mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Merge remote-tracking branch 'upstream/master'

Browse Source
This commit is contained in:
mamontesg
2020-12-04 10:25:59 +01:00
parent 67402848b1 b8855b9599
commit b4d1bfe255
37 changed files with 2916 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+58 -33
View File
@@ -3,42 +3,64 @@ tkDNN is a Deep Neural Network library built with cuDNN and tensorRT primitives,
The main goal of this project is to exploit NVIDIA boards as much as possible to obtain the best inference performance. It does not allow training.
If you use tkDNN in your research, please cite one of the following papers. For use in commercial solutions, write at gattifrancesco@hotmail.it and micaela.verucchi@unimore.it or refer to https://hipert.unimore.it/ .
If you use tkDNN in your research, please cite the [following paper](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9212130&casa_token=sQTJXi7tJNoAAAAA:BguH9xCIY48MxbtDS3LXzIXzO-9sWArm7Hd7y7BwaLmqRuM_Gx8bOYizFPNMNtpo5K0kB-P-). For use in commercial solutions, write at gattifrancesco@hotmail.it and micaela.verucchi@unimore.it or refer to https://hipert.unimore.it/ .
```
Accepted paper @ IRC 2020, will soon be published.
M. Verucchi, L. Bartoli, F. Bagni, F. Gatti, P. Burgio and M. Bertogna, "Real-Time clustering and LiDAR-camera fusion on embedded platforms for self-driving cars", in proceedings in IEEE Robotic Computing (2020)
Accepted paper @ ETFA 2020, will soon be published.
M. Verucchi, G. Brilli, D. Sapienza, M. Verasani, M. Arena, F. Gatti, A. Capotondi, R. Cavicchioli, M. Bertogna, M. Solieri
"A Systematic Assessment of Embedded Neural Networks for Object Detection", in IEEE International Conference on Emerging Technologies and Factory Automation (2020)
@inproceedings{verucchi2020systematic,
title={A Systematic Assessment of Embedded Neural Networks for Object Detection},
author={Verucchi, Micaela and Brilli, Gianluca and Sapienza, Davide and Verasani, Mattia and Arena, Marco and Gatti, Francesco and Capotondi, Alessandro and Cavicchioli, Roberto and Bertogna, Marko and Solieri, Marco},
booktitle={2020 25th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA)},
volume={1},
pages={937--944},
year={2020},
organization={IEEE}
}
```
## Results
Inference FPS of yolov4 with tkDNN, average of 1200 images with the same dimesion as the input size, on
## FPS Results
Inference FPS of yolov4 with tkDNN, average of 1200 images with the same dimension as the input size, on
* RTX 2080Ti (CUDA 10.2, TensorRT 7.0.0, Cudnn 7.6.5);
* Xavier AGX, Jetpack 4.3 (CUDA 10.0, CUDNN 7.6.3, tensorrt 6.0.1 );
* Xavier NX, Jetpack 4.4 (CUDA 10.2, CUDNN 8.0.0, tensorrt 7.1.0 ).
* Tx2, Jetpack 4.2 (CUDA 10.0, CUDNN 7.3.1, tensorrt 5.0.6 );
* Jetson Nano, Jetpack 4.4 (CUDA 10.2, CUDNN 8.0.0, tensorrt 7.1.0 ).
| Platform | Network | FP32, B=1 | FP32, B=4 | FP16, B=1 | FP16, B=4 | INT8, B=1 | INT8, B=4 |
| :------: | :-----: | :-----: | :-----: | :-----: | :-----: | :-----: | :-----: |
| RTX 2080Ti | yolo4 320 | 118,59 |237,31 | 207,81 | 443,32 | 262,37 | 530,93 |
| RTX 2080Ti | yolo4 416 | 104,81 |162,86 | 169,06 | 293,78 | 206,93 | 353,26 |
| RTX 2080Ti | yolo4 512 | 92,98 |132,43 | 140,36 | 215,17 | 165,35 | 254,96 |
| RTX 2080Ti | yolo4 608 | 63,77 |81,53 | 111,39 | 152,89 | 127,79 | 184,72 |
| AGX Xavier | yolo4 320 | 26,78 |32,05 | 57,14 | 79,05 | 73,15 | 97,56 |
| AGX Xavier | yolo4 416 | 19,96 |21,52 | 41,01 | 49,00 | 50,81 | 60,61 |
| AGX Xavier | yolo4 512 | 16,58 |16,98 | 31,12 | 33,84 | 37,82 | 41,28 |
| AGX Xavier | yolo4 608 | 9,45 |10,13 | 21,92 | 23,36 | 27,05 | 28,93 |
| Tx2 | yolo4 320 | 11,18 | 12,07 | 15,32 | 16,31 | - | - |
| Tx2 | yolo4 416 | 7,30 | 7,58 | 9,45 | 9,90 | - | - |
| Tx2 | yolo4 512 | 5,96 | 5,95 | 7,22 | 7,23 | - | - |
| Tx2 | yolo4 608 | 3,63 | 3,65 | 4,67 | 4,70 | - | - |
| Nano | yolo4 320 | 4,23 | 4,55 | 6,14 | 6,53 | - | - |
| Nano | yolo4 416 | 2,88 | 3,00 | 3,90 | 4,04 | - | - |
| Nano | yolo4 512 | 2,32 | 2,34 | 3,02 | 3,04 | - | - |
| Nano | yolo4 608 | 1,40 | 1,41 | 1,92 | 1,93 | - | - |
| RTX 2080Ti | yolo4 320 | 118.59 | 237.31 | 207.81 | 443.32 | 262.37 | 530.93 |
| RTX 2080Ti | yolo4 416 | 104.81 | 162.86 | 169.06 | 293.78 | 206.93 | 353.26 |
| RTX 2080Ti | yolo4 512 | 92.98 | 132.43 | 140.36 | 215.17 | 165.35 | 254.96 |
| RTX 2080Ti | yolo4 608 | 63.77 | 81.53 | 111.39 | 152.89 | 127.79 | 184.72 |
| AGX Xavier | yolo4 320 | 26.78 | 32.05 | 57.14 | 79.05 | 73.15 | 97.56 |
| AGX Xavier | yolo4 416 | 19.96 | 21.52 | 41.01 | 49.00 | 50.81 | 60.61 |
| AGX Xavier | yolo4 512 | 16.58 | 16.98 | 31.12 | 33.84 | 37.82 | 41.28 |
| AGX Xavier | yolo4 608 | 9.45 | 10.13 | 21.92 | 23.36 | 27.05 | 28.93 |
| Xavier NX | yolo4 320 | 14.56 | 16.25 | 30.14 | 41.15 | 42.13 | 53.42 |
| Xavier NX | yolo4 416 | 10.02 | 10.60 | 22.43 | 25.59 | 29.08 | 32.94 |
| Xavier NX | yolo4 512 | 8.10 | 8.32 | 15.78 | 17.13 | 20.51 | 22.46 |
| Xavier NX | yolo4 608 | 5.26 | 5.18 | 11.54 | 12.06 | 15.09 | 15.82 |
| Tx2 | yolo4 320 | 11.18 | 12.07 | 15.32 | 16.31 | - | - |
| Tx2 | yolo4 416 | 7.30 | 7.58 | 9.45 | 9.90 | - | - |
| Tx2 | yolo4 512 | 5.96 | 5.95 | 7.22 | 7.23 | - | - |
| Tx2 | yolo4 608 | 3.63 | 3.65 | 4.67 | 4.70 | - | - |
| Nano | yolo4 320 | 4.23 | 4.55 | 6.14 | 6.53 | - | - |
| Nano | yolo4 416 | 2.88 | 3.00 | 3.90 | 4.04 | - | - |
| Nano | yolo4 512 | 2.32 | 2.34 | 3.02 | 3.04 | - | - |
| Nano | yolo4 608 | 1.40 | 1.41 | 1.92 | 1.93 | - | - |
## MAP Results
Results for COCO val 2017 (5k images), on RTX 2080Ti, with conf threshold=0.001
| | CodaLab | CodaLab | CodaLab | CodaLab | tkDNN map | tkDNN map |
| -------------------- | :-----------: | :-------: | :-----------: | :---------: | :-----------: | :-------: |
| | **tkDNN** | **tkDNN** | **darknet** | **darknet** | **tkDNN** | **tkDNN** |
| | MAP(0.5:0.95) | AP50 | MAP(0.5:0.95) | AP50 | MAP(0.5:0.95) | AP50 |
| Yolov3 (416x416) | 0.381 | 0.675 | 0.380 | 0.675 | 0.372 | 0.663 |
| yolov4 (416x416) | 0.468 | 0.705 | 0.471 | 0.710 | 0.459 | 0.695 |
| yolov3tiny (416x416) | 0.096 | 0.202 | 0.096 | 0.201 | 0.093 | 0.198 |
| yolov4tiny (416x416) | 0.202 | 0.400 | 0.201 | 0.400 | 0.197 | 0.395 |
| Cnet-dla34 (512x512) | 0.366 | 0.543 | \- | \- | 0.361 | 0.535 |
| mv2SSD (512x512) | 0.226 | 0.381 | \- | \- | 0.223 | 0.378 |
## Index
- [tkDNN](#tkdnn)
@@ -155,7 +177,7 @@ tkDNN implement and easy parser for darknet cfg files, a network can be converte
tk::dnn::Network *net = tk::dnn::darknetParser("yolov4.cfg", "yolov4/layers", "coco.names");
net->print();
```
All models from darknet are now parsed directly from cfg, you still need to export the weights with the descripted tools in the previus section.
All models from darknet are now parsed directly from cfg, you still need to export the weights with the described tools in the previous section.
<details>
<summary>Supported layers</summary>
convolutional
@@ -189,11 +211,11 @@ cmake .. -DDEBUG=True
make
```
Once you have succesfully created your rt file, run the demo:
Once you have successfully created your rt file, run the demo:
```
./demo yolo4_fp32.rt ../demo/yolo_test.mp4 y
```
In general the demo program takes 6 parameters:
In general the demo program takes 7 parameters:
```
./demo <network-rt-file> <path-to-video> <kind-of-network> <number-of-classes> <n-batches> <show-flag>
```
@@ -204,6 +226,7 @@ where
* ```<number-of-classes>```is the number of classes the network is trained on
* ```<n-batches>``` number of batches to use in inference (N.B. you should first export TKDNN_BATCHSIZE to the required n_batches and create again the rt file for the network).
* ```<show-flag>``` if set to 0 the demo will not show the visualization but save the video into result.mp4 (if n-batches ==1)
* ```<conf-thresh>``` confidence threshold for the detector. Only bounding boxes with threshold greater than conf-thresh will be displayed.
N.b. By default it is used FP32 inference
@@ -232,7 +255,7 @@ You should provide image_list.txt and label_list.txt, using training images. How
```
bash scripts/download_validation.sh COCO
```
to automatically download COCO2017 validation (inside demo folder) and create those needed file. Use BDD insted of COCO to download BDD validation.
to automatically download COCO2017 validation (inside demo folder) and create those needed file. Use BDD instead of COCO to download BDD validation.
Then a complete example using yolo3 and COCO dataset would be:
```
@@ -254,8 +277,8 @@ N.B.
export TKDNN_BATCHSIZE=2
# build tensorRT files
```
This will create a TensorRT file with the desidered **max** batch size.
The test will still run with a batch of 1, but the created tensorRT can manage the desidered batch size.
This will create a TensorRT file with the desired **max** batch size.
The test will still run with a batch of 1, but the created tensorRT can manage the desired batch size.
### Test batch Inference
This will test the network with random input and check if the output of each batch is the same.
@@ -301,7 +324,7 @@ cd build
./map_demo dla34_cnet_FP32.rt c ../demo/COCO_val2017/all_labels.txt ../demo/config.yaml
```
This demo also creates a json file named ```net_name_COCO_res.json``` containing all the detections computed. The detections are in COCO format, the correct format to subit the results to [CodaLab COCO detection challenge](https://competitions.codalab.org/competitions/20794#participate).
This demo also creates a json file named ```net_name_COCO_res.json``` containing all the detections computed. The detections are in COCO format, the correct format to submit the results to [CodaLab COCO detection challenge](https://competitions.codalab.org/competitions/20794#participate).
## Existing tests and supported networks
@@ -329,7 +352,8 @@ This demo also creates a json file named ```net_name_COCO_res.json``` containing
| csresnext50-panet-spp | Cross Stage Partial Network <sup>7</sup> | [COCO 2014](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/Kcs4xBozwY4wFx8/download) |
| yolo4 | Yolov4 <sup>8</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) |
| yolo4_berkeley | Yolov4 <sup>8</sup> | [BDD100K ](https://bair.berkeley.edu/blog/2018/05/30/bdd/) | 10 | 540x320 | [weights](https://cloud.hipert.unimore.it/s/nkWFa5fgb4NTdnB/download) |
| yolo4tiny | Yolov4 tiny | [COCO 2017](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) |
| yolo4tiny | Yolov4 tiny <sup>9</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) |
| yolo4x | Yolov4x-mish <sup>9</sup> | [COCO 2017](http://cocodataset.org/) | 80 | 672x672 | [weights](https://cloud.hipert.unimore.it/s/BLPpiAigZJLorQD/download) |
## References
@@ -342,3 +366,4 @@ This demo also creates a json file named ```net_name_COCO_res.json``` containing
6. He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
7. Wang, Chien-Yao, et al. "CSPNet: A New Backbone that can Enhance Learning Capability of CNN." arXiv preprint arXiv:1911.11929 (2019).
8. Bochkovskiy, Alexey, Chien-Yao Wang, and Hong-Yuan Mark Liao. "YOLOv4: Optimal Speed and Accuracy of Object Detection." arXiv preprint arXiv:2004.10934 (2020).
9. Bochkovskiy, Alexey, "Yolo v4, v3 and v2 for Windows and Linux" (https://github.com/AlexeyAB/darknet)
demo/config.yaml
+1 -1
View File
@@ -3,5 +3,5 @@ map_points : 101 #number of recall points (0 for all, 101 for COCO, 11 Pascal
map_levels : 10 #number of IoU step for the AP
map_step : 0.05 #step of IoU
IoU_thresh : 0.5 #starting IoU threshold
conf_thresh : 0.0 #threshold on the condifence of the bbox
conf_thresh : 0.001 #threshold on the condifence of the bbox
verbose : false #print on screen information
demo/demo/demo.cpp
+4 -1
View File
@@ -40,6 +40,9 @@ int main(int argc, char *argv[]) {
bool show = true;
if(argc > 6)
show = atoi(argv[6]);
float conf_thresh=0.3;
if(argc > 7)
conf_thresh = atof(argv[7]);
if(n_batch < 1 || n_batch > 64)
FatalError("Batch dim not supported");
@@ -69,7 +72,7 @@ int main(int argc, char *argv[]) {
FatalError("Network type not allowed (3rd parameter)\n");
}
detNN->init(net, n_classes, n_batch);
detNN->init(net, n_classes, n_batch, conf_thresh);
gRun = true;
demo/demo/map.cpp
+1 -1
View File
@@ -105,7 +105,7 @@ int main(int argc, char *argv[])
default:
FatalError("Network type not allowed (3rd parameter)\n");
}
detNN->init(net, n_classes);
detNN->init(net, n_classes, 1, conf_thresh);
//read images
std::ifstream all_labels(labels_path);
include/tkDNN/CenternetDetection.h
+1 -1
View File
@@ -73,7 +73,7 @@ public:
CenternetDetection() {};
~CenternetDetection() {};
bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1);
bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3);
void preprocess(cv::Mat &frame, const int bi=0);
void postprocess(const int bi=0,const bool mAP=false);
};
include/tkDNN/DarknetParser.h
+3
View File
@@ -24,7 +24,10 @@ namespace tk { namespace dnn {
int num = 1;
int pad = 0;
int coords = 4;
int nms_kind = 0;
int new_coords= 0;
float scale_xy = 1;
float nms_thresh = 0.45;
std::vector<int> layers;
std::string activation = "linear";
include/tkDNN/DetectionNN.h
+5 -5
View File
@@ -76,15 +76,15 @@ class DetectionNN {
~DetectionNN(){};
/**
* Method used to inialize the class, allocate memory and compute
* Method used to initialize the class, allocate memory and compute
* needed data.
*
* @param tensor_path path to the rt file og the NN.
* @param tensor_path path to the rt file of the NN.
* @param n_classes number of classes for the given dataset.
* @param n_batches maximum number of batches to use in inference
* @return true if everything is correct, false otherwise.
*/
virtual bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1) = 0;
virtual bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3) = 0;
/**
* This method performs the whole detection of the NN.
@@ -141,9 +141,9 @@ class DetectionNN {
}
/**
* Method to draw boundixg boxes and labels on a frame.
* Method to draw bounding boxes and labels on a frame.
*
* @param frames orginal frame to draw bounding box on.
* @param frames original frame to draw bounding box on.
*/
void draw(std::vector<cv::Mat>& frames) {
tk::dnn::box b;
include/tkDNN/ImuOdom.h
+2 -2
View File
@@ -44,7 +44,7 @@ class ImuOdom {
virtual ~ImuOdom() {}
/**
* Method used for inizialize the class
* Method used for initialize the class
*
* @return Success of the initialization
*/
@@ -141,7 +141,7 @@ class ImuOdom {
//odomPOS = odomPOS + deltaP.cast<double>(); // V2
odomROT = odomROT * q.normalized().toRotationMatrix();
// compute euler
// compute Euler
auto newEULER = odomROT.eulerAngles(0, 1, 2);
for(int i=0; i<3; i++) {
while( fabs(newEULER(i) - odomEULER(i)) > M_PI_2 ) {
include/tkDNN/Layer.h
+19 -15
View File
@@ -171,7 +171,7 @@ public:
/**
Input layer (it doesnt need weigths)
Input layer (it doesn't need weights)
*/
class Input : public Layer {
@@ -207,7 +207,7 @@ public:
/**
Avaible activation functions
Available activation functions
*/
typedef enum {
ACTIVATION_ELU = 100,
@@ -216,7 +216,7 @@ typedef enum {
} tkdnnActivationMode_t;
/**
Activation layer (it doesnt need weigths)
Activation layer (it doesn't need weights)
*/
class Activation : public Layer {
@@ -318,9 +318,9 @@ public:
virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
const bool bidirectional = true; /**> is the net bidir */
bool returnSeq = false; /**> if false return only the result of last timestep */
bool returnSeq = false; /**> if false return only the result of last timestamp */
int stateSize = 0; /**> number of hidden states */
int seqLen = 0; /**> number of timesteps */
int seqLen = 0; /**> number of timestamp */
int numLayers = 1; /**> number of internal layers */
protected:
@@ -367,7 +367,7 @@ public:
/**
Deformable Convolutionl 2d layer
Deformable Convolutional 2d layer
*/
class DeformConv2d : public LayerWgs {
@@ -449,7 +449,7 @@ protected:
/**
Avaible pooling functions (padding on tkDNN is not supported)
Available pooling functions (padding on tkDNN is not supported)
*/
typedef enum {
POOLING_MAX = 0,
@@ -460,7 +460,7 @@ typedef enum {
/**
Pooling layer
currenty supported only 2d pooing (also on 3d input)
currently supported only 2d pooing (also on 3d input)
*/
class Pooling : public Layer {
@@ -526,7 +526,7 @@ public:
/**
Reorg layer
Mantain same dimension but change C*H*W distribution
Maintains same dimension but change C*H*W distribution
*/
class Reorg : public Layer {
@@ -559,7 +559,7 @@ public:
/**
Upsample layer
Mantain same dimension but change C*H*W distribution
Maintains same dimension but change C*H*W distribution
*/
class Upsample : public Layer {
@@ -610,24 +610,28 @@ public:
int sort_class;
};
Yolo(Network *net, int classes, int num, std::string fname_weights,int n_masks=3, float scale_xy=1);
enum nmsKind_t {GREEDY_NMS=0, DIOU_NMS=1};
Yolo(Network *net, int classes, int num, std::string fname_weights,int n_masks=3, float scale_xy=1, double nms_thresh=0.45, nmsKind_t nsm_kind=GREEDY_NMS, int new_coords=0);
virtual ~Yolo();
virtual layerType_t getLayerType() { return LAYER_YOLO; };
int classes, num, n_masks;
int classes, num, n_masks, new_coords;
dnnType *mask_h, *mask_d; //anchors
dnnType *bias_h, *bias_d; //anchors
float scaleXY;
double nms_thresh;
nmsKind_t nsm_kind;
std::vector<std::string> classesNames;
virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
int computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh);
int computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh, int new_coords=0);
dnnType *predictions;
static const int MAX_DETECTIONS = 8192;
static const int MAX_DETECTIONS = 8192*2;
static Yolo::detection *allocateDetections(int nboxes, int classes);
static void mergeDetections(Yolo::detection *dets, int ndets, int classes);
static void mergeDetections(Yolo::detection *dets, int ndets, int classes, double nms_thresh=0.45, nmsKind_t nsm_kind=GREEDY_NMS);
};
/**
include/tkDNN/MobilenetDetection.h
+1 -1
View File
@@ -65,7 +65,7 @@ public:
MobilenetDetection() {};
~MobilenetDetection() {};
bool init(const std::string& tensor_path, const int n_classes, const int n_batches=1);
bool init(const std::string& tensor_path, const int n_classes, const int n_batches=1, const float conf_thresh=0.3);
void preprocess(cv::Mat &frame, const int bi=0);
void postprocess(const int bi=0,const bool mAP=false);
};
include/tkDNN/Network.h
+4 -4
View File
@@ -7,12 +7,12 @@
namespace tk { namespace dnn {
/**
Data rapresentation beetween layers
Data representation between layers
n = batch size
c = channels
h = heigth (lines)
h = height (lines)
w = width (rows)
l = lenght (3rd dimension)
l = length (3rd dimension)
*/
struct dataDim_t {
@@ -43,7 +43,7 @@ public:
void releaseLayers();
/**
Do inferece for every added layer
Do inference for every added layer
*/
dnnType* infer(dataDim_t &dim, dnnType* data);
include/tkDNN/NetworkRT.h
+1 -1
View File
@@ -91,7 +91,7 @@ public:
}
/**
Do inferece
Do inference
*/
dnnType* infer(dataDim_t &dim, dnnType* data);
void enqueue(int batchSize = 1);
include/tkDNN/Yolo3Detection.h
+1 -1
View File
@@ -24,7 +24,7 @@ public:
Yolo3Detection() {};
~Yolo3Detection() {};
bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1);
bool init(const std::string& tensor_path, const int n_classes=80, const int n_batches=1, const float conf_thresh=0.3);
void preprocess(cv::Mat &frame, const int bi=0);
void postprocess(const int bi=0,const bool mAP=false);
};
include/tkDNN/evaluation.h
+4 -4
View File
@@ -73,12 +73,12 @@ double computeMap( std::vector<Frame> &images,const int classes,
* all the recall levels are evaluated, otherwise only
* map_point recall levels are used. For COCO evaluation
* 101 points are used.
* @param map_step step used to increment IoU theshold
* @param map_step step used to increment IoU threshold
* @param map_levels number of IoU step to perform
* @param verbose is set to true, prints on screen additional info
* @param write_on_file if set to true, the results produced by this function
* are written on file
* @param net name of the considerd neural network
* @param net name of the considered neural network
*
* @return mAP IoU_tresh:IoU_tresh+map_step*map_levels (e.g. mAP 0.5:0.95 when
* map_step=0.05 and map_levels=10)
@@ -89,7 +89,7 @@ double computeMapNIoULevels(std::vector<Frame> &images,const int classes,
const int map_levels=10, const bool verbose=false,
const bool write_on_file = false, std::string net = "");
/**
* This method computes the numper of True Positive (TP), False Positive (FP),
* This method computes the number of True Positive (TP), False Positive (FP),
* False Negative (FN), precision, recall and f1-score.
* Those values are computer over all the detections, over all the classes.
*
@@ -101,7 +101,7 @@ double computeMapNIoULevels(std::vector<Frame> &images,const int classes,
* @param verbose is set to true, prints on screen additional info
* @param write_on_file if set to true, the results produced by this function
* are written on file
* @param net name of the considerd neural network
* @param net name of the considered neural network
*/
void computeTPFPFN( std::vector<Frame> &images,const int classes,
const float IoU_thresh=0.5, const float conf_thresh=0.3,
include/tkDNN/pluginsRT/DeformableConvRT.h
+1 -1
View File
@@ -89,7 +89,7 @@ public:
for(int b=0; b<batchSize; b++) {
checkCuda(cudaMemcpy(offset, output_conv + b * 3 * chunk_dim, 2*chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(mask, output_conv + b * 3 * chunk_dim + 2*chunk_dim, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// kernel sigmoide
// kernel sigmoid
activationSIGMOIDForward(mask, mask, chunk_dim);
// deformable convolution
dcnV2CudaForward(stat, handle,
include/tkDNN/pluginsRT/YoloRT.h
+16 -4
View File
@@ -8,12 +8,15 @@ class YoloRT : public IPlugin {
public:
YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr, int n_masks=3, float scale_xy=1) {
YoloRT(int classes, int num, tk::dnn::Yolo *yolo = nullptr, int n_masks=3, float scale_xy=1, float nms_thresh=0.45, int nms_kind=0, int new_coords=0) {
this->classes = classes;
this->num = num;
this->n_masks = n_masks;
this->scaleXY = scale_xy;
this->nms_thresh = nms_thresh;
this->nms_kind = nms_kind;
this->new_coords = new_coords;
mask = new dnnType[n_masks];
bias = new dnnType[num*n_masks*2];
@@ -64,7 +67,10 @@ public:
for (int b = 0; b < batchSize; ++b){
for(int n = 0; n < n_masks; ++n){
int index = entry_index(b, n*w*h, 0);
activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream);
if (new_coords == 1)
activationLOGISTICForward(srcData + index, dstData + index, 4*w*h, stream); //x,y,w,h
else
activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream); //x,y
if (this->scaleXY != 1) scalAdd(dstData + index, 2 * w*h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
@@ -79,7 +85,7 @@ public:
virtual size_t getSerializationSize() override {
return 6*sizeof(int) + sizeof(float)+ n_masks*sizeof(dnnType) + num*n_masks*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char);
return 8*sizeof(int) + 2*sizeof(float)+ n_masks*sizeof(dnnType) + num*n_masks*2*sizeof(dnnType) + YOLORT_CLASSNAME_W*classes*sizeof(char);
}
virtual void serialize(void* buffer) override {
@@ -87,10 +93,13 @@ public:
tk::dnn::writeBUF(buf, classes);
tk::dnn::writeBUF(buf, num);
tk::dnn::writeBUF(buf, n_masks);
tk::dnn::writeBUF(buf, scaleXY);
tk::dnn::writeBUF(buf, nms_thresh);
tk::dnn::writeBUF(buf, nms_kind);
tk::dnn::writeBUF(buf, new_coords);
tk::dnn::writeBUF(buf, c);
tk::dnn::writeBUF(buf, h);
tk::dnn::writeBUF(buf, w);
tk::dnn::writeBUF(buf, scaleXY);
for(int i=0; i<n_masks; i++)
tk::dnn::writeBUF(buf, mask[i]);
for(int i=0; i<n_masks*2*num; i++)
@@ -109,6 +118,9 @@ public:
int c, h, w;
int classes, num, n_masks;
float scaleXY;
float nms_thresh;
int nms_kind;
int new_coords;
std::vector<std::string> classesNames;
dnnType *mask;
include/tkDNN/test.h
+1 -1
View File
@@ -20,7 +20,7 @@ int testInference(std::vector<std::string> input_bins, std::vector<std::string>
}
if(output_bins.size() != outputs.size()) {
std::cout<<output_bins.size()<<" "<<outputs.size()<<"\n";
FatalError("outputs size missmatch");
FatalError("outputs size mismatch");
}
// Load input
scripts/test_all_tests.sh
+1
View File
@@ -73,6 +73,7 @@ do
print_output $? imuodom
test_net yolo4
test_net yolo4x
test_net yolo4_berkeley
test_net yolo4tiny
test_net yolo3
src/CenternetDetection.cpp
+6 -5
View File
@@ -3,11 +3,12 @@
namespace tk { namespace dnn {
bool CenternetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches){
bool CenternetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches, const float conf_thresh){
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
classes = n_classes;
nBatches = n_batches;
confThreshold = conf_thresh;
dim = netRT->input_dim;
@@ -371,10 +372,10 @@ void CenternetDetection::postprocess(const int bi, const bool mAP){
// std::cout<<"th: "<<scores[j]<<" - cl: "<<clses[j]<<" i: "<<i<<std::endl;
//add coco bbox
//det[0:4], i, det[4]
int x0 = target_coords[j*4];
int y0 = target_coords[j*4+1];
int x1 = target_coords[j*4+2];
int y1 = target_coords[j*4+3];
float x0 = target_coords[j*4];
float y0 = target_coords[j*4+1];
float x1 = target_coords[j*4+2];
float y1 = target_coords[j*4+3];
int obj_class = clses[j];
float prob = scores[j];
// std::cout<<"("<<x0<<", "<<y0<<"),("<<x1<<", "<<y1<<")"<<std::endl;
src/DarknetParser.cpp
+13 -3
View File
@@ -37,7 +37,10 @@ namespace tk { namespace dnn {
std::string name,value;
if(!divideNameAndValue(line, name, value))
return false;
if(name.find("width") != std::string::npos)
if(name.find("new_coords") != std::string::npos)
fields.new_coords = std::stoi(value);
else if(name.find("width") != std::string::npos)
fields.width = std::stoi(value);
else if(name.find("height") != std::string::npos)
fields.height = std::stoi(value);
@@ -79,6 +82,13 @@ namespace tk { namespace dnn {
fields.group_id = std::stoi(value);
else if(name.find("scale_x_y") != std::string::npos)
fields.scale_xy = std::stof(value);
else if(name.find("beta_nms") != std::string::npos)
fields.nms_thresh = std::stof(value);
else if(name.find("nms_kind") != std::string::npos){
if(value == "greedynms") fields.nms_kind = 0;
else if(value == "diounms") fields.nms_kind = 1;
else std::cout<<"Not supported nms_kind "<<value<<", setting to greedynms"<<std::endl;
}
else if(name.find("from") != std::string::npos)
fields.layers.push_back(std::stof(value));
else if(name.find("mask") != std::string::npos){
@@ -161,7 +171,7 @@ namespace tk { namespace dnn {
} else if(f.type == "yolo") {
std::string wgs = wgs_path + "/g" + std::to_string(netLayers.size()) + ".bin";
//printf("%d %d %s %d %f\n", f.classes, f.num/f.n_mask, wgs.c_str(), f.n_mask, f.scale_xy);
tk::dnn::Yolo *l = new tk::dnn::Yolo(net, f.classes, f.num/f.n_mask, wgs, f.n_mask, f.scale_xy);
tk::dnn::Yolo *l = new tk::dnn::Yolo(net, f.classes, f.num/f.n_mask, wgs, f.n_mask, f.scale_xy, f.nms_thresh, (tk::dnn::Yolo::nmsKind_t) f.nms_kind, f.new_coords);
if(names.size() != f.classes)
FatalError("Mismatch between number of classes and names");
l->classesNames = names;
@@ -201,7 +211,7 @@ namespace tk { namespace dnn {
tk::dnn::Network *net = nullptr;
// layers without activations to retrive correct id number
// layers without activations to retrieve correct id number
std::vector<tk::dnn::Layer*> netLayers;
std::ifstream if_cfg(cfg_file);
src/DeformConv2d.cpp
+1 -1
View File
@@ -95,7 +95,7 @@ dnnType* DeformConv2d::infer(dataDim_t &dim, dnnType* srcData) {
// split conv2d outputs into offset and mask
checkCuda(cudaMemcpy(offset, output_conv, 2*chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(mask, output_conv + 2*chunk_dim, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// kernel sigmoide
// kernel sigmoid
activationSIGMOIDForward(mask, mask, chunk_dim);
// deformable convolution
src/Dense.cpp
+1 -1
View File
@@ -37,7 +37,7 @@ dnnType* Dense::infer(dataDim_t &dim, dnnType* srcData) {
// place bias into dstData
checkCuda( cudaMemcpy(dstData, bias_d, dim_y*sizeof(dnnType), cudaMemcpyDeviceToDevice) );
//do matrix moltiplication
//do matrix multiplication
checkERROR( cublasSgemv(net->cublasHandle, CUBLAS_OP_T,
dim_x, dim_y,
&alpha,
src/LSTM.cpp
+5 -5
View File
@@ -133,7 +133,7 @@ LSTM::LSTM( Network *net, int hiddensize, bool returnSeq, std::string fname_weig
output_dim = input_dim;
output_dim.c = stateSize*(bidirectional ? 2 : 1);
// if retunseq is disabled only the last timestep is returned
// if retunseq is disabled only the last timestamp is returned
if(!returnSeq) {
output_dim.h = 1;
output_dim.w = 1;
@@ -254,7 +254,7 @@ dnnType* LSTM::infer(dataDim_t &dim, dnnType* srcData) {
rnnDesc,
seqLen, // number of time steps (nT)
x_desc_vec_.data(), // input array of desc (nT*nC_in)
srcF, // input pointer
srcF, // input pointer
hx_desc_, // initial hidden state desc
hx_ptr, // initial hidden state pointer
cx_desc_, // initial cell state desc
@@ -281,7 +281,7 @@ dnnType* LSTM::infer(dataDim_t &dim, dnnType* srcData) {
rnnDesc,
seqLen, // number of time steps (nT)
x_desc_vec_.data(), // input array of desc (nT*nC_in)
srcB, // input pointer
srcB, // input pointer
hx_desc_, // initial hidden state desc
hx_ptr, // initial hidden state pointer
cx_desc_, // initial cell state desc
@@ -289,7 +289,7 @@ dnnType* LSTM::infer(dataDim_t &dim, dnnType* srcData) {
w_desc_, // weights desc
wb_ptr, // weights pointer
y_desc_vec_.data(), // output desc (nT*nC_out)
dstB_NR, // output pointer
dstB_NR, // output pointer
hy_desc_, // final hidden state desc
hy_ptr, // final hidden state pointer
cy_desc_, // final cell state desc
@@ -307,7 +307,7 @@ dnnType* LSTM::infer(dataDim_t &dim, dnnType* srcData) {
one_output_dim.c*sizeof(dnnType), cudaMemcpyDeviceToDevice));
}
// if retunseq is disabled only the last timestep is returned
// if retunseq is disabled only the last timestamp is returned
if(returnSeq) {
// forward transpose
matrixTranspose(net->cublasHandle, dstF, dstData,
src/LayerWgs.cpp
+1 -1
View File
@@ -105,7 +105,7 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
float2half(tmp_d, variance16_d, b_size);
cudaMemcpy(variance16_h, variance16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
//conver scales
//convert scales
float2half(scales_d, scales16_d, b_size);
cudaMemcpy(scales16_h, scales16_d, b_size*sizeof(__half), cudaMemcpyDeviceToHost);
src/MobilenetDetection.cpp
+2 -1
View File
@@ -126,12 +126,13 @@ float MobilenetDetection::iou(const tk::dnn::box &a, const tk::dnn::box &b){
return iou;
}
bool MobilenetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches){
bool MobilenetDetection::init(const std::string& tensor_path, const int n_classes, const int n_batches, const float conf_thresh){
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str());
imageSize = netRT->input_dim.h;
classes = n_classes;
nBatches = n_batches;
confThreshold = conf_thresh;
SSDSpec specs[N_SSDSPEC];
src/MulAdd.cpp
+1 -1
View File
@@ -12,7 +12,7 @@ MulAdd::MulAdd(Network *net, dnnType mul, dnnType add) : Layer(net) {
int size = input_dim.tot();
// create a vector with all value setted to add
// create a vector with all value set to add
dnnType *add_vector_h = new dnnType[size];
for(int i=0; i<size; i++)
add_vector_h[i] = add;
src/NetworkRT.cpp
+9 -5
View File
@@ -163,7 +163,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
// note that indices are guaranteed to be less than IEngine::getNbBindings()
buf_input_idx = engineRT->getBindingIndex("data");
buf_output_idx = engineRT->getBindingIndex("out");
std::cout<<"input idex = "<<buf_input_idx<<" -> output index = "<<buf_output_idx<<"\n";
std::cout<<"input index = "<<buf_input_idx<<" -> output index = "<<buf_output_idx<<"\n";
Dims iDim = engineRT->getBindingDimensions(buf_input_idx);
@@ -529,7 +529,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Yolo *l) {
//std::cout<<"convert Yolo\n";
//std::cout<<"New plugin YOLO\n";
IPlugin *plugin = new YoloRT(l->classes, l->num, l, l->n_masks, l->scaleXY);
IPlugin *plugin = new YoloRT(l->classes, l->num, l, l->n_masks, l->scaleXY, l->nms_thresh, l->nsm_kind, l->new_coords);
IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
checkNULL(lRT);
return lRT;
@@ -739,12 +739,16 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
if(name.find("Yolo") == 0) {
YoloRT *r = new YoloRT(readBUF<int>(buf), //classes
readBUF<int>(buf), //num
nullptr,
readBUF<int>(buf)); //n_masks
nullptr, //yolo
readBUF<int>(buf), //n_masks
readBUF<float>(buf), //scale_xy
readBUF<float>(buf), //nms_thresh
readBUF<int>(buf), //nms_kind
readBUF<int>(buf) //new_coords
);
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
r->scaleXY = readBUF<float>(buf);
for(int i=0; i<r->n_masks; i++)
r->mask[i] = readBUF<dnnType>(buf);
for(int i=0; i<r->n_masks*2*r->num; i++)
src/Region.cpp
+1 -1
View File
@@ -63,7 +63,7 @@ dnnType* Region::infer(dataDim_t &dim, dnnType* srcData) {
}
/* Intepret class */
/* Interpret class */
RegionInterpret::RegionInterpret(dataDim_t input_dim, dataDim_t output_dim,
int classes, int coords, int num, float thresh, std::string fname_weights) {
src/Shortcut.cpp
+1 -1
View File
@@ -13,7 +13,7 @@ Shortcut::Shortcut(Network *net, Layer *backLayer) : Layer(net) {
if( /*backLayer->output_dim.c != input_dim.c ||*/
backLayer->output_dim.w != input_dim.w ||
backLayer->output_dim.h != input_dim.h )
FatalError("Shortcut dim missmatch");
FatalError("Shortcut dim mismatch");
}
Shortcut::~Shortcut() {
src/Yolo.cpp
+54 -14
View File
@@ -11,7 +11,7 @@
namespace tk { namespace dnn {
Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks, float scale_xy) :
Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_masks, float scale_xy, double nms_thresh, nmsKind_t nsm_kind, int new_coords) :
Layer(net) {
this->final = true;
@@ -19,6 +19,9 @@ Yolo::Yolo(Network *net, int classes, int num, std::string fname_weights, int n_
this->num = num;
this->n_masks = n_masks;
this->scaleXY = scale_xy;
this->nms_thresh = nms_thresh;
this->nsm_kind = nsm_kind;
this->new_coords = new_coords;
// load anchors
if(fname_weights != "") {
@@ -59,12 +62,21 @@ int entry_index(int batch, int location, int entry,
entry*input_dim.w*input_dim.h + loc;
}
Yolo::box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride) {
Yolo::box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride, int new_coords) {
Yolo::box b;
b.x = (i + x[index + 0*stride]) / lw;
b.y = (j + x[index + 1*stride]) / lh;
b.w = exp(x[index + 2*stride]) * biases[2*n] / w;
b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h;
if(new_coords == 0){
b.x = (i + x[index + 0*stride]) / lw;
b.y = (j + x[index + 1*stride]) / lh;
b.w = exp(x[index + 2*stride]) * biases[2*n] / w;
b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h;
}
else{
b.x = (i + x[index + 0 * stride] * 2 - 0.5) / lw;
b.y = (j + x[index + 1 * stride] * 2 - 0.5) / lh;
b.w = x[index + 2 * stride] * x[index + 2 * stride] * 4 * biases[2 * n] / w;
b.h = x[index + 3 * stride] * x[index + 3 * stride] * 4 * biases[2 * n + 1] / h;
}
return b;
}
@@ -75,7 +87,10 @@ dnnType* Yolo::infer(dataDim_t &dim, dnnType* srcData) {
for (int b = 0; b < dim.n; ++b){
for(int n = 0; n < n_masks; ++n){
int index = entry_index(b, n*dim.w*dim.h, 0, classes, input_dim, output_dim);
activationLOGISTICForward(srcData + index, dstData + index, 2*dim.w*dim.h);
if (new_coords == 1)
activationLOGISTICForward(srcData + index, dstData + index, 4*dim.w*dim.h);
else
activationLOGISTICForward(srcData + index, dstData + index, 2*dim.w*dim.h);
if (this->scaleXY != 1) scalAdd(dstData + index, 2 * dim.w*dim.h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
@@ -116,7 +131,7 @@ void correct_yolo_boxes(Yolo::detection *dets, int n, int w, int h, int netw, in
}
}
int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh) {
int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int neth, float thresh, int new_coords) {
if(predictions == nullptr)
predictions = new dnnType[output_dim.tot()];
@@ -140,7 +155,7 @@ int Yolo::computeDetections(Yolo::detection *dets, int &ndets, int netw, int net
if(objectness <= thresh) continue;
int box_index = entry_index(0, n*lw*lh + i, 0, classes, input_dim, output_dim);
dets[count].bbox = get_yolo_box(predictions, bias_h, mask_h[n], box_index, col, row, lw, lh, netw, neth, lw*lh);
dets[count].bbox = get_yolo_box(predictions, bias_h, mask_h[n], box_index, col, row, lw, lh, netw, neth, lw*lh, new_coords);
dets[count].objectness = objectness;
dets[count].classes = classes;
for(j = 0; j < classes; ++j){
@@ -193,6 +208,32 @@ float yolo_box_iou(Yolo::box a, Yolo::box b)
return yolo_box_intersection(a, b)/yolo_box_union(a, b);
}
void box_c(const Yolo::box a, const Yolo::box b, float& top, float& bot, float& left, float& right) {
top = std::min(a.y - a.h / 2, b.y - b.h / 2);
bot = std::max(a.y + a.h / 2, b.y + b.h / 2);
left = std::min(a.x - a.w / 2, b.x - b.w / 2);
right = std::max(a.x + a.w / 2, b.x + b.w / 2);
}
// https://github.com/Zzh-tju/DIoU-darknet
// https://arxiv.org/abs/1911.08287
float yolo_box_diou(const Yolo::box a, const Yolo::box b, const float nms_thresh=0.6)
{
float top, bot, left, right;
box_c(a, b, top, bot, left, right);
float w = right - left;
float h = bot - top;
float c = w * w + h * h;
float iou = yolo_box_iou(a, b);
if (c == 0)
return iou;
float d = (a.x - b.x) * (a.x - b.x) + (a.y - b.y) * (a.y - b.y);
float u = pow(d / c, nms_thresh);
float diou_term = u;
return iou - diou_term;
}
int yolo_nms_comparator(const void *pa, const void *pb)
{
Yolo::detection a = *(Yolo::detection *)pa;
@@ -219,8 +260,7 @@ Yolo::detection *Yolo::allocateDetections(int nboxes, int classes) {
return dets;
}
void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes) {
double nms_thresh = 0.45;
void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes, double nms_thresh, nmsKind_t nsm_kind) {
int total = ndets;
int i, j, k;
@@ -246,13 +286,13 @@ void Yolo::mergeDetections(Yolo::detection *dets, int ndets, int classes) {
box a = dets[i].bbox;
for(j = i+1; j < total; ++j){
box b = dets[j].bbox;
if (yolo_box_iou(a, b) > nms_thresh){
if (nsm_kind == GREEDY_NMS && yolo_box_iou(a, b) > nms_thresh)
dets[j].prob[k] = 0;
else if (nsm_kind == DIOU_NMS && yolo_box_diou(a, b, nms_thresh) > nms_thresh)
dets[j].prob[k] = 0;
}
}
}
}
}
}}
src/Yolo3Detection.cpp
+35 -30
View File
@@ -3,13 +3,14 @@
namespace tk { namespace dnn {
bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes, const int n_batches) {
bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes, const int n_batches, const float conf_thresh) {
//convert network to tensorRT
std::cout<<(tensor_path).c_str()<<"\n";
netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str() );
nBatches = n_batches;
confThreshold = conf_thresh;
tk::dnn::dataDim_t idim = netRT->input_dim;
idim.n = nBatches;
@@ -31,6 +32,9 @@ bool Yolo3Detection::init(const std::string& tensor_path, const int n_classes, c
memcpy(yolo[i]->bias_h, yRT->bias, sizeof(dnnType)*num*nMasks*2);
yolo[i]->input_dim = yolo[i]->output_dim = tk::dnn::dataDim_t(1, yRT->c, yRT->h, yRT->w);
yolo[i]->classesNames = yRT->classesNames;
yolo[i]->nms_thresh = yRT->nms_thresh;
yolo[i]->nsm_kind = (tk::dnn::Yolo::nmsKind_t) yRT->nms_kind;
yolo[i]->new_coords = yRT->new_coords;
}
dets = tk::dnn::Yolo::allocateDetections(tk::dnn::Yolo::MAX_DETECTIONS, classes);
@@ -101,46 +105,47 @@ void Yolo3Detection::postprocess(const int bi, const bool mAP){
nDets = 0;
for(int i=0; i<netRT->pluginFactory->n_yolos; i++) {
yolo[i]->dstData = rt_out[i];
yolo[i]->computeDetections(dets, nDets, netRT->input_dim.w, netRT->input_dim.h, confThreshold);
yolo[i]->computeDetections(dets, nDets, netRT->input_dim.w, netRT->input_dim.h, confThreshold, yolo[i]->new_coords);
}
tk::dnn::Yolo::mergeDetections(dets, nDets, classes);
tk::dnn::Yolo::mergeDetections(dets, nDets, classes, yolo[0]->nms_thresh, yolo[0]->nsm_kind);
// fill detected
detected.clear();
for(int j=0; j<nDets; j++) {
tk::dnn::Yolo::box b = dets[j].bbox;
int x0 = (b.x-b.w/2.);
int x1 = (b.x+b.w/2.);
int y0 = (b.y-b.h/2.);
int y1 = (b.y+b.h/2.);
int obj_class = -1;
float prob = 0;
float x0 = (b.x-b.w/2.);
float x1 = (b.x+b.w/2.);
float y0 = (b.y-b.h/2.);
float y1 = (b.y+b.h/2.);
// convert to image coords
x0 = x_ratio*x0;
x1 = x_ratio*x1;
y0 = y_ratio*y0;
y1 = y_ratio*y1;
for(int c=0; c<classes; c++) {
if(dets[j].prob[c] >= confThreshold) {
obj_class = c;
prob = dets[j].prob[c];
int obj_class = c;
float prob = dets[j].prob[c];
tk::dnn::box res;
res.cl = obj_class;
res.prob = prob;
res.x = x0;
res.y = y0;
res.w = x1 - x0;
res.h = y1 - y0;
// FIXME: this shuld be useless
// if(mAP)
// for(int c=0; c<classes; c++)
// res.probs.push_back(dets[j].prob[c]);
detected.push_back(res);
}
}
if(obj_class >= 0) {
// convert to image coords
x0 = x_ratio*x0;
x1 = x_ratio*x1;
y0 = y_ratio*y0;
y1 = y_ratio*y1;
tk::dnn::box res;
res.cl = obj_class;
res.prob = prob;
res.x = x0;
res.y = y0;
res.w = x1 - x0;
res.h = y1 - y0;
if(mAP)
for(int c=0; c<classes; c++)
res.probs.push_back(dets[j].prob[c]);
detected.push_back(res);
}
}
batchDetected.push_back(detected);
}
src/evaluation.cpp
+3 -3
View File
@@ -63,7 +63,7 @@ double computeMap( std::vector<Frame> &images,const int classes,
int gt_checked = 0;
// for each detection comput IoU with groundtruth and match detetcion and
// for each detection compute IoU with groundtruth and match detetcion and
// groundtruth with IoU greater than IoU_thresh
for(auto &img:images){
for(size_t i=0; i<img.det.size(); i++){
@@ -153,7 +153,7 @@ double computeMap( std::vector<Frame> &images,const int classes,
}
}
//compute average precision for each class. Two methods are avaible,
//compute average precision for each class. Two methods are available,
//based on map_points required
double mean_average_precision = 0;
double last_recall, last_precision, delta_recall;
@@ -287,7 +287,7 @@ void computeTPFPFN( std::vector<Frame> &images,const int classes,
}
}
//count all TP, FP, FN and compute precsion, recall and f1-score
//count all TP, FP, FN and compute precision, recall and f1-score
double avg_precision = 0, avg_recall = 0, f1_score = 0;
int TP = 0, FP = 0, FN = 0;
for(size_t i=0; i<classes; i++){
tests/darknet/cfg/yolo4_mmr.cfg
+1158
View File
File diff suppressed because it is too large Load Diff
tests/darknet/cfg/yolo4x.cfg
+1427
View File
File diff suppressed because it is too large Load Diff
tests/darknet/names/mmr.names
+4
View File
@@ -0,0 +1,4 @@
blue-cone
yellow-cone
orange-cone
big-orange-cone
tests/darknet/yolo4_mmr.cpp
+34
View File
@@ -0,0 +1,34 @@
#include<iostream>
#include<vector>
#include "tkdnn.h"
#include "test.h"
#include "DarknetParser.h"
int main() {
std::string bin_path = "yolo4_mmr";
std::vector<std::string> input_bins = {
bin_path + "/layers/input.bin"
};
std::vector<std::string> output_bins = {
bin_path + "/debug/layer139_out.bin",
bin_path + "/debug/layer150_out.bin",
bin_path + "/debug/layer161_out.bin"
};
std::string wgs_path = bin_path + "/layers";
std::string cfg_path = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4_mmr.cfg";
std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/mmr.names";
// downloadWeightsifDoNotExist(input_bins[0], bin_path, "");
// parse darknet network
tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
net->print();
//convert network to tensorRT
tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
int ret = testInference(input_bins, output_bins, net, netRT);
net->releaseLayers();
delete net;
delete netRT;
return ret;
}
tests/darknet/yolo4x.cpp
+36
View File
@@ -0,0 +1,36 @@
#include<iostream>
#include<vector>
#include "tkdnn.h"
#include "test.h"
#include "DarknetParser.h"
int main() {
std::string bin_path = "yolo4x";
std::vector<std::string> input_bins = {
bin_path + "/layers/input.bin"
};
std::vector<std::string> output_bins = {
bin_path + "/debug/layer168_out.bin",
bin_path + "/debug/layer185_out.bin",
bin_path + "/debug/layer202_out.bin"
};
std::string wgs_path = bin_path + "/layers";
std::string cfg_path = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4x.cfg";
std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/coco.names";
downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/BLPpiAigZJLorQD/download");
// parse darknet network
tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
net->print();
//convert network to tensorRT
tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
int ret = testInference(input_bins, output_bins, net, netRT);
net->releaseLayers();
delete net;
delete netRT;
return ret;
}