo.width, networkInfo.height, detectionParams.perClassPreclusterThreshold);
objects.insert(objects.end(), outObjs.begin(), outObjs.end()); objects.insert(objects.end(), outObjs.begin(), outObjs.end());
@@ -165,9 +164,8 @@ NvDsInferParseCustomYoloE(std::vector<NvDsInferLayerInfo> const& outputLayersInf
const NvDsInferLayerInfo& layer = outputLayersInfo[0]; const NvDsInferLayerInfo& layer = outputLayersInfo[0];
const uint outputSize = layer.inferDims.d[0]; const uint outputSize = layer.inferDims.d[0];
const uint count = layer.inferDims.d[1];
std::vector<NvDsInferParseObjectInfo> outObjs = decodeTensorYoloE((const float*) (layer.buffer), outputSize, count, std::vector<NvDsInferParseObjectInfo> outObjs = decodeTensorYoloE((const float*) (layer.buffer), outputSize,
networkInfo.width, networkInfo.height, detectionParams.perClassPreclusterThreshold); networkInfo.width, networkInfo.height, detectionParams.perClassPreclusterThreshold);
objects.insert(objects.end(), outObjs.begin(), outObjs.end()); objects.insert(objects.end(), outObjs.begin(), outObjs.end());

15
nvdsinfer_custom_impl_Yolo/yoloForward.cu
View File

@@ -50,13 +50,12 @@ __global__ void gpuYoloLayer(const float* input, float* output, int* count, cons
int _count = (int)atomicAdd(count, 1); int _count = (int)atomicAdd(count, 1);
output[_count * 7 + 0] = xc; output[_count * 6 + 0] = xc;
output[_count * 7 + 1] = yc; output[_count * 6 + 1] = yc;
output[_count * 7 + 2] = w; output[_count * 6 + 2] = w;
output[_count * 7 + 3] = h; output[_count * 6 + 3] = h;
output[_count * 7 + 4] = maxProb; output[_count * 6 + 4] = maxProb * objectness;
output[_count * 7 + 5] = maxIndex; output[_count * 6 + 5] = maxIndex;
output[_count * 7 + 6] = objectness;
} }
cudaError_t cudaYoloLayer(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize, cudaError_t cudaYoloLayer(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize,
@@ -76,7 +75,7 @@ cudaError_t cudaYoloLayer(const void* input, void* output, void* count, const ui
for (unsigned int batch = 0; batch < batchSize; ++batch) { for (unsigned int batch = 0; batch < batchSize; ++batch) {
gpuYoloLayer<<<number_of_blocks, threads_per_block, 0, stream>>>( gpuYoloLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*> (input) + (batch * inputSize), reinterpret_cast<const float*> (input) + (batch * inputSize),
reinterpret_cast<float*> (output) + (batch * 7 * outputSize), reinterpret_cast<float*> (output) + (batch * 6 * outputSize),
reinterpret_cast<int*> (count) + (batch), reinterpret_cast<int*> (count) + (batch),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY, netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY,
reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask)); reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask));

15
nvdsinfer_custom_impl_Yolo/yoloForward_nc.cu
View File

@@ -47,13 +47,12 @@ __global__ void gpuYoloLayer_nc(const float* input, float* output, int* count, c
int _count = (int)atomicAdd(count, 1); int _count = (int)atomicAdd(count, 1);
output[_count * 7 + 0] = xc; output[_count * 6 + 0] = xc;
output[_count * 7 + 1] = yc; output[_count * 6 + 1] = yc;
output[_count * 7 + 2] = w; output[_count * 6 + 2] = w;
output[_count * 7 + 3] = h; output[_count * 6 + 3] = h;
output[_count * 7 + 4] = maxProb; output[_count * 6 + 4] = maxProb * objectness;
output[_count * 7 + 5] = maxIndex; output[_count * 6 + 5] = maxIndex;
output[_count * 7 + 6] = objectness;
} }
cudaError_t cudaYoloLayer_nc(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize, cudaError_t cudaYoloLayer_nc(const void* input, void* output, void* count, const uint& batchSize, uint64_t& inputSize,
@@ -73,7 +72,7 @@ cudaError_t cudaYoloLayer_nc(const void* input, void* output, void* count, const
for (unsigned int batch = 0; batch < batchSize; ++batch) { for (unsigned int batch = 0; batch < batchSize; ++batch) {
gpuYoloLayer_nc<<<number_of_blocks, threads_per_block, 0, stream>>>( gpuYoloLayer_nc<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*> (input) + (batch * inputSize), reinterpret_cast<const float*> (input) + (batch * inputSize),
reinterpret_cast<float*> (output) + (batch * 7 * outputSize), reinterpret_cast<float*> (output) + (batch * 6 * outputSize),
reinterpret_cast<int*> (count) + (batch), reinterpret_cast<int*> (count) + (batch),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY, netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY,
reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask)); reinterpret_cast<const float*> (anchors), reinterpret_cast<const int*> (mask));

15
nvdsinfer_custom_impl_Yolo/yoloForward_v2.cu
View File

@@ -68,13 +68,12 @@ __global__ void gpuRegionLayer(const float* input, float* softmax, float* output
int _count = (int)atomicAdd(count, 1); int _count = (int)atomicAdd(count, 1);
output[_count * 7 + 0] = xc; output[_count * 6 + 0] = xc;
output[_count * 7 + 1] = yc; output[_count * 6 + 1] = yc;
output[_count * 7 + 2] = w; output[_count * 6 + 2] = w;
output[_count * 7 + 3] = h; output[_count * 6 + 3] = h;
output[_count * 7 + 4] = maxProb; output[_count * 6 + 4] = maxProb * objectness;
output[_count * 7 + 5] = maxIndex; output[_count * 6 + 5] = maxIndex;
output[_count * 7 + 6] = objectness;
} }
cudaError_t cudaRegionLayer(const void* input, void* softmax, void* output, void* count, const uint& batchSize, cudaError_t cudaRegionLayer(const void* input, void* softmax, void* output, void* count, const uint& batchSize,
@@ -93,7 +92,7 @@ cudaError_t cudaRegionLayer(const void* input, void* softmax, void* output, void
gpuRegionLayer<<<number_of_blocks, threads_per_block, 0, stream>>>( gpuRegionLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*> (input) + (batch * inputSize), reinterpret_cast<const float*> (input) + (batch * inputSize),
reinterpret_cast<float*> (softmax) + (batch * inputSize), reinterpret_cast<float*> (softmax) + (batch * inputSize),
reinterpret_cast<float*> (output) + (batch * 7 * outputSize), reinterpret_cast<float*> (output) + (batch * 6 * outputSize),
reinterpret_cast<int*> (count) + (batch), reinterpret_cast<int*> (count) + (batch),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes,
reinterpret_cast<const float*> (anchors)); reinterpret_cast<const float*> (anchors));

4
nvdsinfer_custom_impl_Yolo/yoloPlugins.cpp
View File

@@ -103,7 +103,7 @@ nvinfer1::Dims
YoloLayer::getOutputDimensions(int index, const nvinfer1::Dims* inputs, int nbInputDims) noexcept YoloLayer::getOutputDimensions(int index, const nvinfer1::Dims* inputs, int nbInputDims) noexcept
{ {
assert(index == 0); assert(index == 0);
return nvinfer1::Dims{2, {static_cast<int>(m_OutputSize), 7}}; return nvinfer1::Dims{2, {static_cast<int>(m_OutputSize), 6}};
} }
bool bool
@@ -125,7 +125,7 @@ YoloLayer::enqueue(int batchSize, void const* const* inputs, void* const* output
noexcept noexcept
{ {
void* output = outputs[0]; void* output = outputs[0];
CUDA_CHECK(cudaMemsetAsync((float*) output, 0, sizeof(float) * m_OutputSize * 7 * batchSize, stream)); CUDA_CHECK(cudaMemsetAsync((float*) output, 0, sizeof(float) * m_OutputSize * 6 * batchSize, stream));
void* count = workspace; void* count = workspace;
CUDA_CHECK(cudaMemsetAsync((int*) count, 0, sizeof(int) * batchSize, stream)); CUDA_CHECK(cudaMemsetAsync((int*) count, 0, sizeof(int) * batchSize, stream));

2
utils/export_yoloV5.py
View File

@@ -19,7 +19,7 @@ class DeepStreamOutput(nn.Module):
boxes = x[:, :, :4] boxes = x[:, :, :4]
objectness = x[:, :, 4:5] objectness = x[:, :, 4:5]
scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True) scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True)
return torch.cat((boxes, scores, classes, objectness), dim=2) return torch.cat((boxes, scores * objectness, classes), dim=2)
def suppress_warnings(): def suppress_warnings():

10
utils/export_yoloV6.py
View File

@@ -6,20 +6,24 @@ import onnx
import torch import torch
import torch.nn as nn import torch.nn as nn
from yolov6.utils.checkpoint import load_checkpoint from yolov6.utils.checkpoint import load_checkpoint
from yolov6.layers.common import RepVGGBlock, ConvModule, SiLU from yolov6.layers.common import RepVGGBlock, SiLU
from yolov6.models.effidehead import Detect from yolov6.models.effidehead import Detect
try:
from yolov6.layers.common import ConvModule
except ImportError:
from yolov6.layers.common import Conv as ConvModule
class DeepStreamOutput(nn.Module): class DeepStreamOutput(nn.Module):
def __init__(self): def __init__(self):
super().__init__() super().__init__()
def forward(self, x): def forward(self, x):
print(x)
boxes = x[:, :, :4] boxes = x[:, :, :4]
objectness = x[:, :, 4:5] objectness = x[:, :, 4:5]
scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True) scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True)
return torch.cat((boxes, scores, classes, objectness), dim=2) return torch.cat((boxes, scores * objectness, classes), dim=2)
def suppress_warnings(): def suppress_warnings():

2
utils/export_yoloV7.py
View File

@@ -19,7 +19,7 @@ class DeepStreamOutput(nn.Module):
boxes = x[:, :, :4] boxes = x[:, :, :4]
objectness = x[:, :, 4:5] objectness = x[:, :, 4:5]
scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True) scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True)
return torch.cat((boxes, scores, classes, objectness), dim=2) return torch.cat((boxes, scores * objectness, classes), dim=2)
def suppress_warnings(): def suppress_warnings():

77
utils/export_yoloV7_u6.py Normal file
View File

@@ -0,0 +1,77 @@
import os
import sys
import argparse
import warnings
import onnx
import torch
import torch.nn as nn
from models.experimental import attempt_load
from models.yolo import Detect, V6Detect, IV6Detect
from utils.torch_utils import select_device
class DeepStreamOutput(nn.Module):
def __init__(self):
super().__init__()
def forward(self, x):
x = x.transpose(1, 2)
boxes = x[:, :, :4]
scores, classes = torch.max(x[:, :, 4:], 2, keepdim=True)
return torch.cat((boxes, scores, classes), dim=2)
def suppress_warnings():
warnings.filterwarnings('ignore', category=torch.jit.TracerWarning)
warnings.filterwarnings('ignore', category=UserWarning)
warnings.filterwarnings('ignore', category=DeprecationWarning)
def yolov7_u6_export(weights, device):
model = attempt_load(weights, device=device, inplace=True, fuse=True)
model.eval()
for k, m in model.named_modules():
if isinstance(m, (Detect, V6Detect, IV6Detect)):
m.inplace = False
m.dynamic = False
m.export = True
return model
def main(args):
suppress_warnings()
device = select_device('cpu')
model = yolov7_u6_export(args.weights, device)
model = nn.Sequential(model, DeepStreamOutput())
img_size = args.size * 2 if len(args.size) == 1 else args.size
onnx_input_im = torch.zeros(1, 3, *img_size).to(device)
onnx_output_file = os.path.basename(args.weights).split('.pt')[0] + '.onnx'
torch.onnx.export(model, onnx_input_im, onnx_output_file, verbose=False, opset_version=args.opset,
do_constant_folding=True, input_names=['input'], output_names=['output'], dynamic_axes=None)
if args.simplify:
import onnxsim
model_onnx = onnx.load(onnx_output_file)
model_onnx, _ = onnxsim.simplify(model_onnx)
onnx.save(model_onnx, onnx_output_file)
def parse_args():
parser = argparse.ArgumentParser(description='DeepStream YOLOv7-u6 conversion')
parser.add_argument('-w', '--weights', required=True, help='Input weights (.pt) file path (required)')
parser.add_argument('-s', '--size', nargs='+', type=int, default=[640], help='Inference size [H,W] (default [640])')
parser.add_argument('--opset', type=int, default=12, help='ONNX opset version')
parser.add_argument('--simplify', action='store_true', help='ONNX simplify model')
args = parser.parse_args()
if not os.path.isfile(args.weights):
raise SystemExit('Invalid weights file')
return args
if __name__ == '__main__':
args = parse_args()
sys.exit(main(args))

2
utils/export_yolor.py
View File

@@ -16,7 +16,7 @@ class DeepStreamOutput(nn.Module):
boxes = x[:, :, :4] boxes = x[:, :, :4]
objectness = x[:, :, 4:5] objectness = x[:, :, 4:5]
scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True) scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True)
return torch.cat((boxes, scores, classes, objectness), dim=2) return torch.cat((boxes, scores * objectness, classes), dim=2)
def suppress_warnings(): def suppress_warnings():

2
utils/export_yolox.py
View File

@@ -18,7 +18,7 @@ class DeepStreamOutput(nn.Module):
boxes = x[:, :, :4] boxes = x[:, :, :4]
objectness = x[:, :, 4:5] objectness = x[:, :, 4:5]
scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True) scores, classes = torch.max(x[:, :, 5:], 2, keepdim=True)
return torch.cat((boxes, scores, classes, objectness), dim=2) return torch.cat((boxes, scores * objectness, classes), dim=2)
def suppress_warnings(): def suppress_warnings():