merge with master

Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
2021-07-20 18:38:46 +02:00
parent 4b3731928c 946b1fe5f5
commit 318dfe4b55
68 changed files with 5577 additions and 276 deletions
@@ -12,5 +12,11 @@ build/
 *.hdf5
 *.pk
 *.table
+cmake-build-release/
 demo/COCO_val2017
-demo/BDD100K_val
+demo/BDD100K_val
+/.vs
+cmake-build-minsizerel/*
+scripts/COCO_val2017/*
+scripts/COCO_val2017.zip
+scripts/all_labels.txt
@@ -1,8 +1,15 @@
-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.15)

 project (tkDNN)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -fPIC -Wno-deprecated-declarations -Wno-unused-variable")
+if(UNIX)
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -fPIC -Wno-deprecated-declarations -Wno-unused-variable ")
+endif()
+if(WIN32)
+set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_FLAGS "/O2 /FS /EHsc")
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+endif(WIN32)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include/tkDNN)

 # project specific flags
@@ -10,7 +17,13 @@ if(DEBUG)
    add_definitions(-DDEBUG)
 endif()

-add_definitions(-DTKDNN_PATH="${CMAKE_CURRENT_SOURCE_DIR}")
+if(TKDNN_PATH)
+    message("SET TKDNN_PATH:"${TKDNN_PATH})
+    add_definitions(-DTKDNN_PATH="${TKDNN_PATH}")
+else()
+    add_definitions(-DTKDNN_PATH="${CMAKE_CURRENT_SOURCE_DIR}")
+endif()
+

 #-------------------------------------------------------------------------------
 # CUDA
@@ -28,19 +41,21 @@ include_directories(${CUDNN_INCLUDE_DIR})
 file(GLOB tkdnn_CUSRC "src/kernels/*.cu" "src/sorting.cu")
 cuda_include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${CUDNN_INCLUDE_DIRS})
 cuda_add_library(kernels SHARED ${tkdnn_CUSRC})
+target_link_libraries(kernels ${CUDA_CUBLAS_LIBRARIES})


 #-------------------------------------------------------------------------------
 # External Libraries
 #-------------------------------------------------------------------------------
 find_package(Eigen3 REQUIRED)
+message("Eigen DIR: " ${EIGEN3_INCLUDE_DIR})
 include_directories(${EIGEN3_INCLUDE_DIR})

 find_package(OpenCV REQUIRED)
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DOPENCV")

 # gives problems in cross-compiling, probably malformed cmake config
-#find_package(yaml-cpp REQUIRED)
+find_package(yaml-cpp REQUIRED)

 #-------------------------------------------------------------------------------
 # Build Libraries
@@ -48,7 +63,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DOPENCV")
 file(GLOB tkdnn_SRC "src/*.cpp")
 set(tkdnn_LIBS kernels ${CUDA_LIBRARIES} ${CUDA_CUBLAS_LIBRARIES} ${CUDNN_LIBRARIES} ${OpenCV_LIBS} yaml-cpp)

-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include ${CUDA_INCLUDE_DIRS} ${OPENCV_INCLUDE_DIRS} ${NVINFER_INCLUDES})
 add_library(tkDNN SHARED ${tkdnn_SRC})
 target_link_libraries(tkDNN ${tkdnn_LIBS})
@@ -77,6 +92,7 @@ foreach(test_SRC ${darknet_SRC})
    set(test_NAME test_${test_NAME})
    add_executable(${test_NAME} ${test_SRC})
    target_link_libraries(${test_NAME} tkDNN)
+    install(TARGETS ${test_NAME} DESTINATION bin)
 endforeach()

 # MOBILENET
@@ -103,6 +119,16 @@ target_link_libraries(test_resnet101_cnet tkDNN)
 add_executable(test_dla34_cnet tests/centernet/dla34_cnet/dla34_cnet.cpp)
 target_link_libraries(test_dla34_cnet tkDNN)

+# SHELFNET
+add_executable(test_shelfnet tests/shelfnet/shelfnet.cpp)
+target_link_libraries(test_shelfnet tkDNN)
+
+add_executable(test_shelfnet_berkeley tests/shelfnet/shelfnet_berkeley.cpp)
+target_link_libraries(test_shelfnet_berkeley tkDNN)
+
+add_executable(test_shelfnet_mapillary tests/shelfnet/shelfnet_mapillary.cpp)
+target_link_libraries(test_shelfnet_mapillary tkDNN)
+
 # DEMOS
 add_executable(test_rtinference tests/test_rtinference/rtinference.cpp)
 target_link_libraries(test_rtinference tkDNN)
@@ -113,6 +139,9 @@ target_link_libraries(map_demo tkDNN)
 add_executable(demo demo/demo/demo.cpp)
 target_link_libraries(demo tkDNN)

+add_executable(seg_demo demo/demo/seg_demo.cpp)
+target_link_libraries(seg_demo tkDNN)
+
 #-------------------------------------------------------------------------------
 # Install
 #-------------------------------------------------------------------------------
@@ -123,7 +152,10 @@ target_link_libraries(demo tkDNN)
 message("install dir:" ${CMAKE_INSTALL_PREFIX})
 install(DIRECTORY include/ DESTINATION include/)
 install(TARGETS tkDNN kernels DESTINATION lib)
+install(TARGETS test_simple test_mnist test_mnistRT test_rtinference demo map_demo DESTINATION bin)
 install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/cmake/" # source directory
        DESTINATION "share/tkDNN/cmake/" # target directory
 )
-
+install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/tests/" # source directory
+        DESTINATION "share/tkDNN/tests" # target directory
+)
@@ -0,0 +1 @@
+1)error C2131 @ Yolo3Detection.cpp(97) -> expression doesnt evaluate to a constant caused to read of variable outside its lifetime
@@ -3,42 +3,54 @@ 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}
+}
 ```

+### What's new (20 July 2021)
+- [x] Support to sematic segmentation [README](docs/README_seg.md)
+- [ ] Support to TensorRT8 (WIP)
+
 ## 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
@@ -72,17 +84,30 @@ Results for COCO val 2017 (5k images), on RTX 2080Ti, with conf threshold=0.001
  - [mAP demo](#map-demo)
  - [Existing tests and supported networks](#existing-tests-and-supported-networks)
  - [References](#references)
-
-
-
+  - [tkDNN on Windows 10 (experimental)](#tkdnn-on-windows-10-experimental)
+    - [Dependencies-Windows](#dependencies-windows)
+    - [Compiling tkDNN on Windows](#compiling-tkdnn-on-windows)
+    - [Run the demo on Windows](#run-the-demo-on-windows)
+      - [FP16 inference windows](#fp16-inference-windows)
+      - [INT8 inference windows](#int8-inference-windows)
+    - [Known issues with tkDNN on Windows](#known-issues-with-tkdnn-on-windows)
+  

 ## Dependencies
-This branch works on every NVIDIA GPU that supports the dependencies:
-* CUDA 10.0
-* CUDNN 7.603
-* TENSORRT 6.01
-* OPENCV 3.4
-* yaml-cpp 0.5.2 (sudo apt install libyaml-cpp-dev)
+This branch works on every NVIDIA GPU that supports the following (latest tested) dependencies:
+* CUDA 11.0 (or >= 10)
+* cuDNN 8.0.4 (or >= 7.3)
+* TensorRT 7.2.0 (or >=5)
+* OpenCV 4.5.2 (or >=4)
+* cmake 3.21 (or >= 3.15)
+* yaml-cpp 0.5.2
+* eigen3 3.3.4
+* curl 7.58
+
+```
+sudo apt install libyaml-cpp-dev curl libeigen3-dev
+
+```

 ## About OpenCV
 To compile and install OpenCV4 with contrib us the script ```install_OpenCV4.sh```. It will download and compile OpenCV in Download folder.
@@ -187,6 +212,7 @@ All models from darknet are now parsed directly from cfg, you still need to expo
  relu
  leaky
  mish
+  logistic
 </details>

 ## Run the demo 
@@ -209,7 +235,7 @@ Once you have successfully created your rt file, run the demo:
 ```
 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>
+./demo <network-rt-file> <path-to-video> <kind-of-network> <number-of-classes> <n-batches> <show-flag> <conf-thresh>
 ```
 where
 *  ```<network-rt-file>``` is the rt file generated by a test
@@ -344,7 +370,97 @@ 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        | 640x640       | [weights](https://cloud.hipert.unimore.it/s/5MFjtNtgbDGdJEo/download)     |
+| yolo4x-cps            | Scaled Yolov4 <sup>10</sup>                          | [COCO 2017](http://cocodataset.org/)                          | 80        | 512x512       | [weights](https://cloud.hipert.unimore.it/s/AfzHE4BfTeEm2gH/download)     |
+
+### tkDNN on Windows 10 (experimental)
+
+### Dependencies-Windows 
+This branch should work on every NVIDIA GPU supported in windows with the following dependencies:
+
+* WINDOWS 10 1803 or HIGHER 
+* CUDA 10.0 (Recommended CUDA 11.2 )
+* CUDNN 7.6 (Recommended CUDNN 8.1.1 )
+* TENSORRT 6.0.1 (Recommended TENSORRT 7.2.3.4 )
+* OPENCV 3.4 (Recommended OPENCV 4.2.0 )
+* MSVC 16.7 
+* YAML-CPP 
+* EIGEN3
+* 7ZIP (ADD TO PATH)
+* NINJA 1.10
+
+
+All the above mentioned dependencies except 7ZIP can be installed using Microsoft's [VCPKG](https://github.com/microsoft/vcpkg.git) .
+After bootstrapping VCPKG the dependencies can be built and installed using the following command :
+
+```
+opencv4(normal) - vcpkg.exe install opencv4[tbb,jpeg,tiff,opengl,openmp,png,ffmpeg,eigen]:x64-windows yaml-cpp:x64-windows eigen3:x64-windows --x-install-root=C:\opt --x-buildtrees-root=C:\temp_vcpkg_build
+
+opencv4(cuda) - vcpkg.exe install opencv4[cuda,nonfree,contrib,eigen,tbb,jpeg,tiff,opengl,openmp,png,ffmpeg]:x64-windows yaml-cpp:x64-windows eigen3:x64-windows --x-install-root=C:\opt --x-buildtrees-root=C:\temp_vcpkg_build
+```
+To build opencv4 with cuda and cudnn version corresponding to your cuda version,vcpkg's cudnn portfile needs to be modified by adding ```$ENV{CUDA_PATH}```  at lines 16 and 17 in the portfile.cmake 
+
+After VCPKG finishes building and installing all the packages delete C:\temp_vcpkg_build and add C:\opt\x64-windows\bin and C:\opt\x64-windows\debug\bin to path 
+
+### Compiling tkDNN on Windows 
+
+tkDNN is built with cmake(3.15+) on windows along with ninja.Msbuild and NMake Makefiles are drastically slower when compiling the library compared to windows
+```
+git clone https://github.com/ceccocats/tkDNN.git
+cd tkdnn-windows
+mkdir build
+cd build
+cmake -DCMAKE_BUILD_TYPE=Release -G"Ninja" ..
+ninja -j4
+```
+
+### Run the demo on Windows 
+
+This example uses yolo4_tiny.\
+To run the object detection file create .rt file bu running:
+```
+.\test_yolo4tiny.exe
+```
+
+Once the rt file has been successfully create,run the demo using the following command:
+```
+.\demo.exe yolo4tiny_fp32.rt ..\demo\yolo_test.mp4 y 
+```
+ For general info on more demo paramters,check Run the demo section on top 
+ To run the test_all_tests.sh on windows,use git bash or msys2 
+
+### FP16 inference windows 
+
+This is an untested feature on windows.To run the object detection demo with FP16 interference follow the below steps(example with yolo4tiny):
+```
+set TKDNN_MODE=FP16
+del /f yolo4tiny_fp16.rt
+.\test_yolo4tiny.exe
+.\demo.exe yolo4tiny_fp16.rt ..\demo\yolo_test.mp4
+```
+
+### INT8 inference windows 
+To run object detection demo with INT8 (example with yolo4tiny):
+```
+set TKDNN_MODE=INT8
+set TKDNN_CALIB_LABEL_PATH=..\demo\COCO_val2017\all_labels.txt
+set TKDNN_CALIB_IMG_PATH=..\demo\COCO_val2017\all_images.txt
+del /f  yolo4tiny_int8.rt        # be sure to delete(or move) old tensorRT files
+.\test_yolo4tiny.exe           # run the yolo test (is slow)
+.\demo.exe yolo4tiny_int8.rt ..\demo\yolo_test.mp4 y
+
+```
+
+### Known issues with tkDNN on Windows
+
+Mobilenet and Centernet demos work properly only when built with msvc 16.7 in Release Mode,when built in debug mode for the mentioned networks one might encounter opencv assert errors
+
+All Darknet models work properly with demo using MSVC version(16.7-16.9)
+
+It is recommended to use Nvidia Driver(465+),Cuda unknown errors have been observed when using older drivers on pascal(SM 61) devices.
+
+


 ## References
@@ -357,3 +473,5 @@ 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)
+10. Wang, Chien-Yao, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. "Scaled-YOLOv4: Scaling Cross Stage Partial Network." arXiv preprint arXiv:2011.08036 (2020).
@@ -1,7 +1,7 @@
 #include <iostream>
 #include <signal.h>
 #include <stdlib.h>     /* srand, rand */
-#include <unistd.h>
+//#include <unistd.h>
 #include <mutex>

 #include "CenternetDetection.h"
@@ -22,10 +22,15 @@ int main(int argc, char *argv[]) {
    signal(SIGINT, sig_handler);


-    std::string net = "yolo3_berkeley.rt";
+    std::string net = "yolo4tiny_fp32.rt";
    if(argc > 1)
        net = argv[1]; 
-    std::string input = "../demo/yolo_test.mp4";
+    #ifdef __linux__ 
+        std::string input = "../demo/yolo_test.mp4";
+    #elif _WIN32
+        std::string input = "..\\..\\..\\demo\\yolo_test.mp4";
+    #endif
+
    if(argc > 2)
        input = argv[2]; 
    char ntype = 'y';
@@ -131,7 +136,7 @@ int main(int argc, char *argv[]) {
    double mean = 0; 
    
    std::cout<<COL_GREENB<<"\n\nTime stats:\n";
-    std::cout<<"Min: "<<*std::min_element(detNN->stats.begin(), detNN->stats.end())/n_batch<<" ms\n";    
+   std::cout<<"Min: "<<*std::min_element(detNN->stats.begin(), detNN->stats.end())/n_batch<<" ms\n";    
    std::cout<<"Max: "<<*std::max_element(detNN->stats.begin(), detNN->stats.end())/n_batch<<" ms\n";    
    for(int i=0; i<detNN->stats.size(); i++) mean += detNN->stats[i]; mean /= detNN->stats.size();
    std::cout<<"Avg: "<<mean/n_batch<<" ms\t"<<1000/(mean/n_batch)<<" FPS\n"<<COL_END;   
@@ -2,7 +2,10 @@
 #include <iostream>
 #include <signal.h>
 #include <stdlib.h>     /* srand, rand */
+#ifdef __linux__
 #include <unistd.h>
+#endif
+
 #include <mutex>
 #include "utils.h"

@@ -0,0 +1,148 @@
+#include <iostream>
+#include <signal.h>
+#include <stdlib.h>     /* srand, rand */
+#include <unistd.h>
+#include <mutex>
+
+#include "SegmentationNN.h"
+
+bool gRun;
+bool SAVE_RESULT = true;
+
+void sig_handler(int signo) {
+    std::cout<<"request gateway stop\n";
+    gRun = false;
+}
+
+void writePred(const std::string& images_names, const std::string& gt_folder, const std::string& out_folder, tk::dnn::SegmentationNN& segNN, int& width, int& height, bool show=false){
+    std::ifstream all_gt(images_names);
+    std::string filename;
+    cv::Mat frame;
+    for (; std::getline(all_gt, filename); ) {
+        std::cout<<filename<<std::endl;
+        frame = cv::imread(gt_folder + filename);
+        height = frame.rows;
+        width = frame.cols;
+        segNN.updateOriginal(frame, false);
+        if(show)
+            segNN.draw();
+        cv::imwrite(out_folder + filename, segNN.segmented[0]);
+    }
+}
+
+int main(int argc, char *argv[]) {
+
+    std::cout<<"detection\n";
+    signal(SIGINT, sig_handler);
+
+
+    std::string net = "shelfnet_fp32.rt";
+    if(argc > 1)
+        net = argv[1]; 
+    std::string input = "../demo/yolo_test.mp4";
+    if(argc > 2)
+        input = argv[2]; 
+    int n_batch = 1;
+    if(argc > 3)
+        n_batch = atoi(argv[3]); 
+    int n_classes = 19;
+    if(argc > 4)
+        n_classes = atoi(argv[4]); 
+    bool resize = false;
+    if(argc > 5)
+        resize = atoi(argv[5]); 
+    int baseline_resize = 1024;
+    if(argc > 6)
+        baseline_resize = atoi(argv[6]); 
+    bool show = true;
+    if(argc > 7)
+        show = atoi(argv[7]); 
+    bool write_pred = false;
+    if(argc > 8)
+        write_pred = atoi(argv[8]); 
+
+    if(resize && (baseline_resize < 0 || baseline_resize > 5000)) 
+        FatalError("Problem with baseline resize")
+    if(n_batch < 1 || n_batch > 64)
+        FatalError("Batch dim not supported");
+
+    //net initialization 
+    tk::dnn::SegmentationNN segNN;
+    segNN.init(net, n_classes, n_batch);
+
+    int height = 0, width = 0;
+    int basewidth=baseline_resize,  hsize; 
+    
+    if(write_pred){
+        std::string gt_folder = "../demo/CityScapes_val/images/";
+        std::string images_names = "../demo/CityScapes_val/all_images.txt";
+        std::string out_folder = "seg/";
+
+        writePred(images_names, gt_folder, out_folder, segNN, width, height, show);
+    }
+    else{
+        if(!show)
+            SAVE_RESULT = true;
+
+        gRun = true;
+
+        cv::VideoCapture cap(input);
+        if(!cap.isOpened())
+            gRun = false; 
+        else
+            std::cout<<"camera started\n";
+
+        cv::VideoWriter resultVideo;
+        if(SAVE_RESULT) {
+            int w,h;
+            if(resize){
+                w = basewidth;
+                h = int((float(cap.get(cv::CAP_PROP_FRAME_HEIGHT))*float(basewidth/float(cap.get(cv::CAP_PROP_FRAME_WIDTH)))));
+            }
+            else{
+                w = cap.get(cv::CAP_PROP_FRAME_WIDTH);
+                h = cap.get(cv::CAP_PROP_FRAME_HEIGHT);
+            }
+            resultVideo.open("result.mp4", cv::VideoWriter::fourcc('M','P','4','V'), 30, cv::Size(w, h));
+        }
+
+        cv::Mat frame;
+        while(gRun) {
+            cap >> frame; 
+            if(!frame.data) 
+                break;
+
+            if(resize){
+                hsize  = int((float(frame.rows)*float(basewidth/float(frame.cols))));
+                cv::resize(frame, frame, cv::Size(basewidth, hsize));
+            }
+
+            height = frame.rows;
+            width = frame.cols;
+
+            //inference
+            segNN.updateOriginal(frame, true);
+            if(show)
+                segNN.draw();
+
+            if(SAVE_RESULT)
+                resultVideo << segNN.segmented[0];
+        }
+    }
+
+    std::cout<<"segmentation end\n";   
+    double mean = 0, mean_pre = 0, mean_post = 0;
+    
+    std::cout<<COL_GREENB<<"\n\nTime stats for size ["<<width<<","<<height<<"] :\n";
+    
+    for(int i=0; i<segNN.stats.size(); i++) mean += segNN.stats[i]; mean /= segNN.stats.size();
+    for(int i=0; i<segNN.stats_pre.size(); i++) mean_pre += segNN.stats_pre[i]; mean_pre /= segNN.stats_pre.size();
+    for(int i=0; i<segNN.stats_post.size(); i++) mean_post += segNN.stats_post[i]; mean_post /= segNN.stats_post.size();
+    std::cout<<"Avg pre:\t"<<mean_pre<<" ms\t"<<1000/(mean_pre)<<" FPS\n";   
+    std::cout<<"Avg inf:\t"<<mean<<" ms\t"<<1000/(mean)<<" FPS\n";   
+    std::cout<<"Avg post:\t"<<mean_post<<" ms\t"<<1000/(mean_post)<<" FPS\n\n";   
+    std::cout<<"Avg tot:\t"<<(mean_pre + mean_post + mean) <<" ms\t"<<1000/((mean_pre + mean_post + mean))<<" FPS\n"<<COL_END;   
+    
+    return 0;
+}
+
@@ -0,0 +1,89 @@
+# Semantic Segmentation with tkDNN
+
+Currently tkDNN supports only ShelfNet as semantic segmentation network.
+
+## Export weights from Shelfnet
+To get the weights needed to run Shelfnet tests use [this](https://git.hipert.unimore.it/mverucchi/shelfnet) fork of a Pytorch implementation of Shelfnet network. 
+
+```
+git clone https://git.hipert.unimore.it/mverucchi/shelfnet
+cd shelfnet 
+cd ShelfNet18_realtime
+conda env create --file shelfnet_env.yml
+conda activate shelfnet
+mkdir layer debug
+python export.py
+```
+
+
+## Run the demo
+
+To run the semantic segmentation demo follow these steps (example with shelfnet):
+```
+rm shelfnet_fp32.rt        # be sure to delete(or move) old tensorRT files
+export TKDNN_BATCHSIZE=4   # be sure you have batch size > than 1 if you want to run inference on images bigger than 1024
+./test_shelfnet            # run the yolo test (is slow)
+./demo shelfnet_fp32.rt ../demo/yolo_test.mp4 1 19
+```
+In general the demo program takes the following parameters:
+```
+./seg_demo <network-rt-file> <path-to-video> <n-batches> <number-of-classes> <resize-flag> <baseline-resize> <show-flag> <write-pred>
+```
+where
+*  ```<network-rt-file>``` is the rt file generated by a test
+*  ```<<path-to-video>``` is the path to a video file or a camera input  
+*  ```<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).
+*  ```<number-of-classes>```is the number of classes the network is trained on
+*  ```<resize-flag>``` if set to 0 the demo will not resize the input frames, but use it as it is, otherwise it will resize it.
+*  ```<baseline-resize>``` is ```<resize-flag>``` is set to 1, then the input frames will be proportionally resized using ```<baseline-resize>``` as width baseline.
+*  ```<show-flag>``` if set to 0 the demo will not show the visualization but save the video into result.mp4 (if n-batches ==1)
+*  ```<write-pred>``` if set to 0 (default) the demo will run, otherwise the evaluation of a dataset will run and the output of the segmentation will be saved. Attention: this is under development and paths are embedded, so change them in the code in advance.
+
+NB) By default it is used FP32 inference
+NB) The batching is not used to work on more streams, rather to work on more tiles of the same image. Shelfnet never resized the input image, therefore for images greater than 1024x1024 tiles of 1024x1024 are given in input to the network in batch. 
+
+![gif](output.gif "Results on yolo_test.mp4")  
+
+For other demo videos refer to [this playlist](https://www.youtube.com/playlist?list=PLv0nEQYDD45y5EdSiywwCGPBmJVUzIWwe).
+
+
+## Existing tests and supported networks
+
+| Test Name         | Network                                       | Dataset                                                       | N Classes | Input size    | Weights                                                                   |
+| :---------------- | :-------------------------------------------- | :-----------------------------------------------------------: | :-------: | :-----------: | :------------------------------------------------------------------------ |
+| shelfnet              | ShelfNet18_realtime<sup>1</sup>                           | [Cityscapes](https://www.cityscapes-dataset.com/)                          | 19        | 1024x1024       | [weights](https://cloud.hipert.unimore.it/s/mEDZMRJaGCFWSJF/download)                                                                   |
+| shelfnet_berkeley              | ShelfNet18_realtime<sup>1</sup>                           | [DeepDrive](https://bdd-data.berkeley.edu/)                          | 20        | 1024x1024       | [weights](https://cloud.hipert.unimore.it/s/m92e7QdD9gYMF7f/download)                                                                   |
+
+1. Zhuang, Juntang, et al. "ShelfNet for fast semantic segmentation." Proceedings of the IEEE International Conference on Computer Vision Workshops. 2019.
+
+
+## FPS Results
+
+Inference FPS of shelfnet with tkDNN, average of 1200 images 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 );
+
+| Platform   | Test                     | Phase   | FP32, ms  | FP32, FPS | FP16, ms  |	FP16, FPS  | INT8, ms |	INT8, FPS | 
+| :------:   | :-----:                  | :-----: | :-----:   | :-----:   | :-----:   |	:-----:    | :-----:  |	:-----:   | 
+| RTX 2080Ti | shelfnet 1024x1024 (B=1) | pre     | 6.11863   |  163.435  |   5.81465 |  171.979   |  5.88699 |   169.866 |
+| RTX 2080Ti | shelfnet 1024x1024 (B=1) | inf     | 11.5464   |  86.6074  |   7.35396 |  135.981   |  6.37623 |   156.832 |
+| RTX 2080Ti | shelfnet 1024x1024 (B=1) | post    | 4.09058   |  244.464  |   3.91961 |  255.128   |  4.07343 |   245.493 |
+| RTX 2080Ti | shelfnet 1024x1024 (B=1) | tot     | 21.7556   |  45.9652  |   17.0882 |  58.5199   |  16.3366 |   61.2121 |
+| RTX 2080Ti | shelfnet 2048x2048 (B=4) | pre     | 25.435    |  39.3158  |   25.2953 |  39.5331   |  25.9303 |   38.565  | 
+| RTX 2080Ti | shelfnet 2048x2048 (B=4) | inf     | 36.5015   |  27.3961  |   17.0534 |  58.6395   |  15.6061 |   64.0773 |  
+| RTX 2080Ti | shelfnet 2048x2048 (B=4) | post    | 17.3917   |  57.4985  |   17.1649 |  58.2583   |  17.5539 |   56.9675 |  
+| RTX 2080Ti | shelfnet 2048x2048 (B=4) | tot     | 79.3283   |  12.6058  |   59.5136 |  16.8029   |  59.0903 |   16.9233 |  
+| AGX Xavier | shelfnet 1024x1024 (B=1) | pre     | 8.0174    |  124.729  |   7.5117  |  133.126   |  7.47333 |   133.809 |
+| AGX Xavier | shelfnet 1024x1024 (B=1) | inf     | 72.4173   |  13.8089  |   37.505  |  26.6631   |  31.3286 |   31.9197 |
+| AGX Xavier | shelfnet 1024x1024 (B=1) | post    | 8.89958   |  112.365  |   8.83576 |  113.176   |  9.42655 |   106.083 |
+| AGX Xavier | shelfnet 1024x1024 (B=1) | tot     | 89.3342   |  11.1939  |   53.8525 |  18.5692   |  48.2285 |   20.7346 |
+| AGX Xavier | shelfnet 2048x2048 (B=4) | pre     | 47.1454   |  21.211   |   21.6475 |  46.1947   |  21.4201 |   46.6851 | 
+| AGX Xavier | shelfnet 2048x2048 (B=4) | inf     | 266.537   |  3.75183  |   128.321 |  7.79293   |  107.621 |   9.29185 |  
+| AGX Xavier | shelfnet 2048x2048 (B=4) | post    | 44.0711   |  22.6906  |   40.1732 |  24.8922   |  39.873  |   25.0796 |  
+| AGX Xavier | shelfnet 2048x2048 (B=4) | tot     | 357.753   |  2.79522  |   190.142 |  5.25922   |  168.914 |   5.92016 |  
+
+
+## Known issues
+
+When creating the rt file all the checks returns errors. It is due to a different resize function and handling of the original ShelfNet outputs.
+However, the network is supposed to work. 
@@ -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";

@@ -4,7 +4,10 @@
 #include <iostream>
 #include <signal.h>
 #include <stdlib.h>    
+#ifdef __linux__
 #include <unistd.h>
+#endif 
+
 #include <mutex>
 #include "utils.h"

@@ -14,7 +17,7 @@

 #include "tkdnn.h"

-// #define OPENCV_CUDACONTRIB //if OPENCV has been compiled with CUDA and contrib.
+//#define OPENCV_CUDACONTRIB //if OPENCV has been compiled with CUDA and contrib.

 #ifdef OPENCV_CUDACONTRIB
 #include <opencv2/cudawarping.hpp>
@@ -150,7 +153,6 @@ class DetectionNN {
            int x0, w, x1, y0, h, y1;
            int objClass;
            std::string det_class;
-
            int baseline = 0;
            float font_scale = 0.5;
            int thickness = 2;   
@@ -1,7 +1,14 @@
 #include <iostream>
 #include <signal.h>
 #include <stdlib.h>     /* srand, rand */
+
+#ifdef __linux__
 #include <unistd.h>
+#elif _WIN32
+#define _USE_MATH_DEFINES
+#include <math.h>
+#endif 
+
 #include <mutex>
 #include <Eigen/Dense>
 #include "utils.h"
@@ -11,8 +11,11 @@
 #include <fstream>
 #include <iomanip>
 #include <signal.h>
-#include <stdlib.h>    
+#include <stdlib.h>
+#ifdef __linux__    
 #include <unistd.h>
+#endif
+
 #include <mutex>

 #include "NvInfer.h"
@@ -19,8 +19,10 @@ enum layerType_t {
    LAYER_ACTIVATION_CRELU,
    LAYER_ACTIVATION_LEAKY,
    LAYER_ACTIVATION_MISH,
+    LAYER_ACTIVATION_LOGISTIC,
    LAYER_FLATTEN,
    LAYER_RESHAPE,
+    LAYER_RESIZE,
    LAYER_MULADD,
    LAYER_POOLING,
    LAYER_SOFTMAX,
@@ -72,8 +74,10 @@ public:
            case LAYER_ACTIVATION_CRELU:    return "ActivationCReLU";
            case LAYER_ACTIVATION_LEAKY:    return "ActivationLeaky";
            case LAYER_ACTIVATION_MISH:     return "ActivationMish";
+            case LAYER_ACTIVATION_LOGISTIC: return "ActivationLogistic";
            case LAYER_FLATTEN:             return "Flatten";
            case LAYER_RESHAPE:             return "Reshape";
+            case LAYER_RESIZE:              return "Resize";
            case LAYER_MULADD:              return "MulAdd";
            case LAYER_POOLING:             return "Pooling";
            case LAYER_SOFTMAX:             return "Softmax";
@@ -216,7 +220,8 @@ public:
 typedef enum {
    ACTIVATION_ELU     = 100,
    ACTIVATION_LEAKY   = 101,
-    ACTIVATION_MISH   = 102
+    ACTIVATION_MISH   = 102,
+    ACTIVATION_LOGISTIC   = 103
 } tkdnnActivationMode_t;

 /**
@@ -227,8 +232,9 @@ class Activation : public Layer {
 public:
    int act_mode;
    float ceiling;
+    float slope;

-    Activation(Network *net, int act_mode, const float ceiling=0.0); 
+    Activation(Network *net, int act_mode, const float ceiling=0.0, const float slope=0.1); 
    virtual ~Activation();
    virtual layerType_t getLayerType() { 
        if(act_mode == CUDNN_ACTIVATION_CLIPPED_RELU)
@@ -237,6 +243,8 @@ public:
            return LAYER_ACTIVATION_LEAKY;
        else if (act_mode == ACTIVATION_MISH)
            return LAYER_ACTIVATION_MISH;
+        else if (act_mode == ACTIVATION_LOGISTIC)
+            return LAYER_ACTIVATION_LOGISTIC;
        else
            return LAYER_ACTIVATION;
         };
@@ -431,6 +439,23 @@ public:

 };

+enum ResizeMode_t { NEAREST= 0, 
+                    LINEAR= 1};
+
+/**
+    Resize layer
+*/
+class Resize : public Layer {
+
+public:
+    Resize(Network *net, int scale_c, int scale_h, int scale_w, bool fixed=false, ResizeMode_t mode=NEAREST);
+    virtual ~Resize();
+    virtual layerType_t getLayerType() { return LAYER_RESIZE; };
+
+    virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
+
+    ResizeMode_t mode;
+};

 /**
    MulAdd layer
@@ -551,7 +576,7 @@ public:
 class Shortcut : public Layer {

 public:
-    Shortcut(Network *net, Layer *backLayer); 
+    Shortcut(Network *net, Layer *backLayer, bool mul=false); 
    virtual ~Shortcut();
    virtual layerType_t getLayerType() { return LAYER_SHORTCUT; };

@@ -559,6 +584,7 @@ public:

 public:
    Layer *backLayer;
+    bool mul = false;
 };

 /**
@@ -614,24 +640,28 @@ 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);
 };

 /**
@@ -6,6 +6,7 @@
 #include "Network.h"
 #include "Layer.h"
 #include "NvInfer.h"
+#include <memory>

 namespace tk { namespace dnn {

@@ -24,6 +25,7 @@ template<typename T> T readBUF(const char*& buffer)

 using namespace nvinfer1;
 #include "pluginsRT/ActivationLeakyRT.h"
+#include "pluginsRT/ActivationLogisticRT.h"
 #include "pluginsRT/ActivationReLUCeilingRT.h"
 #include "pluginsRT/ActivationMishRT.h"
 #include "pluginsRT/ReorgRT.h"
@@ -59,6 +61,7 @@ public:
 #if NV_TENSORRT_MAJOR >= 6  
    nvinfer1::IBuilderConfig *configRT;
 #endif
+    
    nvinfer1::ICudaEngine *engineRT;
    nvinfer1::IExecutionContext *contextRT;

@@ -105,6 +108,7 @@ public:
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Route *l);
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Flatten *l);
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Reshape *l);
+    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Resize *l);
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Reorg *l);
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Region *l);
    nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Shortcut *l);
@@ -114,6 +118,9 @@ public:

    bool serialize(const char *filename);
    bool deserialize(const char *filename);
+
+
+
 };

 }}
@@ -5,8 +5,8 @@

 namespace tk { namespace dnn {

-cv::Mat vizFloat2colorMap(cv::Mat map);
-cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim);
+cv::Mat vizFloat2colorMap(cv::Mat map, double min=0, double max=0, int classes=19);
+cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int img_h, int img_w, double min=0, double max=0, int classes=19);
 cv::Mat vizLayer2Mat(tk::dnn::Network *net, int layer, int imgdim = 1000);
    
 }}
@@ -0,0 +1,403 @@
+#ifndef SEGMENTATIONNN_H
+#define SEGMENTATIONNN_H
+
+#include <iostream>
+#include <signal.h>
+#include <stdlib.h>    
+#include <unistd.h>
+#include <mutex>
+#include "utils.h"
+
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+#include <opencv2/core/hal/interface.h>
+
+#include "tkdnn.h"
+#include "NetworkViz.h"
+#include "kernelsThrust.h"
+
+namespace tk { namespace dnn {
+
+class SegmentationNN {
+
+    protected:
+        tk::dnn::NetworkRT *netRT = nullptr;
+        int nBatches = 1;
+
+        std::vector<cv::Size> originalSize;
+        cv::Mat bgr[3];
+        dnnType *input;
+        dnnType *input_d;
+        float* confidences_h;
+
+        float * tmpInputData_d;
+        float *tmpOutData_d;
+        float *tmpOutData_h;
+
+        float *mean_d, *stddev_d;
+
+        cublasHandle_t cublasHandle;
+
+        void computeBorders(const int or_width, const int or_height, int& top, int& bottom, int& left, int&right){
+            top = 0;
+            bottom = 0;
+            left = 0;
+            right = 0;
+
+            if(or_height != or_width){
+                if(or_height < or_width){
+                    top = (or_width - or_height)/2;
+                    bottom = or_width - top - or_height;
+                }
+                else{
+                    left = (or_height - or_width)/2;
+                    right = or_height - left - or_width;
+                }
+            }
+        }
+
+        /**
+         * This method preprocess the image, before feeding it to the NN.
+         *
+         * @param frame original frame to adapt for inference.
+         * @param bi batch index
+         */
+        void preprocess(cv::Mat &frame, const int bi=0) {
+            originalSize[bi] = frame.size();
+
+            frame.convertTo(frame, CV_32FC3, 1 / 255.0, 0);
+            int H = frame.rows;
+            int W = frame.cols;
+            cv::Mat frame_cropped;
+
+            int top, bottom, left, right;
+            computeBorders(W, H, top, bottom, left, right);
+            cv::copyMakeBorder(frame, frame_cropped, top, bottom, left, right, cv::BORDER_CONSTANT, cv::Scalar(0,0,0) );
+
+            tk::dnn::dataDim_t idim = netRT->input_dim;
+
+            resize(frame_cropped, frame_cropped, cv::Size(idim.w, idim.h));
+            
+            cv::split(frame_cropped, bgr);
+            for (int i = 0; i < idim.c; i++){
+                int idx = i * frame_cropped.rows * frame_cropped.cols;
+                int ch = idim.c-1 -i;
+                memcpy((void *)&input[idx + idim.tot()*bi], (void *)bgr[ch].data, frame_cropped.rows * frame_cropped.cols * sizeof(dnnType));
+            }
+
+            checkCuda(cudaMemcpyAsync(input_d+ idim.tot()*bi, input + idim.tot()*bi, idim.tot() * sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
+
+            normalize(input_d + idim.tot()*bi, idim.c, idim.h, idim.w, mean_d, stddev_d);
+        }        
+
+        /**
+         * This method postprocess the output of the NN to obtain the correct 
+         * boundig boxes. 
+         * 
+         * @param bi batch index
+         */
+        void postprocess(const int bi=0, bool appy_colormap = true) {
+            dnnType *rt_out = (dnnType *)netRT->buffersRT[1]+ netRT->buffersDIM[1].tot()*bi;
+
+            dataDim_t odim = netRT->output_dim;
+            
+            matrixTranspose(cublasHandle, rt_out, tmpInputData_d, odim.c, odim.w*odim.h);
+            maxElem(tmpInputData_d, tmpOutData_d, odim.c, odim.h, odim.w);
+            checkCuda(cudaMemcpy(tmpOutData_h, tmpOutData_d, odim.w*odim.h * sizeof(float), cudaMemcpyDeviceToHost));
+
+            dataDim_t vdim = odim;
+            vdim.c = 1;
+
+            cv::Mat colored;
+
+            if(appy_colormap)
+                colored = vizData2Mat(tmpOutData_h, vdim, netRT->input_dim.h, netRT->input_dim.w, 0, classes, classes);
+            else{
+                cv::Mat colored_fp32 (cv::Size(odim.w, odim.h),CV_32FC1, tmpOutData_h);
+                colored_fp32.convertTo(colored, CV_8UC1);
+            }
+            
+            int max_dim = (originalSize[bi].width > originalSize[bi].height) ? originalSize[bi].width : originalSize[bi].height;
+            resize(colored, colored, cv::Size(max_dim, max_dim));
+            int top, bottom, left, right;
+            computeBorders(originalSize[bi].width, originalSize[bi].height, top, bottom, left, right);
+            cv::Rect roi(left,top,originalSize[bi].width, originalSize[bi].height);
+            cv::Mat or_size (colored, roi);
+            segmented[bi] = or_size;
+        };
+
+    public:
+        int classes = 0;
+        std::vector<double> stats; /*keeps track of inference times (ms)*/
+        std::vector<double> stats_pre; 
+        std::vector<double> stats_post; 
+        std::vector<std::string> classesNames;
+        std::vector<cv::Mat> segmented;
+
+        SegmentationNN() {
+            checkERROR( cublasCreate(&cublasHandle) );
+        };
+        ~SegmentationNN(){
+            checkERROR( cublasDestroy(cublasHandle) );
+        };
+
+        /**
+         * Method used to inialize the class, allocate memory and compute 
+         * needed data.
+         * 
+         * @param tensor_path path to the rt file og 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.
+         */
+        bool init(const std::string& tensor_path, const int n_classes=19, const int n_batches=1){
+            std::cout<<(tensor_path).c_str()<<"\n";
+            if(!fileExist(tensor_path.c_str()))
+                FatalError("This file do not exists" + tensor_path );
+
+            netRT = new tk::dnn::NetworkRT(NULL, (tensor_path).c_str());
+            classes = n_classes;
+            nBatches = n_batches;
+
+            checkCuda(cudaMallocHost(&input, sizeof(dnnType) * netRT->input_dim.tot() * nBatches));
+            checkCuda(cudaMalloc(&input_d, sizeof(dnnType) * netRT->input_dim.tot() * nBatches));
+
+            dataDim_t odim = netRT->output_dim;
+
+            checkCuda(cudaMallocHost(&confidences_h, sizeof(float) * odim.tot()));
+            checkCuda(cudaMalloc(&tmpInputData_d, sizeof(float) * odim.tot()));
+            checkCuda(cudaMalloc(&tmpOutData_d, sizeof(float) * odim.w*odim.h));
+            checkCuda(cudaMallocHost(&tmpOutData_h, sizeof(float) * odim.w*odim.h));
+
+            segmented.resize(nBatches);
+            originalSize.resize(nBatches);
+
+            std::vector<float> mean = {0.485, 0.456, 0.406};
+            std::vector<float> stddev = {0.229, 0.224, 0.225};
+
+            checkCuda(cudaMalloc(&mean_d, sizeof(float) * mean.size()));
+            checkCuda(cudaMalloc(&stddev_d, sizeof(float) * stddev.size()));
+
+            checkCuda(cudaMemcpyAsync(mean_d, mean.data(), mean.size() * sizeof(float), cudaMemcpyHostToDevice, netRT->stream));
+            checkCuda(cudaMemcpyAsync(stddev_d, stddev.data(), stddev.size() * sizeof(float), cudaMemcpyHostToDevice, netRT->stream));
+
+        }
+        
+        /**
+         * This method performs the whole detection of the NN.
+         * 
+         * @param frames frames to run detection on.
+         * @param cur_batches number of batches to use in inference
+         * @param save_times if set to true, preprocess, inference and postprocess times 
+         *        are saved on a csv file, otherwise not.
+         * @param times pointer to the output stream where to write times
+         * @param mAP set to true only if all the probabilities for a bounding 
+         *            box are needed, as in some cases for the mAP calculation
+         */
+        void update(std::vector<cv::Mat>& frames, const int cur_batches=1, bool apply_colormap=true){
+            if(cur_batches > nBatches)
+                FatalError("A batch size greater than nBatches cannot be used");
+
+            originalSize.clear();
+            if(TKDNN_VERBOSE) printCenteredTitle(" TENSORRT detection ", '=', 30); 
+            {
+                TKDNN_TSTART
+                for(int bi=0; bi<cur_batches;++bi){
+                    if(!frames[bi].data)
+                        FatalError("No image data feed to detection");
+                    originalSize.push_back(frames[bi].size());
+                    preprocess(frames[bi], bi);    
+                }
+                TKDNN_TSTOP
+                stats_pre.push_back(t_ns);
+            }
+
+            //do inference
+            tk::dnn::dataDim_t dim = netRT->input_dim;
+            dim.n = cur_batches;
+            {
+                if(TKDNN_VERBOSE) dim.print();
+                TKDNN_TSTART
+                netRT->infer(dim, input_d);
+                TKDNN_TSTOP
+                if(TKDNN_VERBOSE) dim.print();
+                stats.push_back(t_ns);
+            }
+
+            {
+                TKDNN_TSTART
+                for(int bi=0; bi<cur_batches;++bi)
+                    postprocess(bi, apply_colormap);
+                TKDNN_TSTOP
+                stats_post.push_back(t_ns);
+            }
+        }     
+
+        void updateOriginal(cv::Mat frame, bool apply_colormap=true){
+
+            std::vector<cv::Mat> splitted_frames;
+            int H, W, net_H, net_W;
+            int top = 0, bottom = 0, left = 0, right = 0;
+            std::vector<std::pair<int,int>> pos;
+
+            {
+                TKDNN_TSTART
+                cv::Size original_size =  frame.size();
+
+                frame.convertTo(frame, CV_32FC3, 1 / 255.0, 0);
+                H = frame.rows;
+                W = frame.cols;
+                net_H = netRT->input_dim.h; 
+                net_W = netRT->input_dim.w;
+
+                cv::Mat frame_cropped;
+                
+                if( H <= net_H && W <= net_W ){ // smaller size wrt network
+                    top = (net_H - H)/2;
+                    bottom = net_H - H - top ;
+                    left = (net_W - W)/2;
+                    right = net_W - W - left ;
+                    cv::copyMakeBorder(frame, frame_cropped, top, bottom, left, right, cv::BORDER_CONSTANT, cv::Scalar(0,0,0) );
+                    splitted_frames.push_back(frame_cropped);
+                }
+                else{ //bigger size wrt network
+
+                    
+                    if(H < net_H  ||  W < net_W){
+                        if(H < net_H){
+                            top = (net_H - H)/2;
+                            bottom = net_H - H - top ;
+                        }
+                        else{
+                            left = (net_W - W)/2;
+                            right = net_W - W - left ;
+                        }
+                        cv::copyMakeBorder(frame, frame_cropped, top, bottom, left, right, cv::BORDER_CONSTANT, cv::Scalar(0,0,0));
+                    }
+
+                    for(int x=0; x+net_W<=W ;){
+                        for(int y=0; y+net_H <=H ; ){
+                            cv::Rect roi(x, y, net_W, net_H);
+                            cv::Mat image_roi = frame(roi);
+                            splitted_frames.push_back(image_roi);
+                            pos.push_back(std::make_pair(x,y));
+                            
+                            y += net_H;
+                            if(y == H)
+                                break;
+                            if(y + net_H > H) y = H - net_H;
+                        }
+                        x += net_W;
+                        if(x == W)
+                            break;
+                        if(x + net_W > W) x = W - net_W;
+                    }
+                }
+
+                tk::dnn::dataDim_t idim = netRT->input_dim;
+
+                if(splitted_frames.size()> nBatches)
+                    FatalError(std::to_string(splitted_frames.size()) + " min batches required");
+
+                for(int bi=0; bi<splitted_frames.size();++bi){
+                    cv::split(splitted_frames[bi], bgr);
+                    for (int i = 0; i < idim.c; i++){
+                        int idx = i * splitted_frames[bi].rows * splitted_frames[bi].cols;
+                        int ch = idim.c-1 -i;
+                        memcpy((void *)&input[idx + idim.tot()*bi], (void *)bgr[ch].data, splitted_frames[bi].rows * splitted_frames[bi].cols * sizeof(dnnType));
+                    }
+
+                    checkCuda(cudaMemcpyAsync(input_d+ idim.tot()*bi, input + idim.tot()*bi, idim.tot() * sizeof(dnnType), cudaMemcpyHostToDevice, netRT->stream));
+                    normalize(input_d + idim.tot()*bi, idim.c, idim.h, idim.w, mean_d, stddev_d);
+                }
+                TKDNN_TSTOP
+                stats_pre.push_back(t_ns);
+            }
+
+            tk::dnn::dataDim_t dim = netRT->input_dim;
+            dim.n = splitted_frames.size();
+            {
+                if(TKDNN_VERBOSE) dim.print();
+                TKDNN_TSTART
+                netRT->infer(dim, input_d);
+                TKDNN_TSTOP
+                if(TKDNN_VERBOSE) dim.print();
+                stats.push_back(t_ns);
+            }
+
+            dataDim_t odim = netRT->output_dim;
+
+            std::vector<cv::Mat> out_img;
+
+            {
+                TKDNN_TSTART
+
+                for(int bi=0; bi<splitted_frames.size();++bi){
+
+                    dnnType *rt_out = (dnnType *)netRT->buffersRT[1]+ netRT->buffersDIM[1].tot()*bi;               
+                    
+                    matrixTranspose(cublasHandle, rt_out, tmpInputData_d, odim.c, odim.w*odim.h);
+                    maxElem(tmpInputData_d, tmpOutData_d, odim.c, odim.h, odim.w);
+                    checkCuda(cudaMemcpy(tmpOutData_h, tmpOutData_d, odim.w*odim.h * sizeof(float), cudaMemcpyDeviceToHost));
+
+                    dataDim_t vdim = odim;
+                    vdim.c = 1;
+
+                    cv::Mat colored;
+
+                    if(apply_colormap)
+                        colored = vizData2Mat(tmpOutData_h, vdim, netRT->input_dim.h, netRT->input_dim.w, 0, classes, classes);
+                    else{
+                        cv::Mat colored_fp32 (cv::Size(odim.w, odim.h),CV_32FC1, tmpOutData_h);
+                        colored_fp32.convertTo(colored, CV_8UC1);
+                    }
+                    out_img.push_back(colored);
+                }
+
+
+                cv::Mat seg(frame.size(), out_img[0].type());
+                if(out_img.size() == 1)
+                {
+                    cv::Rect roi(left, top, W, H);
+                    seg = out_img[0](roi);
+                }
+                else{
+                    int bi=0;
+
+                    if(top == 0 && left == 0){
+
+                        for(int i=0; i<out_img.size(); ++i){
+                            cv::Mat roi_collage = seg(cv::Rect( pos[i].first ,pos[i].second,out_img[i].cols,out_img[i].rows));
+                            out_img[i].copyTo(roi_collage);
+                        }
+                    }
+                    else{
+                        FatalError("Not handled case")
+                    }
+                }
+                segmented[0] = seg;
+                
+                TKDNN_TSTOP
+                stats_post.push_back(t_ns);
+            }
+        } 
+
+        /**
+         * Method to draw boundixg boxes and labels on a frame.
+         */
+        cv::Mat draw(const int cur_batches=1) {
+            for(int i=0; i<cur_batches; ++i){
+
+                cv::imshow("segmented", segmented[i]);
+                cv::resizeWindow("segmented", cv::Size(512,288));
+                cv::waitKey(1);
+            }
+            return segmented[0];
+        }
+
+};
+
+}}
+
+#endif /* SEGMENTATIONNN_H*/
@@ -4,7 +4,7 @@
 #include "utils.h"

 void activationELUForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0));
-void activationLEAKYForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0));
+void activationLEAKYForward(dnnType *srcData, dnnType *dstData, int size, float slope, cudaStream_t stream = cudaStream_t(0));
 void activationReLUCeilingForward(dnnType *srcData, dnnType *dstData, int size, const float ceiling, cudaStream_t stream = cudaStream_t(0));
 void activationLOGISTICForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0));
 void activationSIGMOIDForward(dnnType *srcData, dnnType *dstData, int size, cudaStream_t stream = cudaStream_t(0));
@@ -24,7 +24,7 @@ void softmaxForward(float *input, int n, int batch, int batch_offset,
                    int groups, int group_offset, int stride, float temp, float *output, cudaStream_t stream = cudaStream_t(0));

 void shortcutForward(dnnType *srcData, dnnType *dstData, int n1, int c1, int h1, int w1, int s1,
-                     int n2, int c2, int h2, int w2, int s2,
+                     int n2, int c2, int h2, int w2, int s2, bool mul,
                     cudaStream_t stream = cudaStream_t(0));

 void upsampleForward(dnnType *srcData, dnnType *dstData,
@@ -2,6 +2,7 @@
 #define KERNELSTHRUST_H


+#include <thrust/extrema.h>
 #include <thrust/sort.h>
 #include <thrust/execution_policy.h>
 #include <thrust/functional.h>
@@ -9,6 +10,8 @@
 #include <thrust/iterator/constant_iterator.h>
 #include <thrust/gather.h>
 #include <thrust/copy.h>
+#include <thrust/device_ptr.h>
+

 #include "tkdnn.h"

@@ -36,4 +39,6 @@ void topKxyAddOffset(int * ids_begin, const int K, const int size, int *intxs_be
 void bboxes(int * ids_begin, const int K, const int size, float *xs_begin, float *ys_begin, 
            dnnType *src_begin, float *bbx0, float *bbx1, float *bby0, float *bby1, float *src_out, int *ids_out);

+void maxElem(dnnType *src_begin, dnnType *dst_begin, const int c, const int h, const int w);
+
 #endif //KERNELSTHRUST_H
@@ -4,9 +4,8 @@
 class ActivationLeakyRT : public IPlugin {

 public:
-	ActivationLeakyRT() {
-
-
+	ActivationLeakyRT(float s) {
+		slope = s;
 	}

 	~ActivationLeakyRT(){
@@ -42,19 +41,21 @@ public:
 	virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override {

 		activationLEAKYForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]), 
-											reinterpret_cast<dnnType*>(outputs[0]), batchSize*size, stream);
+											reinterpret_cast<dnnType*>(outputs[0]), batchSize*size, slope, stream);
 		return 0;
 	}


 	virtual size_t getSerializationSize() override {
-		return 1*sizeof(int);
+		return 1*sizeof(int) + 1*sizeof(float);
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, size);
+		assert(buf == a + getSerializationSize());
 	}

 	int size;
+	float slope;
 };
@@ -0,0 +1,60 @@
+#include<cassert>
+#include "../kernels.h"
+
+class ActivationLogisticRT : public IPlugin {
+
+public:
+    ActivationLogisticRT() {
+
+
+    }
+
+    ~ActivationLogisticRT(){
+
+    }
+
+    int getNbOutputs() const override {
+        return 1;
+    }
+
+    Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override {
+        return inputs[0];
+    }
+
+    void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override {
+        size = 1;
+        for(int i=0; i<outputDims[0].nbDims; i++)
+            size *= outputDims[0].d[i];
+    }
+
+    int initialize() override {
+
+        return 0;
+    }
+
+    virtual void terminate() override {
+    }
+
+    virtual size_t getWorkspaceSize(int maxBatchSize) const override {
+        return 0;
+    }
+
+    virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override {
+
+        activationLOGISTICForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]),
+                                  reinterpret_cast<dnnType*>(outputs[0]), batchSize*size, stream);
+        return 0;
+    }
+
+
+    virtual size_t getSerializationSize() override {
+        return 1*sizeof(int);
+    }
+
+    virtual void serialize(void* buffer) override {
+        char *buf = reinterpret_cast<char*>(buffer);
+        tk::dnn::writeBUF(buf, size);
+    }
+
+    int size;
+};
@@ -52,8 +52,9 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, size);
+		assert(buf == a + getSerializationSize());
 	}

 	int size;
@@ -51,9 +51,10 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, ceiling);
 		tk::dnn::writeBUF(buf, size);
+		assert(buf = a + getSerializationSize());
 		
 	}

@@ -52,8 +52,9 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, size);
+		assert(buf == a + getSerializationSize());
 	}

 	int size;
@@ -116,7 +116,7 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, chunk_dim);
 		tk::dnn::writeBUF(buf, kh);
 		tk::dnn::writeBUF(buf, kw);
@@ -163,6 +163,7 @@ public:
        for(int i=0; i<dim_ones; i++)
    		tk::dnn::writeBUF(buf, aus[i]);
 		free(aus);
+		assert(buf == a + getSerializationSize());
 	}

 	cublasStatus_t stat; 
@@ -65,12 +65,13 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a = buf;
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
 		tk::dnn::writeBUF(buf, rows);
 		tk::dnn::writeBUF(buf, cols);
+		assert(buf == a + getSerializationSize());
 	}

 	int c, h, w;
@@ -55,7 +55,7 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;

 		tk::dnn::writeBUF(buf, this->c);
 		tk::dnn::writeBUF(buf, this->h);
@@ -65,6 +65,7 @@ public:
 		tk::dnn::writeBUF(buf, this->stride_W);
 		tk::dnn::writeBUF(buf, this->winSize);
 		tk::dnn::writeBUF(buf, this->padding);
+		assert(buf == a + getSerializationSize());
 	}

 	int n, c, h, w;
@@ -73,13 +73,14 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, classes);
 		tk::dnn::writeBUF(buf, coords);
 		tk::dnn::writeBUF(buf, num);
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 	}

 	int c, h, w;
@@ -52,11 +52,12 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, stride);
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 	}

 	int c, h, w, stride;
@@ -50,11 +50,12 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a = buf;
 		tk::dnn::writeBUF(buf, n);
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 	}

 	int n, c, h, w;
@@ -52,7 +52,7 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;

 		tk::dnn::writeBUF(buf, o_c);
 		tk::dnn::writeBUF(buf, o_h);
@@ -61,6 +61,7 @@ public:
 		tk::dnn::writeBUF(buf, i_c);
 		tk::dnn::writeBUF(buf, i_h);
 		tk::dnn::writeBUF(buf, i_w);
+		assert(buf == a + getSerializationSize());
 	}

 	int i_c, i_h, i_w, o_c, o_h, o_w;
@@ -75,7 +75,7 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, groups);
 		tk::dnn::writeBUF(buf, group_id);
 		tk::dnn::writeBUF(buf, in);
@@ -85,6 +85,7 @@ public:
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 	}

 	static const int MAX_INPUTS = 4;
@@ -4,10 +4,11 @@
 class ShortcutRT : public IPlugin {

 public:
-	ShortcutRT(tk::dnn::dataDim_t bdim) {
+	ShortcutRT(tk::dnn::dataDim_t bdim, bool mul) {
 		this->bc = bdim.c;
 		this->bh = bdim.h;
 		this->bw = bdim.w;
+		this->mul = mul;
 	}

 	~ShortcutRT(){
@@ -47,28 +48,30 @@ public:
 		dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);

 		checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
-		for(int b=0; b < batchSize; ++b)
-			shortcutForward(srcDataBack + b*bc*bh*bw, dstData + b*c*h*w, 1, c, h, w, 1, 1, bc, bh, bw, 1, stream);
+		shortcutForward(srcDataBack, dstData, batchSize, c, h, w, 1, batchSize, bc, bh, bw, 1, mul, stream);

 		return 0;
 	}


 	virtual size_t getSerializationSize() override {
-		return 6*sizeof(int);
+		return 6*sizeof(int) + sizeof(bool);
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, bc);
 		tk::dnn::writeBUF(buf, bh);
 		tk::dnn::writeBUF(buf, bw);
+		tk::dnn::writeBUF(buf, mul);
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 		
 	}

 	int c, h, w;
 	int bc, bh, bw;
+	bool mul;
 };
@@ -54,11 +54,12 @@ public:
 	}

 	virtual void serialize(void* buffer) override {
-		char *buf = reinterpret_cast<char*>(buffer);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
 		tk::dnn::writeBUF(buf, stride);
 		tk::dnn::writeBUF(buf, c);
 		tk::dnn::writeBUF(buf, h);
 		tk::dnn::writeBUF(buf, w);
+		assert(buf == a + getSerializationSize());
 	}

 	int c, h, w, stride;
@@ -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];
@@ -61,17 +64,23 @@ public:

 		checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));

-		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 (this->scaleXY != 1) scalAdd(dstData + index, 2 * w*h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
-				
-				index = entry_index(b, n*w*h, 4);
-				activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*w*h, stream);
-			}
-		}
+        for (int b = 0; b < batchSize; ++b){
+            for(int n = 0; n < n_masks; ++n){
+                int index = entry_index(b, n*w*h, 0);
+                if (new_coords == 1){
+                    if (this->scaleXY != 1) scalAdd(dstData + index, 2 * w*h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
+                }
+                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);
+
+                    index = entry_index(b, n*w*h, 4);
+                    activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*w*h, stream);
+                }
+            }
+        }

 		//std::cout<<"YOLO END\n";
 		return 0;
@@ -79,22 +88,29 @@ 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 {
-		char *buf = reinterpret_cast<char*>(buffer);
-		tk::dnn::writeBUF(buf, classes);
-		tk::dnn::writeBUF(buf, num);
-		tk::dnn::writeBUF(buf, n_masks);
-		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++)
-    		tk::dnn::writeBUF(buf, bias[i]);
+		char *buf = reinterpret_cast<char*>(buffer),*a=buf;
+		tk::dnn::writeBUF(buf, classes); 	//std::cout << "Classes :" << classes << std::endl;
+		tk::dnn::writeBUF(buf, num); 		//std::cout << "Num : " << num << std::endl;
+		tk::dnn::writeBUF(buf, n_masks); 	//std::cout << "N_Masks" << n_masks << std::endl;
+		tk::dnn::writeBUF(buf, scaleXY); 	//std::cout << "ScaleXY :" << scaleXY << std::endl;
+		tk::dnn::writeBUF(buf, nms_thresh); //std::cout << "nms_thresh :" << nms_thresh << std::endl;
+		tk::dnn::writeBUF(buf, nms_kind); 	//std::cout << "nms_kind : " << nms_kind << std::endl;
+		tk::dnn::writeBUF(buf, new_coords); //std::cout << "new_coords : " << new_coords << std::endl;
+		tk::dnn::writeBUF(buf, c); 			//std::cout << "C : " << c << std::endl;
+		tk::dnn::writeBUF(buf, h); 			//std::cout << "H : " << h << std::endl;
+		tk::dnn::writeBUF(buf, w); 			//std::cout << "C : " << c << std::endl;
+		for (int i = 0; i < n_masks; i++)
+		{
+			tk::dnn::writeBUF(buf, mask[i]); //std::cout << "mask[i] : " << mask[i] << std::endl;
+		}
+		for (int i = 0; i < n_masks * 2 * num; i++)
+		{
+			tk::dnn::writeBUF(buf, bias[i]); //std::cout << "bias[i] : " << bias[i] << std::endl;
+		}

 		// save classes names
 		for(int i=0; i<classes; i++) {
@@ -104,11 +120,15 @@ public:
 				tk::dnn::writeBUF(buf, tmp[j]);
 			}
 		}
+		assert(buf == a + getSerializationSize());
 	}

 	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;
@@ -29,7 +29,8 @@ int testInference(std::vector<std::string> input_bins, std::vector<std::string>
    readBinaryFile(input_bins[0], net->input_dim.tot(), &input_h, &data);

    // outputs
-    dnnType *cudnn_out[outputs.size()], *rt_out[outputs.size()]; 
+    //dnnType *cudnn_out[outputs.size()], *rt_out[outputs.size()];
+    std::vector<dnnType *> cudnn_out,rt_out;

    tk::dnn::dataDim_t dim1 =  net->input_dim; //input dim
    printCenteredTitle(" CUDNN inference ", '=', 30); {
@@ -39,7 +40,7 @@ int testInference(std::vector<std::string> input_bins, std::vector<std::string>
        TKDNN_TSTOP
        dim1.print();   
    }
-    for(int i=0; i<outputs.size(); i++) cudnn_out[i] = outputs[i]->dstData;
+    for(int i=0; i<outputs.size(); i++) cudnn_out.push_back(outputs[i]->dstData);

    if(netRT != nullptr) {
        tk::dnn::dataDim_t dim2 = net->input_dim;
@@ -50,7 +51,7 @@ int testInference(std::vector<std::string> input_bins, std::vector<std::string>
            TKDNN_TSTOP
            dim2.print();
        }
-        for(int i=0; i<outputs.size(); i++) rt_out[i] = (dnnType*)netRT->buffersRT[i+1];
+        for(int i=0; i<outputs.size(); i++) rt_out.push_back((dnnType*)netRT->buffersRT[i+1]);
    }

    int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0; 
@@ -12,8 +12,12 @@
 #include <cublas_v2.h>
 #include <cudnn.h>

+#ifdef __linux__
 #include <unistd.h>
+#endif
+
 #include <ios>
+#include <chrono>


 #define dnnType float
@@ -39,6 +43,7 @@
 #define TKDNN_VERBOSE 0

 // Simple Timer 
+#ifdef __linux__
 #define TKDNN_TSTART timespec start, end;                               \
                    clock_gettime(CLOCK_MONOTONIC, &start);            

@@ -48,6 +53,14 @@
    if(show) std::cout<<col<<"Time:"<<std::setw(16)<<t_ns<<" ms\n"<<COL_END; 

 #define TKDNN_TSTOP TKDNN_TSTOP_C(COL_CYANB, TKDNN_VERBOSE)
+#elif _WIN32
+#define TKDNN_TSTART auto start = std::chrono::high_resolution_clock::now();
+#define TKDNN_TSTOP auto stop = std::chrono::high_resolution_clock::now(); \
+std::chrono::duration<double> duration = stop -start;                      \
+auto time_ms = std::chrono::duration_cast<std::chrono::milliseconds>(duration);\
+double t_ns = time_ms.count();
+#endif
+

 /********************************************************
 * Prints the error message, and exits
@@ -0,0 +1,39 @@
+import os
+import urllib.request as dowReq
+import zipfile
+
+val = input("Enter BDD or COCO :")
+if(val == "COCO"):
+    url = "https://cloud.hipert.unimore.it/s/LNxBDk4wzqXPL8c/download"
+    lib = "..\demo\COCO_val2017"
+    lib_zip = "COCO_val2017.zip"
+elif(val == "BDD"):
+    url = "https://cloud.hipert.unimore.it/s/bikqk3FzCq2tg4D/download"
+    lib = "..\demo\BDD100k_val"
+    lib_zip = "BDD100k_val.zip"
+
+dowReq.urlretrieve(url,lib_zip)
+
+with zipfile.ZipFile(lib_zip,'r') as zip_ref:
+    zip_ref.extractall(lib)
+
+labelFolder = lib + "\labels"
+imageFolder = lib + "\images"
+
+file1 = open(".\\..\\demo\\all_labels.txt","a")
+path1 = os.path.realpath(labelFolder)
+for file in os.listdir(labelFolder):
+    valTemp = path1 + "\\" + file
+    valTemp = valTemp + '\n'
+    file1.write(valTemp)
+file1.close()
+
+file2 = open(".\\..\\demo\\all_images.txt","a")
+path2 = os.path.realpath(imageFolder)
+for file in os.listdir(imageFolder):
+    pathtemp = path2 + "\\" + file
+    pathtemp = pathtemp + '\n'
+    file2.write(pathtemp)
+file2.close()
+
+print("Completed")
@@ -69,12 +69,21 @@ do
    echo -e "${ORANGE}Batch $TKDNN_BATCHSIZE ${NC}"
    
    test_net mnist    
-    ./test_imuodom &>> $out_file
-    print_output $? imuodom
+    # ./test_imuodom &>> $out_file
+    # print_output $? imuodom

+    test_net shelfnet
+    test_net shelfnet_berkeley
    test_net yolo4
+    test_net yolo4_320
+    test_net yolo4_320_coco2
+    test_net yolo4_512
+    test_net yolo4_608
+    test_net yolo4-csp
+    test_net yolo4x
    test_net yolo4_berkeley
    test_net yolo4tiny
+    test_net yolo4tiny_512
    test_net yolo3
    test_net yolo3_berkeley
    test_net yolo3_coco4
@@ -42,7 +42,7 @@ do
    echo -e "${ORANGE}Batch $TKDNN_BATCHSIZE ${NC}"
    
    test_inference yolo4_320 $mode
-    test_inference yolo4_416 $mode
+    test_inference yolo4 $mode
    test_inference yolo4_512 $mode
    test_inference yolo4_608 $mode
    test_inference yolo4tiny $mode
@@ -5,11 +5,12 @@

 namespace tk { namespace dnn {

-Activation::Activation(Network *net, int act_mode, const float ceiling) : 
+Activation::Activation(Network *net, int act_mode, const float ceiling, const float slope) : 
    Layer(net) {

-    this->act_mode = act_mode;
-    this->ceiling = ceiling;
+    this->act_mode  = act_mode;
+    this->ceiling   = ceiling;
+    this->slope     = slope;
    checkCuda( cudaMalloc(&dstData, input_dim.tot()*sizeof(dnnType)) );

    if(int(act_mode) < 100) {
@@ -46,12 +47,15 @@ Activation::~Activation() {

 dnnType* Activation::infer(dataDim_t &dim, dnnType* srcData) {
    if(act_mode == ACTIVATION_LEAKY) {
-        activationLEAKYForward(srcData, dstData, dim.tot());
-
+        activationLEAKYForward(srcData, dstData, dim.tot(), this->slope);
    }
    else if(act_mode == ACTIVATION_MISH) {
        activationMishForward(srcData, dstData, dim.tot());

+    }
+    else if(act_mode == ACTIVATION_LOGISTIC) {
+        activationLOGISTICForward(srcData, dstData, dim.tot());
+
    } else {
        dnnType alpha = dnnType(1);
        dnnType beta  = dnnType(0);
@@ -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;
@@ -177,6 +187,7 @@ namespace tk { namespace dnn {
            if(f.activation == "relu") act = tkdnnActivationMode_t(CUDNN_ACTIVATION_RELU);
            else if(f.activation == "leaky") act = tk::dnn::ACTIVATION_LEAKY;
            else if(f.activation == "mish") act = tk::dnn::ACTIVATION_MISH;
+            else if(f.activation == "logistic") act = tk::dnn::ACTIVATION_LOGISTIC;
            else { FatalError("activation not supported: " + f.activation); }
            netLayers[netLayers.size()-1] = new tk::dnn::Activation(net, act);
        };
@@ -87,17 +87,22 @@ LSTM::LSTM( Network *net, int hiddensize, bool returnSeq, std::string fname_weig
    checkCUDNN(cudnnCreateRNNDescriptor(&rnnDesc));

 #if CUDNN_MAJOR > 7
-    checkCUDNN(cudnnSetRNNDescriptor_v6(net->cudnnHandle,
+    checkCUDNN(cudnnSetRNNDescriptor_v6(net->cudnnHandle,rnnDesc, stateSize, numLayers, dropoutDesc,
+                                        cudnnRNNInputMode_t::CUDNN_LINEAR_INPUT,
+            //(bidirectional ? cudnnDirectionMode_t::CUDNN_BIDIRECTIONAL : cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL),
+                                        cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL,
+                                        cudnnRNNMode_t::CUDNN_LSTM,
+                                        cudnnRNNAlgo_t::CUDNN_RNN_ALGO_STANDARD,
+                                        net->dataType));
 #else
-    checkCUDNN(cudnnSetRNNDescriptor(net->cudnnHandle,
-#endif
-        rnnDesc, stateSize, numLayers, dropoutDesc,
+    checkCUDNN(cudnnSetRNNDescriptor(net->cudnnHandle,rnnDesc, stateSize, numLayers, dropoutDesc,
        cudnnRNNInputMode_t::CUDNN_LINEAR_INPUT,
        //(bidirectional ? cudnnDirectionMode_t::CUDNN_BIDIRECTIONAL : cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL),
        cudnnDirectionMode_t::CUDNN_UNIDIRECTIONAL,
        cudnnRNNMode_t::CUDNN_LSTM,
        cudnnRNNAlgo_t::CUDNN_RNN_ALGO_STANDARD,
        net->dataType));
+#endif


    // Get temp space sizes
@@ -32,6 +32,7 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,

    this->batchnorm = batchnorm;
    if(batchnorm) {
+        
        readBinaryFile(weights_path.c_str(), outputs, &scales_h, &scales_d, seek);
        seek += outputs;
        readBinaryFile(weights_path.c_str(), outputs, &mean_h, &mean_d, seek);
@@ -140,6 +140,7 @@ NetworkRT::NetworkRT(Network *net, const char *name) {
        engineRT = builderRT->buildEngineWithConfig(*networkRT, *configRT);
 #else 
        engineRT = builderRT->buildCudaEngine(*networkRT);
+        //engineRT = std::shared_ptr<nvinfer1::ICudaEngine>(builderRT->buildCudaEngine(*networkRT));
 #endif
        if(engineRT == nullptr)
            FatalError("cloud not build cuda engine")
@@ -226,7 +227,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) {
        return convert_layer(input, (Conv2d*) l);
    if(type == LAYER_POOLING)
        return convert_layer(input, (Pooling*) l);
-    if(type == LAYER_ACTIVATION || type == LAYER_ACTIVATION_CRELU || type == LAYER_ACTIVATION_LEAKY || type == LAYER_ACTIVATION_MISH)
+    if(type == LAYER_ACTIVATION || type == LAYER_ACTIVATION_CRELU || type == LAYER_ACTIVATION_LEAKY || type == LAYER_ACTIVATION_MISH || type == LAYER_ACTIVATION_LOGISTIC)
        return convert_layer(input, (Activation*) l);
    if(type == LAYER_SOFTMAX)
        return convert_layer(input, (Softmax*) l);
@@ -236,6 +237,8 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) {
        return convert_layer(input, (Flatten*) l);
    if(type == LAYER_RESHAPE)
        return convert_layer(input, (Reshape*) l);
+    if(type == LAYER_RESIZE)
+        return convert_layer(input, (Resize*) l);
    if(type == LAYER_REORG)
        return convert_layer(input, (Reorg*) l);
    if(type == LAYER_REGION)
@@ -389,13 +392,13 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) {
        
 #if NV_TENSORRT_MAJOR < 6                
        // plugin version
-        IPlugin *plugin = new ActivationLeakyRT();
+        IPlugin *plugin = new ActivationLeakyRT(l->slope);
        IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
        checkNULL(lRT);
        return lRT;
 #else 
        IActivationLayer *lRT = networkRT->addActivation(*input, ActivationType::kLEAKY_RELU);
-        lRT->setAlpha(0.1);
+        lRT->setAlpha(l->slope);
        checkNULL(lRT);
        return lRT;
 #endif
@@ -421,6 +424,12 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Activation *l) {
        checkNULL(lRT);
        return lRT;
    }
+    else if(l->act_mode == ACTIVATION_LOGISTIC) {
+        IPlugin *plugin = new ActivationLogisticRT();
+        IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
+        checkNULL(lRT);
+        return lRT;
+    }
    else {
        FatalError("this Activation mode is not yet implemented");
        return NULL;
@@ -472,13 +481,23 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Flatten *l) {
 ILayer* NetworkRT::convert_layer(ITensor *input, Reshape *l) {
    // std::cout<<"convert Reshape\n";

-    l->output_dim.print();
    IPlugin *plugin = new ReshapeRT(l->output_dim);
    IPluginLayer *lRT = networkRT->addPlugin(&input, 1, *plugin);
    checkNULL(lRT);
    return lRT;
 }

+ILayer* NetworkRT::convert_layer(ITensor *input, Resize *l) {
+    // std::cout<<"convert Resize\n";
+
+    IResizeLayer *lRT = networkRT->addResize(*input); //default is kNEAREST
+    checkNULL(lRT);
+    Dims d{};
+    lRT->setResizeMode(ResizeMode(l->mode));
+    lRT->setOutputDimensions(DimsCHW{l->output_dim.c, l->output_dim.h, l->output_dim.w});
+    return lRT;
+}
+
 ILayer* NetworkRT::convert_layer(ITensor *input, Reorg *l) {
    //std::cout<<"convert Reorg\n";

@@ -506,7 +525,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Shortcut *l) {
    
    ITensor *back_tens = tensors[l->backLayer];

-    if(l->backLayer->output_dim.c == l->output_dim.c)
+    if(l->backLayer->output_dim.c == l->output_dim.c && !l->mul) 
    {
        IElementWiseLayer *lRT = networkRT->addElementWise(*input, *back_tens, ElementWiseOperation::kSUM);
        checkNULL(lRT);
@@ -515,7 +534,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Shortcut *l) {
    else
    {
        // plugin version
-        IPlugin *plugin = new ShortcutRT(l->backLayer->output_dim);
+        IPlugin *plugin = new ShortcutRT(l->backLayer->output_dim, l->mul);
        ITensor **inputs = new ITensor*[2];
        inputs[0] = input;
        inputs[1] = back_tens; 
@@ -529,7 +548,7 @@ 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;
@@ -561,7 +580,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, DeformConv2d *l) {
    IPluginLayer *lRT = networkRT->addPlugin(inputs, 2, *plugin);
    checkNULL(lRT);
    lRT->setName( ("Deformable" + std::to_string(l->id)).c_str() );
-    delete(inputs);
+    delete[](inputs);
    // batchnorm
    void *bias_b, *power_b, *mean_b, *variance_b, *scales_b;
    if(dtRT == DataType::kHALF) {
@@ -638,43 +657,61 @@ bool NetworkRT::deserialize(const char *filename) {


 IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialData, size_t serialLength) {
-    const char * buf = reinterpret_cast<const char*>(serialData);
+    const char * buf = reinterpret_cast<const char*>(serialData),*bufCheck = buf;

    std::string name(layerName);
    //std::cout<<name<<std::endl;

    if(name.find("ActivationLeaky") == 0) {
-        ActivationLeakyRT *a = new ActivationLeakyRT();
+        ActivationLeakyRT *a = new ActivationLeakyRT(readBUF<float>(buf));
        a->size = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return a;
    }
    if(name.find("ActivationMish") == 0) {
        ActivationMishRT *a = new ActivationMishRT();
        a->size = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
+        return a;
+    }
+    if(name.find("ActivationLogistic") == 0) {
+        ActivationLogisticRT *a = new ActivationLogisticRT();
+        a->size = readBUF<int>(buf);
+        return a;
+    }
+    if(name.find("ActivationLogistic") == 0) {
+        ActivationLogisticRT *a = new ActivationLogisticRT();
+        a->size = readBUF<int>(buf);
        return a;
    }
    if(name.find("ActivationCReLU") == 0) {
-        ActivationReLUCeiling *a = new ActivationReLUCeiling(readBUF<float>(buf));
+        float activationReluTemp = readBUF<float>(buf);
+        ActivationReLUCeiling* a = new ActivationReLUCeiling(activationReluTemp);
        a->size = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return a;
    }

    if(name.find("Region") == 0) {
-        RegionRT *r = new RegionRT(readBUF<int>(buf),    //classes
-                                    readBUF<int>(buf),    //coords
-                                    readBUF<int>(buf));   //num
+        int classesTemp = readBUF<int>(buf);
+        int coordsTemp = readBUF<int>(buf);
+        int numTemp = readBUF<int>(buf);
+        RegionRT* r = new RegionRT(classesTemp, coordsTemp, numTemp);

        r->c = readBUF<int>(buf);
        r->h = readBUF<int>(buf);
        r->w = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

    if(name.find("Reorg") == 0) {
-        ReorgRT *r = new ReorgRT(readBUF<int>(buf)); //stride
+        int strideTemp = readBUF<int>(buf);
+        ReorgRT *r = new ReorgRT(strideTemp);
        r->c = readBUF<int>(buf);
        r->h = readBUF<int>(buf);
        r->w = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

@@ -685,32 +722,39 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
        bdim.w = readBUF<int>(buf);
        bdim.l = 1;

-        ShortcutRT *r = new ShortcutRT(bdim);
+        ShortcutRT *r = new ShortcutRT(bdim, readBUF<bool>(buf));
        r->c = readBUF<int>(buf);
        r->h = readBUF<int>(buf);
        r->w = readBUF<int>(buf);
        return r;
+        assert(buf == bufCheck + serialLength);
    } 

    if(name.find("Pooling") == 0) {
-        MaxPoolFixedSizeRT *r = new MaxPoolFixedSizeRT( readBUF<int>(buf), //c
-                                                        readBUF<int>(buf), //h
-                                                        readBUF<int>(buf), //w
-                                                        readBUF<int>(buf), //n
-                                                        readBUF<int>(buf), //strideH
-                                                        readBUF<int>(buf), //strideW
-                                                        readBUF<int>(buf), //winSize
-                                                        readBUF<int>(buf)); //padding
+        int cTemp = readBUF<int>(buf);
+        int hTemp = readBUF<int>(buf);
+        int wTemp = readBUF<int>(buf);
+        int nTemp = readBUF<int>(buf);
+        int strideHTemp = readBUF<int>(buf);
+        int strideWTemp = readBUF<int>(buf);
+        int winSizeTemp = readBUF<int>(buf);
+        int paddingTemp = readBUF<int>(buf);
+
+        MaxPoolFixedSizeRT* r = new MaxPoolFixedSizeRT(cTemp, hTemp, wTemp, nTemp, strideHTemp, strideWTemp, winSizeTemp, paddingTemp);
+        assert(buf == bufCheck + serialLength);
        return r;
    }

    if(name.find("Resize") == 0) {
-        ResizeLayerRT *r = new ResizeLayerRT(readBUF<int>(buf), //o_c
-                                            readBUF<int>(buf), //o_h
-                                            readBUF<int>(buf)); //o_w
+        int o_cTemp = readBUF<int>(buf);
+        int o_hTemp = readBUF<int>(buf);
+        int o_wTemp = readBUF<int>(buf);
+        ResizeLayerRT* r = new ResizeLayerRT(o_cTemp, o_hTemp, o_wTemp);
+
        r->i_c = readBUF<int>(buf);
        r->i_h = readBUF<int>(buf);
        r->i_w = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

@@ -721,6 +765,7 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
        r->w = readBUF<int>(buf);
        r->rows = readBUF<int>(buf);
        r->cols = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

@@ -732,19 +777,28 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
        new_dim.h = readBUF<int>(buf);
        new_dim.w = readBUF<int>(buf);
        ReshapeRT *r = new ReshapeRT(new_dim); 
+        assert(buf == bufCheck + serialLength);
        
        return r;
    } 

    if(name.find("Yolo") == 0) {
-        YoloRT *r = new YoloRT(readBUF<int>(buf),    //classes
-                                readBUF<int>(buf),   //num
-                                nullptr,
-                                readBUF<int>(buf));   //n_masks
+
+        int classes_temp = readBUF<int>(buf);
+        int num_temp = readBUF<int>(buf);
+        int n_masks_temp = readBUF<int>(buf);
+        float scale_xy_temp = readBUF<float>(buf);
+        float nms_thresh_temp = readBUF<float>(buf);
+        int nms_kind_temp = readBUF<int>(buf);
+        int new_coords_temp = readBUF<int>(buf);
+
+       YoloRT *r = new YoloRT(classes_temp,num_temp,nullptr,n_masks_temp,scale_xy_temp,nms_thresh_temp,nms_kind_temp,new_coords_temp);  
+
+
+
        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++)
@@ -758,36 +812,54 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
 				tmp[j] = readBUF<char>(buf);
            r->classesNames[i] = std::string(tmp);
 		}
+        assert(buf == bufCheck + serialLength);

        yolos[n_yolos++] = r;
        return r;
    } 
    if(name.find("Upsample") == 0) {
-        UpsampleRT *r = new UpsampleRT(readBUF<int>(buf)); //stride
+        int strideTemp = readBUF<int>(buf);
+        UpsampleRT* r = new UpsampleRT(strideTemp);
        r->c = readBUF<int>(buf);
        r->h = readBUF<int>(buf);
        r->w = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

    if(name.find("Route") == 0) {
-        RouteRT *r = new RouteRT(readBUF<int>(buf),readBUF<int>(buf));
+        int groupsTemp = readBUF<int>(buf);
+        int group_idTemp = readBUF<int>(buf);
+        RouteRT* r = new RouteRT(groupsTemp, group_idTemp);
        r->in = readBUF<int>(buf);
        for(int i=0; i<RouteRT::MAX_INPUTS; i++)
            r->c_in[i] = readBUF<int>(buf);
        r->c = readBUF<int>(buf);
        r->h = readBUF<int>(buf);
        r->w = readBUF<int>(buf);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

    if(name.find("Deformable") == 0) {
-        DeformableConvRT *r = new DeformableConvRT(readBUF<int>(buf), readBUF<int>(buf), readBUF<int>(buf),
-                                                    readBUF<int>(buf), readBUF<int>(buf), readBUF<int>(buf),
-                                                    readBUF<int>(buf), readBUF<int>(buf),
-                                                    readBUF<int>(buf),readBUF<int>(buf),readBUF<int>(buf),readBUF<int>(buf),
-                                                    readBUF<int>(buf),readBUF<int>(buf),readBUF<int>(buf),readBUF<int>(buf),
-                                                    nullptr); 
+        int chuck_dimTemp = readBUF<int>(buf);
+        int khTemp = readBUF<int>(buf);
+        int kwTemp = readBUF<int>(buf);
+        int shTemp = readBUF<int>(buf);
+        int swTemp = readBUF<int>(buf);
+        int phTemp = readBUF<int>(buf);
+        int pwTemp = readBUF<int>(buf);
+        int deformableGroupTemp = readBUF<int>(buf);
+        int i_nTemp = readBUF<int>(buf);
+        int i_cTemp = readBUF<int>(buf);
+        int i_hTemp = readBUF<int>(buf);
+        int i_wTemp = readBUF<int>(buf);
+        int o_nTemp = readBUF<int>(buf);
+        int o_cTemp = readBUF<int>(buf);
+        int o_hTemp = readBUF<int>(buf);
+        int o_wTemp = readBUF<int>(buf);
+
+        DeformableConvRT* r = new DeformableConvRT(chuck_dimTemp, khTemp, kwTemp, shTemp, swTemp, phTemp, pwTemp, deformableGroupTemp, i_nTemp, i_cTemp, i_hTemp, i_wTemp, o_nTemp, o_cTemp, o_hTemp, o_wTemp, nullptr);
        dnnType *aus = new dnnType[r->chunk_dim*2];
        for(int i=0; i<r->chunk_dim*2; i++)
    		aus[i] = readBUF<dnnType>(buf);
@@ -818,6 +890,7 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
    		aus[i] = readBUF<dnnType>(buf);
 		checkCuda( cudaMemcpy(r->ones_d2, aus, sizeof(dnnType)*r->dim_ones, cudaMemcpyHostToDevice) );
        free(aus);
+        assert(buf == bufCheck + serialLength);
        return r;
    } 

@@ -6,23 +6,389 @@

 namespace tk { namespace dnn {

-cv::Mat vizFloat2colorMap(cv::Mat map) {
+cv::Mat mapillary_15_map(cv::Mat adjMap){
+
+    // cv::imshow("test", adjMap);
+    // cv::waitKey(0);
+    cv::Mat M1(1, 256, CV_8UC1), M2(1, 256, CV_8UC1), M3(1, 256, CV_8UC1);
+
+    //animal
+    M3.at<uchar>(0)=165;
+    M2.at<uchar>(0)=42;
+    M1.at<uchar>(0)=45;
+
+    //curb
+    M3.at<uchar>(1)=196;
+    M2.at<uchar>(1)=196;
+    M1.at<uchar>(1)=196;
+
+    //barrier
+    M3.at<uchar>(2)=90;
+    M2.at<uchar>(2)=120;
+    M1.at<uchar>(2)=150;
+
+    //road
+    M3.at<uchar>(3)=128;
+    M2.at<uchar>(3)=64;
+    M1.at<uchar>(3)=128;
+
+    //building
+    M3.at<uchar>(4)=70;
+    M2.at<uchar>(4)=70;
+    M1.at<uchar>(4)=70;
+
+    //person
+    M3.at<uchar>(5)=220;
+    M2.at<uchar>(5)=20;
+    M1.at<uchar>(5)=60;
+
+    //roadmark
+    M3.at<uchar>(6)=255;
+    M2.at<uchar>(6)=255;
+    M1.at<uchar>(6)=255;
+
+    //nature
+    M3.at<uchar>(7)=107;
+    M2.at<uchar>(7)=142;
+    M1.at<uchar>(7)=35;
+
+    //sky
+    M3.at<uchar>(8)=70;
+    M2.at<uchar>(8)=130;
+    M1.at<uchar>(8)=180;
+
+    //billboard
+    M3.at<uchar>(9)=220;
+    M2.at<uchar>(9)=220;
+    M1.at<uchar>(9)=220;
+
+    //pole
+    M3.at<uchar>(10)=153;
+    M2.at<uchar>(10)=153;
+    M1.at<uchar>(10)=153;
+
+    //traffic sign
+    M3.at<uchar>(11)=128;
+    M2.at<uchar>(11)=128;
+    M1.at<uchar>(11)=128;
+
+    //bike
+    M3.at<uchar>(12)=119;
+    M2.at<uchar>(12)=11;
+    M1.at<uchar>(12)=32;
+
+    //vehicle
+    M3.at<uchar>(13)=0;
+    M2.at<uchar>(13)=0;
+    M1.at<uchar>(13)=142;
+
+    //void
+    for(int i=14;i<256;i++)
+    {
+        M1.at<uchar>(i)=0;
+        M2.at<uchar>(i)=0;
+        M3.at<uchar>(i)=0;
+    }
+
+    cv::Mat r1,r2,r3;
+
+    cv::LUT(adjMap,M1,r1);
+    cv::LUT(adjMap,M2,r2);
+    cv::LUT(adjMap,M3,r3);
+
+    std::vector<cv::Mat> planes;
+    planes.push_back(r1);
+    planes.push_back(r2);
+    planes.push_back(r3);
+
+    cv::Mat dst;
+    cv::merge(planes,dst);
+    return dst;
+
+
+}
+
+cv::Mat berkeley_20_map(cv::Mat adjMap){
+
+    cv::Mat M1(1, 256, CV_8UC1), M2(1, 256, CV_8UC1), M3(1, 256, CV_8UC1);
+
+    //road
+    M3.at<uchar>(0)=128;
+    M2.at<uchar>(0)=64;
+    M1.at<uchar>(0)=128;
+
+    //sidewalk
+    M3.at<uchar>(1)=244;
+    M2.at<uchar>(1)=35;
+    M1.at<uchar>(1)=232;
+
+    //building
+    M3.at<uchar>(2)=70;
+    M2.at<uchar>(2)=70;
+    M1.at<uchar>(2)=70;
+
+    //wall
+    M3.at<uchar>(3)=102;
+    M2.at<uchar>(3)=102;
+    M1.at<uchar>(3)=156;
+
+    //fence
+    M3.at<uchar>(4)=90;
+    M2.at<uchar>(4)=120;
+    M1.at<uchar>(4)=150;
+
+    //pole
+    M3.at<uchar>(5)=153;
+    M2.at<uchar>(5)=153;
+    M1.at<uchar>(5)=153;
+
+    //traffic light
+    M3.at<uchar>(6)=250;
+    M2.at<uchar>(6)=170;
+    M1.at<uchar>(6)=30;
+
+    //traffic sign
+    M3.at<uchar>(7)=128;
+    M2.at<uchar>(7)=128;
+    M1.at<uchar>(7)=128;
+
+    //nature
+    M3.at<uchar>(8)=107;
+    M2.at<uchar>(8)=142;
+    M1.at<uchar>(8)=35;
+
+    //ground
+    M3.at<uchar>(9)=0;
+    M2.at<uchar>(9)=192;
+    M1.at<uchar>(9)=0;
+
+    //sky
+    M3.at<uchar>(10)=70;
+    M2.at<uchar>(10)=130;
+    M1.at<uchar>(10)=180;
+
+    //person
+    M3.at<uchar>(11)=220;
+    M2.at<uchar>(11)=20;
+    M1.at<uchar>(11)=60;
+
+    //rider
+    M3.at<uchar>(12)=255;
+    M2.at<uchar>(12)=0;
+    M1.at<uchar>(12)=100;
+
+    //car
+    M3.at<uchar>(13)=0;
+    M2.at<uchar>(13)=0;
+    M1.at<uchar>(13)=142;
+
+    //truck
+    M3.at<uchar>(14)=0;
+    M2.at<uchar>(14)=0;
+    M1.at<uchar>(14)=70;
+
+    //bus
+    M3.at<uchar>(15)=0;
+    M2.at<uchar>(15)=60;
+    M1.at<uchar>(15)=100;
+
+    //train
+    M3.at<uchar>(16)=0;
+    M2.at<uchar>(16)=0;
+    M1.at<uchar>(16)=192;
+
+    //motorbike
+    M3.at<uchar>(17)=0;
+    M2.at<uchar>(17)=0;
+    M1.at<uchar>(17)=230;
+
+    //bike
+    M3.at<uchar>(18)=119;
+    M2.at<uchar>(18)=11;
+    M1.at<uchar>(18)=32;
+
+    //void
+    for(int i=19;i<256;i++)
+    {
+        M1.at<uchar>(i)=0;
+        M2.at<uchar>(i)=0;
+        M3.at<uchar>(i)=0;
+    }
+
+    cv::Mat r1,r2,r3;
+
+    cv::LUT(adjMap,M1,r1);
+    cv::LUT(adjMap,M2,r2);
+    cv::LUT(adjMap,M3,r3);
+
+    std::vector<cv::Mat> planes;
+    planes.push_back(r1);
+    planes.push_back(r2);
+    planes.push_back(r3);
+
+    cv::Mat dst;
+    cv::merge(planes,dst);
+    return dst;
+
+}
+
+cv::Mat cityscapes_19_map(cv::Mat adjMap){
+
+    cv::Mat M1(1, 256, CV_8UC1), M2(1, 256, CV_8UC1), M3(1, 256, CV_8UC1);
+
+    //road
+    M3.at<uchar>(0)=128;
+    M2.at<uchar>(0)=64;
+    M1.at<uchar>(0)=128;
+
+    //sidewalk
+    M3.at<uchar>(1)=244;
+    M2.at<uchar>(1)=35;
+    M1.at<uchar>(1)=232;
+
+    //building
+    M3.at<uchar>(2)=70;
+    M2.at<uchar>(2)=70;
+    M1.at<uchar>(2)=70;
+
+    //wall
+    M3.at<uchar>(3)=102;
+    M2.at<uchar>(3)=102;
+    M1.at<uchar>(3)=156;
+
+    //fence
+    M3.at<uchar>(4)=190;
+    M2.at<uchar>(4)=153;
+    M1.at<uchar>(4)=153;
+
+    //pole
+    M3.at<uchar>(5)=153;
+    M2.at<uchar>(5)=153;
+    M1.at<uchar>(5)=153;
+
+    //traffic light
+    M3.at<uchar>(6)=250;
+    M2.at<uchar>(6)=170;
+    M1.at<uchar>(6)=30;
+
+    //traffic sign
+    M3.at<uchar>(7)=220;
+    M2.at<uchar>(7)=220;
+    M1.at<uchar>(7)=0;
+
+    //vegetation
+    M3.at<uchar>(8)=107;
+    M2.at<uchar>(8)=142;
+    M1.at<uchar>(8)=35;
+
+    //terrain
+    M3.at<uchar>(9)=152;
+    M2.at<uchar>(9)=251;
+    M1.at<uchar>(9)=152;
+
+    //sky
+    M3.at<uchar>(10)=70;
+    M2.at<uchar>(10)=130;
+    M1.at<uchar>(10)=180;
+
+    //person
+    M3.at<uchar>(11)=220;
+    M2.at<uchar>(11)=20;
+    M1.at<uchar>(11)=60;
+
+    //rider
+    M3.at<uchar>(12)=255;
+    M2.at<uchar>(12)=0;
+    M1.at<uchar>(12)=0;
+
+    //car
+    M3.at<uchar>(13)=0;
+    M2.at<uchar>(13)=0;
+    M1.at<uchar>(13)=142;
+
+    //truck
+    M3.at<uchar>(14)=0;
+    M2.at<uchar>(14)=0;
+    M1.at<uchar>(14)=70;
+
+    //bus
+    M3.at<uchar>(15)=0;
+    M2.at<uchar>(15)=60;
+    M1.at<uchar>(15)=100;
+
+    //train
+    M3.at<uchar>(16)=0;
+    M2.at<uchar>(16)=80;
+    M1.at<uchar>(16)=100;
+
+    //motorcycle
+    M3.at<uchar>(17)=0;
+    M2.at<uchar>(17)=0;
+    M1.at<uchar>(17)=230;
+
+    //bicycle
+    M3.at<uchar>(18)=119;
+    M2.at<uchar>(18)=11;
+    M1.at<uchar>(18)=32;
+
+    //void
+    for(int i=19;i<256;i++)
+    {
+        M1.at<uchar>(i)=0;
+        M2.at<uchar>(i)=0;
+        M3.at<uchar>(i)=0;
+    }
+
+    cv::Mat r1,r2,r3;
+
+    cv::LUT(adjMap,M1,r1);
+    cv::LUT(adjMap,M2,r2);
+    cv::LUT(adjMap,M3,r3);
+
+    std::vector<cv::Mat> planes;
+    planes.push_back(r1);
+    planes.push_back(r2);
+    planes.push_back(r3);
+
+    cv::Mat dst;
+    cv::merge(planes,dst);
+    return dst;
+
+}
+
+
+cv::Mat vizFloat2colorMap(cv::Mat map,double min, double max, int classes) {
+
+    if(min == 0 && max == 0)
+        cv::minMaxIdx(map, &min, &max);

-    double min;
-    double max;
-    cv::minMaxIdx(map, &min, &max);
    cv::Mat adjMap;
-    // expand your range to 0..255. Similar to histEq();
-    map.convertTo(adjMap,CV_8UC1, 255 / (max-min), -min); 
-    //return adjMap;
-
-    
    cv::Mat falseColorsMap;
-    applyColorMap(adjMap, falseColorsMap, cv::COLORMAP_HOT);
+    
+    switch (classes)
+    {
+    case 15:
+        map.convertTo(adjMap,CV_8UC1);
+        falseColorsMap = mapillary_15_map(adjMap);
+        break;
+    case 20:
+        map.convertTo(adjMap,CV_8UC1);
+        falseColorsMap = berkeley_20_map(adjMap);
+        break;
+    case 19:
+        map.convertTo(adjMap,CV_8UC1);
+        falseColorsMap = cityscapes_19_map(adjMap);
+        break;
+    
+    default:
+        // expand your range to 0..255. Similar to histEq();
+        map.convertTo(adjMap,CV_8UC1, 255 / (max-min), -min); 
+        applyColorMap(adjMap, falseColorsMap, cv::COLORMAP_JET);
+    }
    return falseColorsMap;
 }

-cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim) {
+cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int img_h, int img_w, double min, double max, int classes) {
    dnnType *data = nullptr;

    // copy to CPU
@@ -38,14 +404,13 @@ cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim) {
    cv::Mat grid = cv::Mat(gridSize, CV_8UC3, cv::Scalar(0));    

    for(int i=0; i<dim.c;i++) {
-        cv::Mat raw = vizFloat2colorMap(cv::Mat(cv::Size(dim.w, dim.h),CV_32FC1, data + dim.w*dim.h*i));
+        cv::Mat raw = vizFloat2colorMap(cv::Mat(cv::Size(dim.w, dim.h),CV_32FC1, data + dim.w*dim.h*i), min, max, classes);
        int r = i / gridDim;
        int c = i - r * gridDim;
        raw.copyTo(grid.rowRange(r*dim.h, r*dim.h + dim.h).colRange(c*dim.w, c*dim.w + dim.w));
    }

-    float ar = float(dim.w)/dim.h;
-    cv::Size vdim(ar*imgdim, imgdim);
+    cv::Size vdim(img_w, img_h);
    cv::Mat viz;
    cv::resize(grid, viz, vdim, 0, 0, 0);
    
@@ -59,7 +424,7 @@ cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim) {
 cv::Mat vizLayer2Mat(tk::dnn::Network *net, int layer, int imgdim) {
    if(layer >= net->num_layers)
        FatalError("Could not viz layer\n");
-    return vizData2Mat(net->layers[layer]->dstData, net->layers[layer]->output_dim, imgdim);
+    return vizData2Mat(net->layers[layer]->dstData, net->layers[layer]->output_dim, imgdim, imgdim);

    //cv::imwrite("viz/layer" + std::to_string(layer) + ".png", viz);
    //cv::imshow("layer", viz);
@@ -15,6 +15,11 @@ Reshape::Reshape(Network *net, dataDim_t new_dim) : Layer(net) {
    output_dim.w = new_dim.w;
    output_dim.l = new_dim.l;

+    output_dim = new_dim;
+
+    if(input_dim.tot() != output_dim.tot())
+        FatalError("Reshape dimension mismatch");
+
 }

 Reshape::~Reshape() {
@@ -0,0 +1,39 @@
+#include <iostream>
+
+#include "Layer.h"
+#include "kernels.h"
+
+namespace tk { namespace dnn {
+
+Resize::Resize(Network *net, int scale_c, int scale_h, int scale_w, bool fixed, ResizeMode_t mode) : Layer(net) {
+
+    this->mode = mode;
+    if(fixed){
+        output_dim.c = scale_c;
+        output_dim.h = scale_h;
+        output_dim.w = scale_w;
+    }
+    else{
+        output_dim.c *= scale_c;
+        output_dim.h *= scale_h;
+        output_dim.w *= scale_w;
+    }
+
+    checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) );
+}
+
+Resize::~Resize() {
+
+    checkCuda( cudaFree(dstData) );
+}
+
+dnnType* Resize::infer(dataDim_t &dim, dnnType* srcData) {
+
+    resizeForward(srcData, dstData, dim.n, dim.c, dim.h, dim.w,
+                   output_dim.c, output_dim.h, output_dim.w);
+    dim = output_dim;
+
+    return dstData;
+}
+
+}}
@@ -5,15 +5,16 @@

 namespace tk { namespace dnn {

-Shortcut::Shortcut(Network *net, Layer *backLayer) : Layer(net) {
+Shortcut::Shortcut(Network *net, Layer *backLayer, bool mul) : Layer(net) {

    this->backLayer = backLayer;
+    this->mul = mul;
    checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) );

-    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 mismatch");
+    if( ( backLayer->output_dim.c != input_dim.c && mul ) ||
+        (( backLayer->output_dim.w != input_dim.w ||  backLayer->output_dim.h != input_dim.h ) && !mul ) )  
+        FatalError("Shortcut dim missmatch");
+    
 }

 Shortcut::~Shortcut() {
@@ -26,7 +27,7 @@ dnnType* Shortcut::infer(dataDim_t &dim, dnnType* srcData) {
    dataDim_t bdim = this->backLayer->output_dim;

    checkCuda(cudaMemcpy(dstData, srcData, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
-    shortcutForward(this->backLayer->dstData, dstData, dim.n, dim.c, dim.h, dim.w, 1, bdim.n, bdim.c, bdim.h, bdim.w, 1);
+    shortcutForward(this->backLayer->dstData, dstData, dim.n, dim.c, dim.h, dim.w, 1, bdim.n, bdim.c, bdim.h, bdim.w, 1, mul);

    //update data dimensions    
    dim = output_dim;
@@ -9,9 +9,10 @@
 #include "Layer.h"
 #include "kernels.h"

+
 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 +20,9 @@ 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 +63,21 @@ 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] ) / lw;
+        b.y = (j + x[index + 1 * stride] ) / 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;
 }

@@ -75,12 +88,16 @@ 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){
+                if (this->scaleXY != 1) scalAdd(dstData + index, 2 * dim.w*dim.h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
+            }
+            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);
-            
-            index = entry_index(b, n*dim.w*dim.h, 4, classes, input_dim, output_dim);
-            activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*dim.w*dim.h);
+                if (this->scaleXY != 1) scalAdd(dstData + index, 2 * dim.w*dim.h, this->scaleXY, -0.5*(this->scaleXY - 1), 1);
+                index = entry_index(b, n*dim.w*dim.h, 4, classes, input_dim, output_dim);
+                activationLOGISTICForward(srcData + index, dstData + index, (1+classes)*dim.w*dim.h);
+            }
        }
    }

@@ -116,7 +133,7 @@ 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 +157,7 @@ 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 +210,32 @@ 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 +262,7 @@ 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 +288,13 @@ 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;
-                }
            }
        }
    }
-
 }

 }}
@@ -32,6 +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);
@@ -91,9 +94,10 @@ void Yolo3Detection::preprocess(cv::Mat &frame, const int bi){
 void Yolo3Detection::postprocess(const int bi, const bool mAP){

    //get yolo outputs
-    dnnType *rt_out[netRT->pluginFactory->n_yolos]; 
-    for(int i=0; i<netRT->pluginFactory->n_yolos; i++) 
-        rt_out[i] = (dnnType*)netRT->buffersRT[i+1] + netRT->buffersDIM[i+1].tot()*bi;
+    std::vector<float *> rt_out;
+    //dnnType *rt_out[netRT->pluginFactory->n_yolos];
+    for(int i=0; i<netRT->pluginFactory->n_yolos; i++)
+        rt_out.push_back((dnnType*)netRT->buffersRT[i+1] + netRT->buffersDIM[i+1].tot()*bi);

    float x_ratio =  float(originalSize[bi].width) / float(netRT->input_dim.w);
    float y_ratio =  float(originalSize[bi].height) / float(netRT->input_dim.h);
@@ -102,9 +106,9 @@ 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();
@@ -1,7 +1,7 @@
 #include "kernels.h"

 __global__
-void activation_leaky(dnnType *input, dnnType *output, int size) {
+void activation_leaky(dnnType *input, dnnType *output, int size, float slope) {

    int i = blockDim.x*blockIdx.x + threadIdx.x;

@@ -9,7 +9,7 @@ void activation_leaky(dnnType *input, dnnType *output, int size) {
        if (input[i]>0)
            output[i] = input[i];
        else
-            output[i] = 0.1f*input[i];
+            output[i] = slope*input[i];
    }
 }

@@ -17,12 +17,12 @@ void activation_leaky(dnnType *input, dnnType *output, int size) {
 /**
    ELU activation function
 */
-void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, cudaStream_t stream)
+void activationLEAKYForward(dnnType* srcData, dnnType* dstData, int size, float slope, cudaStream_t stream)
 {
    int blocks = (size+255)/256;
    int threads = 256;
    
-    activation_leaky<<<blocks, threads, 0, stream>>>(srcData, dstData, size);
+    activation_leaky<<<blocks, threads, 0, stream>>>(srcData, dstData, size, slope);
 }


@@ -18,7 +18,7 @@ inline int GET_BLOCKS(const int N)
 }


-__device__ float dmcn_im2col_bilinear(const float *bottom_data, const int data_width,
+__device__ __host__  float dmcn_im2col_bilinear(const float *bottom_data, const int data_width,
  const int height, const int width, float h, float w) {
 int h_low = floor(h);
 int w_low = floor(w);
@@ -34,6 +34,25 @@ void sortAndTopKonDevice(dnnType *src_begin, int *idsrc, float *topk_scores, int
    sortAndTopK_kernel<<<blocks, threads, 0>>>(src_begin, idsrc, topk_scores, topk_inds, topk_ys, topk_xs, size, K);   
 }

+__global__ 
+void maxElem_kernel(float *src_begin, float *dst_begin, const int n_classes, const int size){
+    int i = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i > size)
+        return;
+
+    thrust::device_ptr<float> dPbeg ( &src_begin[i*n_classes] ) ;
+    thrust::device_ptr<float> dPend = dPbeg + n_classes;
+    thrust::device_ptr<float> result = thrust::max_element(thrust::device,dPbeg, dPend);
+
+    dst_begin[i] = result - dPbeg;
+}
+
+void maxElem(dnnType *src_begin, dnnType *dst_begin, const int c, const int h, const int w){
+    int blocks = (h*w)/32+1;
+    int threads = 32;
+    maxElem_kernel<<<blocks, threads, 0>>>(src_begin, dst_begin, c, h*w);   
+}
+
 void topKxyclasses(int *ids_begin, int *ids_end, const int K, const int size, const int wh, int *clses, int *xs, int *ys){    
    thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(wh), clses, thrust::divides<int>());
    thrust::transform(thrust::device, ids_begin, ids_end, thrust::make_constant_iterator(wh), ids_begin, thrust::modulus<int>());
@@ -1,46 +1,33 @@
 #include "kernels.h"
 #include <stdio.h>
-#define MIN(a,b) (((a)<(b))?(a):(b))
-#define MAX(a,b) (((a)>(b))?(a):(b))

-__global__ void resize_kernel(  int i_N,float *x, int i_w, int i_h, int i_c,  
+__global__ void resize_kernel(  int size,float *x, int i_w, int i_h, int i_c,  
                                int o_w, int o_h, int o_c, int batch, float *out)
 {
-    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
-    if(i >= i_N) return;
+    int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+    if(id >= size) return;

-    int out_index = i;
-    int out_w = i%o_w;
-    i = i/o_w;
-    int out_h = i%o_h;
-    i = i/o_h;
-    int out_c = i%o_c;
-    i = i/o_c;
+    int i = id % o_w;
+    id /= o_w;
+    int j = id % o_h;
+    id /= o_h;
+    int k = id % o_c;
+    id /= o_c;
+    int b = id % batch;

-    //copying last column/last row as padding
-    int in_index = ((i*i_c + MIN(out_c,i_c-1))*i_h +  MIN(out_h,i_h-1))*i_w + MIN(out_w, i_w-1);
-    out[out_index] = x[in_index];
+    int out_index = i + o_w*(j + o_h*(k + o_c*b));
+    int add_index = i/(o_w/i_w) + i_w*(j/(o_h/i_h) + i_h*(k + i_c*b));
+    out[out_index] = x[add_index];
 }


 void resizeForward(  dnnType* srcData, dnnType* dstData, int n, int i_c, int i_h, int i_w,
     int o_c, int o_h, int o_w, cudaStream_t stream )
 {
-    int i_size = n*i_c*i_h*i_w;
    int o_size = n*o_c*o_h*o_w;

    int blocks = (o_size+255)/256;
    int threads = 256;

-    if(i_c == o_c && i_h == o_h && i_w == o_w )
-    {
-        checkCuda(cudaMemcpy(dstData, srcData, i_size*sizeof(dnnType), cudaMemcpyDeviceToDevice));
-    }
-    else
-    {
-        checkCuda(cudaMemset(dstData, 0, o_size*sizeof(dnnType)));
-        resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
-        // printDeviceVector(i_size, srcData);
-        // printDeviceVector(o_size, dstData);
-    }
+    resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
 }
@@ -21,27 +21,60 @@ __global__ void shortcut_kernel(int size, int minw, int minh, int minc, int stri
    //out[out_index] += add[add_index];
 }

+__global__ void shortcut_mul_kernel(int size, int minw, int minh, int minc, int sample, int batch, 
+                                int w1, int h1, int c1, dnnType *mul, 
+                                int w2, int h2, int c2, float s1, float s2, dnnType *out)
+{
+    int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+    if (id >= size) return;
+    int i = id % minw;
+    id /= minw;
+    int j = id % minh;
+    id /= minh;
+    int k = id % minc;
+    id /= minc;
+    int b = id % batch;
+
+    int out_index = i*sample + w1*(j*sample + h1*(k + c1*b));
+    out[out_index] = out[out_index] * mul[k + c2*b];
+}
+
 void shortcutForward(dnnType* srcData, dnnType* dstData, int n1, int c1, int h1, int w1, int s1,
                                                         int n2, int c2, int h2, int w2, int s2, 
-                     cudaStream_t stream)
+                                                         bool mul, cudaStream_t stream)
 {
    assert(n1 == n2);
    int batch = n1;

-    int minw = (w1 < w2) ? w1 : w2;
-    int minh = (h1 < h2) ? h1 : h2;
-    int minc = (c1 < c2) ? c1 : c2;
+    if(!mul){
+        int minw = (w1 < w2) ? w1 : w2;
+        int minh = (h1 < h2) ? h1 : h2;
+        int minc = (c1 < c2) ? c1 : c2;
+        int stride = w1/w2;
+        int sample = w2/w1;
+        assert(stride == h1/h2);
+        assert(sample == h2/h1);
+        if(stride < 1) stride = 1;
+        if(sample < 1) sample = 1;

-    int stride = w1/w2;
-    int sample = w2/w1;
-    assert(stride == h1/h2);
-    assert(sample == h2/h1);
-    if(stride < 1) stride = 1;
-    if(sample < 1) sample = 1;
+        int size = batch * minw * minh * minc;
+        int blocks = (size+255)/256;
+        int threads = 256;
+        
+        shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch, 
+            w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);  
+    } 
+    else{
+        int minw = w1;
+        int minh = h1;
+        int minc = c1;
+        int sample = 1;

-    int size = batch * minw * minh * minc;
-    int blocks = (size+255)/256;
-    int threads = 256;
-    shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch, 
-        w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
+        int size = batch * minw * minh * minc;
+        int blocks = (size+255)/256;
+        int threads = 256;
+
+        shortcut_mul_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, sample, batch, 
+            w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);  
+    }
 }
@@ -23,14 +23,23 @@ bool fileExist(const char *fname) {
 void downloadWeightsifDoNotExist(const std::string& input_bin, const std::string& test_folder, const std::string& weights_url){
    if(!fileExist(input_bin.c_str())){
        std::string mkdir_cmd = "mkdir " + test_folder; 
-        std::string wget_cmd = "wget " + weights_url + " -O " + test_folder + "/weights.zip"; 
+        std::string wget_cmd = "curl " + weights_url + " --output " + test_folder + "/weights.zip";
+#ifdef __linux__
        std::string unzip_cmd = "unzip " + test_folder + "/weights.zip -d" + test_folder;
        std::string rm_cmd = "rm " + test_folder + "/weights.zip";
+
+#elif _WIN32
+
+        std::string unzip_cmd = "7z x " + test_folder + "/weights.zip -o" + test_folder;
+#endif 
        int err = 0;
        err = system(mkdir_cmd.c_str());
        err = system(wget_cmd.c_str());
        err = system(unzip_cmd.c_str());
+#ifdef __linux__
        err = system(rm_cmd.c_str());
+#endif
+
    }
 }

@@ -102,6 +111,7 @@ int checkResult(int size, dnnType *data_d, dnnType *correct_d, bool device, int
    }
    int diffs = 0;
    for(int i=0; i<size; i++) {
+        // data_h[i] = data_h[i]*1e-2;
        if(data_h[i] != data_h[i] || correct_h[i] != correct_h[i] || //nan control
           fabs(data_h[i] - correct_h[i]) > eps) {
            diffs += 1;
@@ -193,8 +203,12 @@ void getMemUsage(double& vm_usage_kb, double& resident_set_kb){
               >> O >> itrealvalue >> starttime >> vsize >> rss;

   stat_stream.close();
-
+#ifdef __linux__
   long page_size_kb = sysconf(_SC_PAGE_SIZE) / 1024; // in case x86-64 is configured to use 2MB pages
+#elif _WIN32
+ long page_size_kb = 4096/1024;
+#endif
+
   vm_usage_kb     = vsize / 1024.0;  
   resident_set_kb = rss * page_size_kb;
 }
@@ -1,12 +1,11 @@
 [net]
-# Testing
 batch=1
 subdivisions=1
 # Training
-#batch=64
-#subdivisions=8
-width=416
-height=416
+width=512
+height=512
+# width=608
+# height=608
 channels=3
 momentum=0.949
 decay=0.0005
@@ -15,11 +14,11 @@ saturation = 1.5
 exposure = 1.5
 hue=.1

-learning_rate=0.00261
+learning_rate=0.0013
 burn_in=1000
-max_batches = 500500
+max_batches = 16000
 policy=steps
-steps=400000,450000
+steps=12800,14400
 scales=.1,.1

 #cutmix=1
@@ -959,14 +958,14 @@ activation=leaky
 size=1
 stride=1
 pad=1
-filters=255
+filters=27
 activation=linear


 [yolo]
 mask = 0,1,2
 anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
-classes=80
+classes=4
 num=9
 jitter=.3
 ignore_thresh = .7
@@ -978,6 +977,7 @@ iou_normalizer=0.07
 iou_loss=ciou
 nms_kind=greedynms
 beta_nms=0.6
+max_delta=5


 [route]
@@ -1046,14 +1046,14 @@ activation=leaky
 size=1
 stride=1
 pad=1
-filters=255
+filters=27
 activation=linear


 [yolo]
 mask = 3,4,5
 anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
-classes=80
+classes=4
 num=9
 jitter=.3
 ignore_thresh = .7
@@ -1065,6 +1065,7 @@ iou_normalizer=0.07
 iou_loss=ciou
 nms_kind=greedynms
 beta_nms=0.6
+max_delta=5


 [route]
@@ -1133,14 +1134,14 @@ activation=leaky
 size=1
 stride=1
 pad=1
-filters=255
+filters=27
 activation=linear


 [yolo]
 mask = 6,7,8
 anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
-classes=80
+classes=4
 num=9
 jitter=.3
 ignore_thresh = .7
@@ -1153,4 +1154,5 @@ iou_normalizer=0.07
 iou_loss=ciou
 nms_kind=greedynms
 beta_nms=0.6
+max_delta=5

@@ -0,0 +1,4 @@
+blue-cone
+yellow-cone
+orange-cone
+big-orange-cone
@@ -0,0 +1,34 @@
+#include<iostream>
+#include<vector>
+#include "tkdnn.h"
+#include "test.h"
+#include "DarknetParser.h"
+
+int main() {
+    std::string bin_path  = "yolo4-csp";
+    std::vector<std::string> input_bins = { 
+        bin_path + "/layers/input.bin"
+    };
+    std::vector<std::string> output_bins = {
+        bin_path + "/debug/layer144_out.bin",
+        bin_path + "/debug/layer159_out.bin",
+        bin_path + "/debug/layer174_out.bin"
+    };
+    std::string wgs_path  = bin_path + "/layers";
+    std::string cfg_path  = std::string(TKDNN_PATH) + "/tests/darknet/cfg/yolo4-csp.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/AfzHE4BfTeEm2gH/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;
+}
@@ -5,7 +5,7 @@
 #include "DarknetParser.h"

 int main() {
-    std::string bin_path  = "yolo4_416";
+    std::string bin_path  = "yolo4_mmr";
    std::vector<std::string> input_bins = { 
        bin_path + "/layers/input.bin"
    };
@@ -15,9 +15,9 @@ int main() {
        bin_path + "/debug/layer161_out.bin"
    };
    std::string wgs_path  = bin_path + "/layers";
-    std::string cfg_path  = "../tests/darknet/cfg/yolo4_416.cfg";
-    std::string name_path = "../tests/darknet/names/coco.names";
-    downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/982LxTQcNQfFQc4/download");
+    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);
@@ -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/5MFjtNtgbDGdJEo/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;
+}
@@ -0,0 +1,295 @@
+#include <iostream>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+#include "tkdnn.h"
+#include "NetworkViz.h"
+
+
+const char *input_bin = "shelfnet/debug/input.bin";
+
+const char *backbone[] = {
+    "shelfnet/layers/backbone-conv1.bin",
+    "shelfnet/layers/backbone-layer1-0-conv1.bin",
+    "shelfnet/layers/backbone-layer1-0-conv2.bin",
+    "shelfnet/layers/backbone-layer1-1-conv1.bin",
+    "shelfnet/layers/backbone-layer1-1-conv2.bin",
+    "shelfnet/layers/backbone-layer2-0-conv1.bin",
+    "shelfnet/layers/backbone-layer2-0-conv2.bin",
+    "shelfnet/layers/backbone-layer2-0-downsample-0.bin",
+    "shelfnet/layers/backbone-layer2-1-conv1.bin",
+    "shelfnet/layers/backbone-layer2-1-conv2.bin",
+    "shelfnet/layers/backbone-layer3-0-conv1.bin",
+    "shelfnet/layers/backbone-layer3-0-conv2.bin",
+    "shelfnet/layers/backbone-layer3-0-downsample-0.bin",
+    "shelfnet/layers/backbone-layer3-1-conv1.bin",
+    "shelfnet/layers/backbone-layer3-1-conv2.bin",
+    "shelfnet/layers/backbone-layer4-0-conv1.bin",
+    "shelfnet/layers/backbone-layer4-0-conv2.bin",
+    "shelfnet/layers/backbone-layer4-0-downsample-0.bin",
+    "shelfnet/layers/backbone-layer4-1-conv1.bin",
+    "shelfnet/layers/backbone-layer4-1-conv2.bin"};
+
+const char *conv_out[] = {
+    "shelfnet/layers/conv_out-conv-conv.bin",
+    "shelfnet/layers/conv_out-conv_out.bin",
+    "shelfnet/layers/conv_out16-conv-conv.bin",
+    "shelfnet/layers/conv_out16-conv_out.bin",
+    "shelfnet/layers/conv_out32-conv-conv.bin",
+    "shelfnet/layers/conv_out32-conv_out.bin"
+    };
+
+const char *decoder[] = {
+    "shelfnet/layers/decoder-bottom-conv1.bin",
+    "shelfnet/layers/decoder-bottom-conv12.bin",
+    "shelfnet/layers/decoder-up_conv_list-0-conv-conv.bin",
+    "shelfnet/layers/decoder-up_conv_list-0-conv_atten.bin",
+    "shelfnet/layers/decoder-up_dense_list-0-conv.bin",
+    "shelfnet/layers/decoder-up_conv_list-1-conv-conv.bin",
+    "shelfnet/layers/decoder-up_conv_list-1-conv_atten.bin",
+    "shelfnet/layers/decoder-up_dense_list-1-conv.bin"
+    };
+
+    
+const char *ladder[] = {
+    "shelfnet/layers/ladder-inconv-conv1.bin",
+    "shelfnet/layers/ladder-inconv-conv12.bin",
+    "shelfnet/layers/ladder-down_module_list-0-conv1.bin",
+    "shelfnet/layers/ladder-down_module_list-0-conv12.bin",
+    "shelfnet/layers/ladder-down_conv_list-0.bin",
+
+    "shelfnet/layers/ladder-down_module_list-1-conv1.bin",
+    "shelfnet/layers/ladder-down_module_list-1-conv12.bin",
+    "shelfnet/layers/ladder-down_conv_list-1.bin",
+
+    "shelfnet/layers/ladder-bottom-conv1.bin",
+    "shelfnet/layers/ladder-bottom-conv12.bin",
+    
+    
+    
+    "shelfnet/layers/ladder-up_conv_list-0-conv-conv.bin",
+    "shelfnet/layers/ladder-up_conv_list-0-conv_atten.bin",
+    "shelfnet/layers/ladder-up_dense_list-0-conv.bin",
+
+    
+    "shelfnet/layers/ladder-up_conv_list-1-conv-conv.bin",
+    "shelfnet/layers/ladder-up_conv_list-1-conv_atten.bin",
+    "shelfnet/layers/ladder-up_dense_list-1-conv.bin"};
+
+const char *trans[] = {
+    "shelfnet/layers/trans1-conv.bin",
+    "shelfnet/layers/trans2-conv.bin",
+    "shelfnet/layers/trans3-conv.bin"};
+int main()
+{
+
+    downloadWeightsifDoNotExist(input_bin, "shelfnet", "https://cloud.hipert.unimore.it/s/mEDZMRJaGCFWSJF/download");
+
+    int classes = 19;
+
+    // Network layout    
+    tk::dnn::dataDim_t dim(1, 3, 1024, 1024, 1);
+    tk::dnn::Network net(dim);
+
+    int bi = 0, di = 0, li = 0, ci = 0;
+    new tk::dnn::Conv2d(&net, 64, 7, 7, 2, 2, 3, 3, backbone[bi++], true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    tk::dnn::Layer* last = new tk::dnn::Pooling (&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
+
+
+    
+    for(int i=0; i<2; ++i){
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    }
+
+    std::vector<tk::dnn::Layer*> features;
+    for(int i=0;i<3;++i){
+        int out_channel = pow(2,7+i);
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 2, 2, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        tk::dnn::Layer* bn2 = new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 2, 2, 0, 0, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, bn2);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        features.push_back(last);
+    }
+
+    for(int i=0; i<features.size(); ++i){
+        new tk::dnn::Route(&net, &features[i], 1);
+        int out_channel = pow(2,6+i);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
+        features[i] = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    }
+
+    //DECODER
+    
+    last = features[2];
+    std::vector<tk::dnn::Layer*> up_out;
+    //bottom
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, decoder[di++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, features[1-i]);
+
+        //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+    //LADDER
+
+    std::vector<tk::dnn::Layer*> down_out;
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    
+    for(int i=0; i<2;++i){
+        int out_channel = pow(2,6+i);
+        tk::dnn::Layer* l_last = new tk::dnn::Shortcut(&net, up_out[2-i]);
+    
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Shortcut(&net, l_last);
+        l_last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        down_out.push_back(l_last);
+
+        new tk::dnn::Conv2d (&net, out_channel*2, 3, 3, 2, 2, 1, 1, ladder[li++], false);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.0f); //should be ReLU
+    }
+
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.clear();
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, ladder[li++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, down_out[1-i]);
+
+        // //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+
+    // for(int i=2;i>=0;--i){
+        // new tk::dnn::Route(&net, &up_out[i], 1);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, conv_out[ci++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, 19, 3, 3, 1, 1, 1, 1, conv_out[ci++], false);
+        /*up_out[i] =*/ new tk::dnn::Resize(&net, 19, net.input_dim.h, net.input_dim.w, true, tk::dnn::ResizeMode_t::LINEAR);
+    // }
+
+    new tk::dnn::Softmax(&net);
+    
+    const char *output_bin = "shelfnet/debug/softmax.bin";
+    
+    // Load input
+    dnnType *data;
+    dnnType *input_h;
+    readBinaryFile(input_bin, dim.tot(), &input_h, &data);
+    std::cout<<"Input:"<<std::endl;
+
+    //print network model
+    net.print();
+
+    // // convert network to tensorRT
+    tk::dnn::NetworkRT netRT(&net, net.getNetworkRTName("shelfnet"));
+
+    tk::dnn::dataDim_t dim1 = dim; //input dim
+    dnnType *cudnn_out = nullptr;
+    printCenteredTitle(" CUDNN inference ", '=', 30);
+    {
+        dim1.print();
+        TKDNN_TSTART
+        cudnn_out = net.infer(dim1, data);
+        TKDNN_TSTOP
+        dim1.print();
+    }
+
+    tk::dnn::dataDim_t dim2 = dim;
+    printCenteredTitle(" TENSORRT inference ", '=', 30);
+    {
+        dim2.print();
+        TKDNN_TSTART
+        netRT.infer(dim2, data);
+        TKDNN_TSTOP
+        dim2.print();
+    }
+
+    dnnType *rt_out1 = (dnnType *)netRT.buffersRT[1];
+
+    printCenteredTitle(std::string(" CHECK RESULTS ").c_str(), '=', 30);
+    dnnType *out1, *out1_h;
+    int odim1 = dim1.tot();
+    readBinaryFile(output_bin, odim1, &out1_h, &out1);
+
+    int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0; 
+    std::cout << "CUDNN vs correct" << std::endl;
+    ret_cudnn |= checkResult(odim1, cudnn_out, out1, true, 20) == 0 ? 0 : ERROR_CUDNN;
+
+    std::cout << "TRT   vs correct" << std::endl;
+    ret_tensorrt |=checkResult(odim1, rt_out1, out1) == 0 ? 0 : ERROR_TENSORRT;
+
+    std::cout << "CUDNN vs TRT    " << std::endl;
+    ret_cudnn_tensorrt |= checkResult(odim1, cudnn_out, rt_out1) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
+    
+    cv::Mat viz = vizLayer2Mat(&net, net.num_layers-1);
+    cv::imwrite("test.png", viz);
+
+    return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
+}
@@ -0,0 +1,295 @@
+#include <iostream>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+#include "tkdnn.h"
+#include "NetworkViz.h"
+
+
+const char *input_bin = "shelfnet_berkeley/debug/input.bin";
+
+const char *backbone[] = {
+    "shelfnet_berkeley/layers/backbone-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer1-0-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer1-0-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer1-1-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer1-1-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer2-0-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer2-0-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer2-0-downsample-0.bin",
+    "shelfnet_berkeley/layers/backbone-layer2-1-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer2-1-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer3-0-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer3-0-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer3-0-downsample-0.bin",
+    "shelfnet_berkeley/layers/backbone-layer3-1-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer3-1-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer4-0-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer4-0-conv2.bin",
+    "shelfnet_berkeley/layers/backbone-layer4-0-downsample-0.bin",
+    "shelfnet_berkeley/layers/backbone-layer4-1-conv1.bin",
+    "shelfnet_berkeley/layers/backbone-layer4-1-conv2.bin"};
+
+const char *conv_out[] = {
+    "shelfnet_berkeley/layers/conv_out-conv-conv.bin",
+    "shelfnet_berkeley/layers/conv_out-conv_out.bin",
+    "shelfnet_berkeley/layers/conv_out16-conv-conv.bin",
+    "shelfnet_berkeley/layers/conv_out16-conv_out.bin",
+    "shelfnet_berkeley/layers/conv_out32-conv-conv.bin",
+    "shelfnet_berkeley/layers/conv_out32-conv_out.bin"
+    };
+
+const char *decoder[] = {
+    "shelfnet_berkeley/layers/decoder-bottom-conv1.bin",
+    "shelfnet_berkeley/layers/decoder-bottom-conv12.bin",
+    "shelfnet_berkeley/layers/decoder-up_conv_list-0-conv-conv.bin",
+    "shelfnet_berkeley/layers/decoder-up_conv_list-0-conv_atten.bin",
+    "shelfnet_berkeley/layers/decoder-up_dense_list-0-conv.bin",
+    "shelfnet_berkeley/layers/decoder-up_conv_list-1-conv-conv.bin",
+    "shelfnet_berkeley/layers/decoder-up_conv_list-1-conv_atten.bin",
+    "shelfnet_berkeley/layers/decoder-up_dense_list-1-conv.bin"
+    };
+
+    
+const char *ladder[] = {
+    "shelfnet_berkeley/layers/ladder-inconv-conv1.bin",
+    "shelfnet_berkeley/layers/ladder-inconv-conv12.bin",
+    "shelfnet_berkeley/layers/ladder-down_module_list-0-conv1.bin",
+    "shelfnet_berkeley/layers/ladder-down_module_list-0-conv12.bin",
+    "shelfnet_berkeley/layers/ladder-down_conv_list-0.bin",
+
+    "shelfnet_berkeley/layers/ladder-down_module_list-1-conv1.bin",
+    "shelfnet_berkeley/layers/ladder-down_module_list-1-conv12.bin",
+    "shelfnet_berkeley/layers/ladder-down_conv_list-1.bin",
+
+    "shelfnet_berkeley/layers/ladder-bottom-conv1.bin",
+    "shelfnet_berkeley/layers/ladder-bottom-conv12.bin",
+    
+    
+    
+    "shelfnet_berkeley/layers/ladder-up_conv_list-0-conv-conv.bin",
+    "shelfnet_berkeley/layers/ladder-up_conv_list-0-conv_atten.bin",
+    "shelfnet_berkeley/layers/ladder-up_dense_list-0-conv.bin",
+
+    
+    "shelfnet_berkeley/layers/ladder-up_conv_list-1-conv-conv.bin",
+    "shelfnet_berkeley/layers/ladder-up_conv_list-1-conv_atten.bin",
+    "shelfnet_berkeley/layers/ladder-up_dense_list-1-conv.bin"};
+
+const char *trans[] = {
+    "shelfnet_berkeley/layers/trans1-conv.bin",
+    "shelfnet_berkeley/layers/trans2-conv.bin",
+    "shelfnet_berkeley/layers/trans3-conv.bin"};
+int main()
+{
+
+    downloadWeightsifDoNotExist(input_bin, "shelfnet_berkeley", "https://cloud.hipert.unimore.it/s/m92e7QdD9gYMF7f/download");
+
+    int classes = 20;
+
+    // Network layout    
+    tk::dnn::dataDim_t dim(1, 3, 736, 1280, 1);
+    tk::dnn::Network net(dim);
+
+    int bi = 0, di = 0, li = 0, ci = 0;
+    new tk::dnn::Conv2d(&net, 64, 7, 7, 2, 2, 3, 3, backbone[bi++], true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    tk::dnn::Layer* last = new tk::dnn::Pooling (&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
+
+
+    
+    for(int i=0; i<2; ++i){
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    }
+
+    std::vector<tk::dnn::Layer*> features;
+    for(int i=0;i<3;++i){
+        int out_channel = pow(2,7+i);
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 2, 2, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        tk::dnn::Layer* bn2 = new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 2, 2, 0, 0, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, bn2);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        features.push_back(last);
+    }
+
+    for(int i=0; i<features.size(); ++i){
+        new tk::dnn::Route(&net, &features[i], 1);
+        int out_channel = pow(2,6+i);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
+        features[i] = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    }
+
+    //DECODER
+    
+    last = features[2];
+    std::vector<tk::dnn::Layer*> up_out;
+    //bottom
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, decoder[di++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, features[1-i]);
+
+        //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+    //LADDER
+
+    std::vector<tk::dnn::Layer*> down_out;
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    
+    for(int i=0; i<2;++i){
+        int out_channel = pow(2,6+i);
+        tk::dnn::Layer* l_last = new tk::dnn::Shortcut(&net, up_out[2-i]);
+    
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Shortcut(&net, l_last);
+        l_last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        down_out.push_back(l_last);
+
+        new tk::dnn::Conv2d (&net, out_channel*2, 3, 3, 2, 2, 1, 1, ladder[li++], false);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.0f); //should be ReLU
+    }
+
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.clear();
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, ladder[li++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, down_out[1-i]);
+
+        // //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+
+    // for(int i=2;i>=0;--i){
+        // new tk::dnn::Route(&net, &up_out[i], 1);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, conv_out[ci++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, classes, 3, 3, 1, 1, 1, 1, conv_out[ci++], false);
+        /*up_out[i] =*/ new tk::dnn::Resize(&net, classes, net.input_dim.h, net.input_dim.w, true, tk::dnn::ResizeMode_t::LINEAR);
+    // }
+
+    new tk::dnn::Softmax(&net);
+    
+    const char *output_bin = "shelfnet_berkeley/debug/softmax.bin";
+    
+    // Load input
+    dnnType *data;
+    dnnType *input_h;
+    readBinaryFile(input_bin, dim.tot(), &input_h, &data);
+    std::cout<<"Input:"<<std::endl;
+
+    //print network model
+    net.print();
+
+    // // convert network to tensorRT
+    tk::dnn::NetworkRT netRT(&net, net.getNetworkRTName("shelfnet_berkeley"));
+
+    tk::dnn::dataDim_t dim1 = dim; //input dim
+    dnnType *cudnn_out = nullptr;
+    printCenteredTitle(" CUDNN inference ", '=', 30);
+    {
+        dim1.print();
+        TKDNN_TSTART
+        cudnn_out = net.infer(dim1, data);
+        TKDNN_TSTOP
+        dim1.print();
+    }
+
+    tk::dnn::dataDim_t dim2 = dim;
+    printCenteredTitle(" TENSORRT inference ", '=', 30);
+    {
+        dim2.print();
+        TKDNN_TSTART
+        netRT.infer(dim2, data);
+        TKDNN_TSTOP
+        dim2.print();
+    }
+
+    dnnType *rt_out1 = (dnnType *)netRT.buffersRT[1];
+
+    printCenteredTitle(std::string(" CHECK RESULTS ").c_str(), '=', 30);
+    dnnType *out1, *out1_h;
+    int odim1 = dim1.tot();
+    readBinaryFile(output_bin, odim1, &out1_h, &out1);
+
+    int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0; 
+    std::cout << "CUDNN vs correct" << std::endl;
+    ret_cudnn |= checkResult(odim1, cudnn_out, out1, true, 20) == 0 ? 0 : ERROR_CUDNN;
+
+    std::cout << "TRT   vs correct" << std::endl;
+    ret_tensorrt |=checkResult(odim1, rt_out1, out1) == 0 ? 0 : ERROR_TENSORRT;
+
+    std::cout << "CUDNN vs TRT    " << std::endl;
+    ret_cudnn_tensorrt |= checkResult(odim1, cudnn_out, rt_out1) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
+    
+    cv::Mat viz = vizLayer2Mat(&net, net.num_layers-1);
+    cv::imwrite("test.png", viz);
+
+    return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
+}
@@ -0,0 +1,297 @@
+#include <iostream>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+#include "tkdnn.h"
+#include "NetworkViz.h"
+
+
+const char *input_bin = "shelfnet_mapillary/debug/input.bin";
+
+const char *backbone[] = {
+    "shelfnet_mapillary/layers/backbone-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer1-0-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer1-0-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer1-1-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer1-1-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer2-0-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer2-0-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer2-0-downsample-0.bin",
+    "shelfnet_mapillary/layers/backbone-layer2-1-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer2-1-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer3-0-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer3-0-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer3-0-downsample-0.bin",
+    "shelfnet_mapillary/layers/backbone-layer3-1-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer3-1-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer4-0-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer4-0-conv2.bin",
+    "shelfnet_mapillary/layers/backbone-layer4-0-downsample-0.bin",
+    "shelfnet_mapillary/layers/backbone-layer4-1-conv1.bin",
+    "shelfnet_mapillary/layers/backbone-layer4-1-conv2.bin"};
+
+const char *conv_out[] = {
+    "shelfnet_mapillary/layers/conv_out-conv-conv.bin",
+    "shelfnet_mapillary/layers/conv_out-conv_out.bin",
+    "shelfnet_mapillary/layers/conv_out16-conv-conv.bin",
+    "shelfnet_mapillary/layers/conv_out16-conv_out.bin",
+    "shelfnet_mapillary/layers/conv_out32-conv-conv.bin",
+    "shelfnet_mapillary/layers/conv_out32-conv_out.bin"
+    };
+
+const char *decoder[] = {
+    "shelfnet_mapillary/layers/decoder-bottom-conv1.bin",
+    "shelfnet_mapillary/layers/decoder-bottom-conv12.bin",
+    "shelfnet_mapillary/layers/decoder-up_conv_list-0-conv-conv.bin",
+    "shelfnet_mapillary/layers/decoder-up_conv_list-0-conv_atten.bin",
+    "shelfnet_mapillary/layers/decoder-up_dense_list-0-conv.bin",
+    "shelfnet_mapillary/layers/decoder-up_conv_list-1-conv-conv.bin",
+    "shelfnet_mapillary/layers/decoder-up_conv_list-1-conv_atten.bin",
+    "shelfnet_mapillary/layers/decoder-up_dense_list-1-conv.bin"
+    };
+
+    
+const char *ladder[] = {
+    "shelfnet_mapillary/layers/ladder-inconv-conv1.bin",
+    "shelfnet_mapillary/layers/ladder-inconv-conv12.bin",
+    "shelfnet_mapillary/layers/ladder-down_module_list-0-conv1.bin",
+    "shelfnet_mapillary/layers/ladder-down_module_list-0-conv12.bin",
+    "shelfnet_mapillary/layers/ladder-down_conv_list-0.bin",
+
+    "shelfnet_mapillary/layers/ladder-down_module_list-1-conv1.bin",
+    "shelfnet_mapillary/layers/ladder-down_module_list-1-conv12.bin",
+    "shelfnet_mapillary/layers/ladder-down_conv_list-1.bin",
+
+    "shelfnet_mapillary/layers/ladder-bottom-conv1.bin",
+    "shelfnet_mapillary/layers/ladder-bottom-conv12.bin",
+    
+    
+    
+    "shelfnet_mapillary/layers/ladder-up_conv_list-0-conv-conv.bin",
+    "shelfnet_mapillary/layers/ladder-up_conv_list-0-conv_atten.bin",
+    "shelfnet_mapillary/layers/ladder-up_dense_list-0-conv.bin",
+
+    
+    "shelfnet_mapillary/layers/ladder-up_conv_list-1-conv-conv.bin",
+    "shelfnet_mapillary/layers/ladder-up_conv_list-1-conv_atten.bin",
+    "shelfnet_mapillary/layers/ladder-up_dense_list-1-conv.bin"};
+
+const char *trans[] = {
+    "shelfnet_mapillary/layers/trans1-conv.bin",
+    "shelfnet_mapillary/layers/trans2-conv.bin",
+    "shelfnet_mapillary/layers/trans3-conv.bin"};
+int main()
+{
+
+    // downloadWeightsifDoNotExist(input_bin, "shelfnet_mapillary", "");
+    // download the weights from here:  https://cloud.hipert.unimore.it/f/652476
+
+    // Mapillary Vistas has originally 66 classes, but we reduced them to 15 to improve the results on the categories of our interest.
+    int classes = 15;
+
+    // Network layout    
+    tk::dnn::dataDim_t dim(1, 3, 1024, 1024, 1);
+    tk::dnn::Network net(dim);
+
+    int bi = 0, di = 0, li = 0, ci = 0;
+    new tk::dnn::Conv2d(&net, 64, 7, 7, 2, 2, 3, 3, backbone[bi++], true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    tk::dnn::Layer* last = new tk::dnn::Pooling (&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
+
+
+    
+    for(int i=0; i<2; ++i){
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    }
+
+    std::vector<tk::dnn::Layer*> features;
+    for(int i=0;i<3;++i){
+        int out_channel = pow(2,7+i);
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 2, 2, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        tk::dnn::Layer* bn2 = new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 2, 2, 0, 0, backbone[bi++], true);
+        new tk::dnn::Shortcut(&net, bn2);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
+        
+        new tk::dnn::Shortcut(&net, last);
+        last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        features.push_back(last);
+    }
+
+    for(int i=0; i<features.size(); ++i){
+        new tk::dnn::Route(&net, &features[i], 1);
+        int out_channel = pow(2,6+i);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
+        features[i] = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    }
+
+    //DECODER
+    
+    last = features[2];
+    std::vector<tk::dnn::Layer*> up_out;
+    //bottom
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        std::cout<<out_channel<<std::endl;
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, decoder[di++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, features[1-i]);
+
+        //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+    //LADDER
+
+    std::vector<tk::dnn::Layer*> down_out;
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    
+    for(int i=0; i<2;++i){
+        int out_channel = pow(2,6+i);
+        tk::dnn::Layer* l_last = new tk::dnn::Shortcut(&net, up_out[2-i]);
+    
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+        new tk::dnn::Shortcut(&net, l_last);
+        l_last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+        down_out.push_back(l_last);
+
+        new tk::dnn::Conv2d (&net, out_channel*2, 3, 3, 2, 2, 1, 1, ladder[li++], false);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.0f); //should be ReLU
+    }
+
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+    new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
+    new tk::dnn::Shortcut(&net, last);
+    last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
+    up_out.clear();
+    up_out.push_back(last);
+
+    for(int i=0; i<2; ++i){
+        int out_channel = pow(2,7-i);
+        //up-conv
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        
+        new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
+        new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, ladder[li++], true);
+        
+        tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
+        new tk::dnn::Route(&net, &last, 1);
+        new tk::dnn::Shortcut(&net, act, true);
+
+        //interpolate
+        new tk::dnn::Resize(&net, 1,2,2);
+        new tk::dnn::Shortcut(&net, down_out[1-i]);
+
+        // //up-dense
+        new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
+        last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        up_out.push_back(last);
+    }
+
+
+    // for(int i=2;i>=0;--i){
+        // new tk::dnn::Route(&net, &up_out[i], 1);
+        new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, conv_out[ci++], true);
+        new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY, 0.0f, 0.01);
+        new tk::dnn::Conv2d (&net, classes, 3, 3, 1, 1, 1, 1, conv_out[ci++], false);
+        /*up_out[i] =*/ new tk::dnn::Resize(&net, classes, net.input_dim.h, net.input_dim.w, true, tk::dnn::ResizeMode_t::LINEAR);
+    // }
+
+    new tk::dnn::Softmax(&net);
+    
+    const char *output_bin = "shelfnet_mapillary/debug/softmax.bin";
+    
+    // Load input
+    dnnType *data;
+    dnnType *input_h;
+    readBinaryFile(input_bin, dim.tot(), &input_h, &data);
+    std::cout<<"Input:"<<std::endl;
+
+    //print network model
+    net.print();
+
+    // // convert network to tensorRT
+    tk::dnn::NetworkRT netRT(&net, net.getNetworkRTName("shelfnet_mapillary"));
+
+    tk::dnn::dataDim_t dim1 = dim; //input dim
+    dnnType *cudnn_out = nullptr;
+    printCenteredTitle(" CUDNN inference ", '=', 30);
+    {
+        dim1.print();
+        TKDNN_TSTART
+        cudnn_out = net.infer(dim1, data);
+        TKDNN_TSTOP
+        dim1.print();
+    }
+
+    tk::dnn::dataDim_t dim2 = dim;
+    printCenteredTitle(" TENSORRT inference ", '=', 30);
+    {
+        dim2.print();
+        TKDNN_TSTART
+        netRT.infer(dim2, data);
+        TKDNN_TSTOP
+        dim2.print();
+    }
+
+    dnnType *rt_out1 = (dnnType *)netRT.buffersRT[1];
+
+    printCenteredTitle(std::string(" CHECK RESULTS ").c_str(), '=', 30);
+    dnnType *out1, *out1_h;
+    int odim1 = dim1.tot();
+    readBinaryFile(output_bin, odim1, &out1_h, &out1);
+
+    int ret_cudnn = 0, ret_tensorrt = 0, ret_cudnn_tensorrt = 0; 
+    std::cout << "CUDNN vs correct" << std::endl;
+    ret_cudnn |= checkResult(odim1, cudnn_out, out1, true, 20) == 0 ? 0 : ERROR_CUDNN;
+
+    std::cout << "TRT   vs correct" << std::endl;
+    ret_tensorrt |=checkResult(odim1, rt_out1, out1) == 0 ? 0 : ERROR_TENSORRT;
+
+    std::cout << "CUDNN vs TRT    " << std::endl;
+    ret_cudnn_tensorrt |= checkResult(odim1, cudnn_out, rt_out1) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
+    
+    cv::Mat viz = vizLayer2Mat(&net, net.num_layers-1);
+    cv::imwrite("test.png", viz);
+
+    return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
+}
@@ -35,7 +35,7 @@ int main(int argc, char *argv[]) {
    std::vector<double> stats;
    printCenteredTitle(" TENSORRT inference ", '=', 30); 
    float total_time = 0;
-    for(int i=0; i<1200; i++) {
+    for(int i=0; i<64; i++) {

        // generate input
        for(int j=0; j<netRT.input_dim.tot(); j++) {
				`@@ -0,0 +1 @@`
				`1)error C2131 @ Yolo3Detection.cpp(97) -> expression doesnt evaluate to a constant caused to read of variable outside its lifetime`