diff --git a/CMakeLists.txt b/CMakeLists.txt index 0a390c6..b959c19 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -3,10 +3,10 @@ cmake_minimum_required(VERSION 3.15) project (tkDNN) set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake) if(UNIX) -set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -fPIC -Wno-deprecated-declarations -Wno-unused-variable ") +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++14 -fPIC -Wno-deprecated-declarations") endif() if(WIN32) -set(CMAKE_CXX_STANDARD 11) +set(CMAKE_CXX_STANDARD 14) set(CMAKE_CXX_FLAGS "/O2 /FS /EHsc") set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) endif(WIN32) @@ -140,6 +140,9 @@ target_link_libraries(test_shelfnet_berkeley tkDNN) add_executable(test_shelfnet_mapillary tests/shelfnet/shelfnet_mapillary.cpp) target_link_libraries(test_shelfnet_mapillary tkDNN) +add_executable(test_shelfnet_coco tests/shelfnet/shelfnet_coco.cpp) +target_link_libraries(test_shelfnet_coco tkDNN) + # DEMOS add_executable(test_rtinference tests/test_rtinference/rtinference.cpp) target_link_libraries(test_rtinference tkDNN) diff --git a/README.md b/README.md index b630fec..ddfd4ef 100644 --- a/README.md +++ b/README.md @@ -17,10 +17,15 @@ If you use tkDNN in your research, please cite the [following paper](https://iee } ``` -### What's new (20 July 2021) +### What's new +#### 20 July 2021 - [x] Support to sematic segmentation [README](docs/README_seg.md) - [x] Support 2D/3D Object Detection and Tracking [README](docs/README_2d3dtracking.md) -- [ ] Support to TensorRT8 (WIP) +#### 24 November 2021 +- [x] Support to sematic segmentation on cuda 11 +- [x] Support to TensorRT8 (thanks to [Harshvardhan Chandirasekar](https://github.com/perseusdg)). + +TensorRT8 (and therefore Jetpack 4.6) is currently supported only on the branch tensorrt8 due to [performance issue with TensorRT8](https://docs.nvidia.com/deeplearning/tensorrt/release-notes/tensorrt-8.html)). We will merge it to the master as soon as those issues are fixed (probably in future minor releases). ## FPS Results Inference FPS of yolov4 with tkDNN, average of 1200 images with the same dimension as the input size, on @@ -168,11 +173,10 @@ For specific details on how to run tkDNN on Windows 10 see [HERE](./docs/windows | yolo4_320 | Yolov4 8 | [COCO 2017](http://cocodataset.org/) | 80 | 320x320 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) | | yolo4_512 | Yolov4 8 | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) | | yolo4_608 | Yolov4 8 | [COCO 2017](http://cocodataset.org/) | 80 | 608x608 | [weights](https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download) | -| yolo4_berkeley | Yolov4 8 | [BDD100K ](https://bair.berkeley.edu/blog/2018/05/30/bdd/) | 10 | 540x320 | [weights](https://cloud.hipert.unimore.it/s/nkWFa5fgb4NTdnB/download) | +| yolo4_berkeley | Yolov4 8 | [BDD100K ](https://bair.berkeley.edu/blog/2018/05/30/bdd/) | 10 | 544x320 | [weights](https://cloud.hipert.unimore.it/s/nkWFa5fgb4NTdnB/download) | | yolo4tiny | Yolov4 tiny 9 | [COCO 2017](http://cocodataset.org/) | 80 | 416x416 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) | -| yolo4x | Yolov4x-mish 9 | [COCO 2017](http://cocodataset.org/) | +| yolo4x | Yolov4x-mish 9 | [COCO 2017](http://cocodataset.org/) | 80 | 640x640 | [weights](https://cloud.hipert.unimore.it/s/5MFjtNtgbDGdJEo/download) | | yolo4tiny_512 | Yolov4 tiny 9 | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/iRnc4pSqmx78gJs/download) | -80 | 640x640 | [weights](https://cloud.hipert.unimore.it/s/5MFjtNtgbDGdJEo/download) | | yolo4x-cps | Scaled Yolov4 10 | [COCO 2017](http://cocodataset.org/) | 80 | 512x512 | [weights](https://cloud.hipert.unimore.it/s/AfzHE4BfTeEm2gH/download) | | shelfnet | ShelfNet18_realtime11 | [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x1024 | [weights](https://cloud.hipert.unimore.it/s/mEDZMRJaGCFWSJF/download) | | shelfnet_berkeley | ShelfNet18_realtime11 | [DeepDrive](https://bdd-data.berkeley.edu/) | 20 | 1024x1024 | [weights](https://cloud.hipert.unimore.it/s/m92e7QdD9gYMF7f/download) | diff --git a/demo/demo/demo.cpp b/demo/demo/demo.cpp index 317a574..5445d86 100644 --- a/demo/demo/demo.cpp +++ b/demo/demo/demo.cpp @@ -9,7 +9,6 @@ #include "Yolo3Detection.h" bool gRun; -bool SAVE_RESULT = false; void sig_handler(int signo) { std::cout<<"request gateway stop\n"; @@ -18,43 +17,53 @@ void sig_handler(int signo) { int main(int argc, char *argv[]) { - std::cout<<"detection\n"; signal(SIGINT, sig_handler); - - std::string net = "yolo4tiny_fp32.rt"; - if(argc > 1) - net = argv[1]; + // get config file path and read it #ifdef __linux__ - std::string input = "../demo/yolo_test.mp4"; + std::string config_file = "../demo/demoConfig.yaml"; #elif _WIN32 - std::string input = "..\\..\\..\\demo\\yolo_test.mp4"; + std::string config_file = "..\\..\\..\\demo\\demoConfig.yaml"; #endif + if(argc > 1) + config_file = argv[1]; + + YAML::Node conf = YAMLloadConf(config_file); + if(!conf) + FatalError("Problem with config file"); - if(argc > 2) - input = argv[2]; - char ntype = 'y'; - if(argc > 3) - ntype = argv[3][0]; - int n_classes = 80; - if(argc > 4) - n_classes = atoi(argv[4]); - int n_batch = 1; - if(argc > 5) - n_batch = atoi(argv[5]); - bool show = true; - if(argc > 6) - show = atoi(argv[6]); - float conf_thresh=0.3; - if(argc > 7) - conf_thresh = atof(argv[7]); + // read settings from config file + std::string net = YAMLgetConf(conf, "net", "yolo4tiny_fp32.rt"); + if(!fileExist(net.c_str())) + FatalError("The given network does not exist. Create the rt first."); + #ifdef __linux__ + std::string input = YAMLgetConf(conf, "input", "../demo/yolo_test.mp4"); + #elif _WIN32 + std::string input = YAMLgetConf(conf, "win_input", "..\\..\\..\\demo\\yolo_test.mp4"); + #endif + if(!fileExist(input.c_str())) + FatalError("The given input video does not exist."); + + char ntype = YAMLgetConf(conf, "ntype", 'y'); + int n_classes = YAMLgetConf(conf, "n_classes", 80); + int n_batch = YAMLgetConf(conf, "n_batch", 1); if(n_batch < 1 || n_batch > 64) FatalError("Batch dim not supported"); + float conf_thresh = YAMLgetConf(conf, "conf_thresh", 0.3); + bool show = YAMLgetConf(conf, "show", true); + bool save = YAMLgetConf(conf, "save", false); - if(!show) - SAVE_RESULT = true; - + std::cout <<"Net settings - net: "<< net + <<", ntype: "<< ntype + <<", n_classes: "<< n_classes + <<", n_batch: "<< n_batch + <<", conf_thresh: "<< conf_thresh<<"\n"; + std::cout <<"Demo settings - input: "<< input + <<", show: "<< show + <<", save: "<< save<<"\n\n"; + + // create detection network tk::dnn::Yolo3Detection yolo; tk::dnn::CenternetDetection cnet; tk::dnn::MobilenetDetection mbnet; @@ -79,8 +88,7 @@ int main(int argc, char *argv[]) { detNN->init(net, n_classes, n_batch, conf_thresh); - gRun = true; - + // open video stream cv::VideoCapture cap(input); if(!cap.isOpened()) gRun = false; @@ -88,19 +96,21 @@ int main(int argc, char *argv[]) { std::cout<<"camera started\n"; cv::VideoWriter resultVideo; - if(SAVE_RESULT) { + if(save) { int w = cap.get(cv::CAP_PROP_FRAME_WIDTH); int 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; if(show) cv::namedWindow("detection", cv::WINDOW_NORMAL); + cv::Mat frame; std::vector batch_frame; std::vector batch_dnn_input; + // start detection loop + gRun = true; while(gRun) { batch_dnn_input.clear(); batch_frame.clear(); @@ -128,19 +138,18 @@ int main(int argc, char *argv[]) { cv::waitKey(1); } } - if(n_batch == 1 && SAVE_RESULT) + if(n_batch == 1 && save) resultVideo << frame; } std::cout<<"detection end\n"; - double mean = 0; + double mean = 0; std::cout<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; istats.size(); i++) mean += detNN->stats[i]; mean /= detNN->stats.size(); std::cout<<"Avg: "< ``` -In general the demo program takes 7 parameters: -``` -./demo -``` -where +In general the demo program takes 1 parameter, the `````` that is the path to che configuration file. The parameter is optional and its default value is ```"../demo/demoConfig.yaml"```. -* `````` is the rt file generated by a test -* ```<``` is the path to a video file or a camera input -* `````` is the type of network. Thee types are currently supported: ```y``` (YOLO family), ```c``` (CenterNet family) and ```m``` (MobileNet-SSD family) -* ``````is the number of classes the network is trained on -* `````` 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). -* `````` if set to 0 the demo will not show the visualization but save the video into result.mp4 (if n-batches ==1) -* `````` confidence threshold for the detector. Only bounding boxes with threshold greater than conf-thresh will be displayed. +The config file is a yaml file with the following attributes: +* ```net``` is the rt file generated by a test +* ```input``` is the path to a video file or a camera input (on Linux) +* ```win_input``` is the path to a video file or a camera input (on Windows) +* ```ntype``` is the type of network. Thee types are currently supported: ```y``` (YOLO family), ```c``` (CenterNet family) and ```m``` (MobileNet-SSD family) +* ```n_classes``` is the number of classes the network is trained on +* ```n_batch``` 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). +* ```conf_thresh``` confidence threshold for the detector. Only bounding boxes with threshold greater than conf-thresh will be displayed. +* ```show``` if set to 0 the demo will not show the visualization (if n-batches ==1) +* ```save``` if set to 1 the demo will save the video of the demo into result.mp4 (if n-batches ==1) N.B. By default it is used FP32 inference @@ -61,7 +60,8 @@ To run the demo with FP16 inference follow these steps (example with yolov3): export TKDNN_MODE=FP16 # set the half floating point optimization rm yolo3_fp16.rt # be sure to delete(or move) old tensorRT files ./test_yolo3 # run the yolo test (is slow) -./demo yolo3_fp16.rt ../demo/yolo_test.mp4 y +# set net: yolo3_fp16.rt in the config-file +./demo ``` N.B. Using FP16 inference will lead to some errors in the results (first or second decimal). @@ -86,7 +86,8 @@ export TKDNN_CALIB_LABEL_PATH=../demo/COCO_val2017/all_labels.txt export TKDNN_CALIB_IMG_PATH=../demo/COCO_val2017/all_images.txt rm yolo3_int8.rt # be sure to delete(or move) old tensorRT files ./test_yolo3 # run the yolo test (is slow) -./demo yolo3_int8.rt ../demo/yolo_test.mp4 y +# set net: yolo3_int8.rt in the config-file +./demo ``` N.B. diff --git a/include/tkDNN/SegmentationNN.h b/include/tkDNN/SegmentationNN.h index b691cbc..b73ffa2 100644 --- a/include/tkDNN/SegmentationNN.h +++ b/include/tkDNN/SegmentationNN.h @@ -182,6 +182,7 @@ class SegmentationNN { 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)); + return true; } /** diff --git a/include/tkDNN/utils.h b/include/tkDNN/utils.h index 65375ba..4965a7a 100644 --- a/include/tkDNN/utils.h +++ b/include/tkDNN/utils.h @@ -21,6 +21,7 @@ #include #include +#include #define dnnType float @@ -137,4 +138,19 @@ static inline bool isCudaPointer(void *data) { cudaPointerAttributes attr; return cudaPointerGetAttributes(&attr, data) == 0; } + +inline YAML::Node YAMLloadConf(const std::string& conf_file) { + std::cerr<<"Loading YAML: "< +inline T YAMLgetConf(YAML::Node conf, std::string key, T defaultVal) { + T val = defaultVal; + if(conf && conf[key]) { + val = conf[key].as(); + } + return val; +} + #endif //UTILS_H diff --git a/scripts/checkExecTimes.py b/scripts/checkExecTimes.py new file mode 100644 index 0000000..c3e1b0b --- /dev/null +++ b/scripts/checkExecTimes.py @@ -0,0 +1,37 @@ +import sys +import pandas as pd + +if len(sys.argv) < 3: + print("Error: two csv files are needed, old first new second") + exit(1) + +old_perf_file = str(sys.argv[1]) +new_perf_file = str(sys.argv[2]) + +verbose = False +if len(sys.argv) == 4: + verbose = bool(sys.argv[3]) + +print("Comparing {} vs {}".format(old_perf_file, new_perf_file)) + +df_old = pd.read_csv (old_perf_file, sep=';', header=None, index_col=0) +df_new = pd.read_csv (new_perf_file, sep=';', header=None, index_col=0) + +for index, row in df_new.iterrows(): + if index in df_old.index: + if verbose: + print("New: ",row[1], row[2], row[3]) + print("Old: ",df_old.loc[index][1], df_old.loc[index][2], df_old.loc[index][3]) + + print(index, end=': ') + if abs(row[1] - df_old.loc[index][1]) < df_old.loc[index][1]*0.1: + print("similar performance") + elif (row[1] < df_old.loc[index][1]): + print('\x1b[3;30;42m' + 'faster' + '\x1b[0m') + elif (row[1] > df_old.loc[index][1]): + if row[1] > df_old.loc[index][1] + df_old.loc[index][1] * 0.5 : + print('\x1b[3;30;41m' + 'WAY SLOWER' + '\x1b[0m') + else: + print('\x1b[3;30;41m' + 'slower' + '\x1b[0m') + + diff --git a/src/CenterTrack.cpp b/src/CenterTrack.cpp index dc15823..991ff69 100644 --- a/src/CenterTrack.cpp +++ b/src/CenterTrack.cpp @@ -17,6 +17,8 @@ bool CenterTrack::init(const std::string& tensor_path, const int n_classes, cons init_pre_inf(); init_postprocessing(); init_visualization(n_classes); + + return true; } bool CenterTrack::init_preprocessing(){ @@ -59,6 +61,8 @@ bool CenterTrack::init_preprocessing(){ checkCuda( cudaMalloc(&input_d, sizeof(dnnType)*netRT->input_dim.tot() * nBatches)); checkCuda( cudaMalloc(&input_pre_inf_d, sizeof(dnnType)*dim.tot())); checkCuda( cudaMalloc(&d_ptrs, dim.tot() * sizeof(float)) ); + + return true; } bool CenterTrack::init_pre_inf(){ @@ -202,6 +206,8 @@ bool CenterTrack::init_postprocessing(){ trRes.resize(nBatches); countTr.resize(nBatches, 0); trackId.resize(nBatches, 0); + + return true; } bool CenterTrack::init_visualization(const int n_classes){ @@ -274,6 +280,8 @@ bool CenterTrack::init_visualization(const int n_classes){ faceId.push_back({3,0,4,7}); faceId.push_back({2,3,7,6}); // ([[0,1,5,4], [1,2,6, 5], [2,3,7,6], [3,0,4,7]]); + + return true; } void CenterTrack::_get_additional_inputs(){ @@ -308,7 +316,7 @@ void CenterTrack::preprocess(cv::Mat &frame, const int bi){ } float c[] = {new_width / 2.0f, new_height /2.0f}; - float s[] = {dim.w, dim.h}; + float s[] = {float(dim.w), float(dim.h)}; // float s = new_width >= new_height ? new_width : new_height; // ----------- get_affine_transform // rot_rad = pi * 0 / 100 --> 0 diff --git a/src/CenternetDetection.cpp b/src/CenternetDetection.cpp index 46757f4..dce1455 100644 --- a/src/CenternetDetection.cpp +++ b/src/CenternetDetection.cpp @@ -119,6 +119,7 @@ bool CenternetDetection::init(const std::string& tensor_path, const int n_classe dst2.at(2,0)=dst2.at(1,0) + (-dst2.at(0,1)+dst2.at(1,1) ); dst2.at(2,1)=dst2.at(1,1) + (dst2.at(0,0)-dst2.at(1,0) ); + return true; } diff --git a/src/CenternetDetection3D.cpp b/src/CenternetDetection3D.cpp index 653624b..14f674d 100644 --- a/src/CenternetDetection3D.cpp +++ b/src/CenternetDetection3D.cpp @@ -167,6 +167,8 @@ bool CenternetDetection3D::init(const std::string& tensor_path, const int n_clas faceId.push_back({2,3,7,6}); faceId.push_back({3,0,4,7}); // ([[0,1,5,4], [1,2,6, 5], [2,3,7,6], [3,0,4,7]]); + + return true; } void CenternetDetection3D::preprocess(cv::Mat &frame, const int bi){ diff --git a/src/MobilenetDetection.cpp b/src/MobilenetDetection.cpp index 3c54e28..13b836f 100644 --- a/src/MobilenetDetection.cpp +++ b/src/MobilenetDetection.cpp @@ -198,7 +198,7 @@ bool MobilenetDetection::init(const std::string& tensor_path, const int n_classe "bottle" , "wine glass" , "cup" , "fork" , "knife" , "spoon" , "bowl" , "banana" , "apple" , "sandwich" , "orange" , "broccoli" , "carrot" , "hot dog" , "pizza" , "donut" , "cake" , "chair" , "sofa" , "pottedplant" , "bed" , "diningtable" , - "toilet" , "tvmonitor" , "laptop" , "mouse" , "remote" , "keyboard" , + "toilet" , "tvmonitor" , "laptop" , "mouse" , "remote" , "keyboard" , "cell phone" , "microwave" , "oven" , "toaster" , "sink" , "refrigerator" , "book" , "clock" , "vase" , "scissors" , "teddy bear" , "hair drier" , "toothbrush"}; classesNames = std::vector(classes_names_, std::end(classes_names_)); @@ -207,7 +207,7 @@ bool MobilenetDetection::init(const std::string& tensor_path, const int n_classe else{ FatalError("Number of classes not supported for mobilenet"); } - return 1; + return true; } void MobilenetDetection::preprocess(cv::Mat &frame, const int bi){ diff --git a/tests/darknet/cfg/yolo4-csp_crowd.cfg b/tests/darknet/cfg/yolo4-csp_crowd.cfg new file mode 100644 index 0000000..4c50f4d --- /dev/null +++ b/tests/darknet/cfg/yolo4-csp_crowd.cfg @@ -0,0 +1,1279 @@ +[net] +# Testing +#batch=1 +#subdivisions=1 +# Training +batch=64 +subdivisions=16 +width=512 +height=512 +channels=3 +momentum=0.949 +decay=0.0005 +angle=0 +saturation = 1.5 +exposure = 1.5 +hue=.1 + +learning_rate=0.001 +burn_in=1000 +max_batches = 8000 +policy=steps +steps=6400,7200 +scales=.1,.1 + +mosaic=1 + +letter_box=1 + +ema_alpha=0.9998 + +#optimized_memory=1 + +#23:104x104 54:52x52 85:26x26 104:13x13 for 416 + + + +[convolutional] +batch_normalize=1 +filters=32 +size=3 +stride=1 +pad=1 +activation=mish + +# Downsample + +[convolutional] +batch_normalize=1 +filters=64 +size=3 +stride=2 +pad=1 +activation=mish + +#[convolutional] +#batch_normalize=1 +#filters=64 +#size=1 +#stride=1 +#pad=1 +#activation=mish + +#[route] +#layers = -2 + +#[convolutional] +#batch_normalize=1 +#filters=64 +#size=1 +#stride=1 +#pad=1 +#activation=mish + +[convolutional] +batch_normalize=1 +filters=32 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=64 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +#[convolutional] +#batch_normalize=1 +#filters=64 +#size=1 +#stride=1 +#pad=1 +#activation=mish + +#[route] +#layers = -1,-7 + +#[convolutional] +#batch_normalize=1 +#filters=64 +#size=1 +#stride=1 +#pad=1 +#activation=mish + +# Downsample + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=2 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=64 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=64 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=64 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=64 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=64 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=64 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=64 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1,-10 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +# Downsample + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=2 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1,-28 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +# Downsample + +[convolutional] +batch_normalize=1 +filters=512 +size=3 +stride=2 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1,-28 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +# Downsample + +[convolutional] +batch_normalize=1 +filters=1024 +size=3 +stride=2 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=3 +stride=1 +pad=1 +activation=mish + +[shortcut] +from=-3 +activation=linear + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1,-16 + +[convolutional] +batch_normalize=1 +filters=1024 +size=1 +stride=1 +pad=1 +activation=mish + +########################## + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=512 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +### SPP ### +[maxpool] +stride=1 +size=5 + +[route] +layers=-2 + +[maxpool] +stride=1 +size=9 + +[route] +layers=-4 + +[maxpool] +stride=1 +size=13 + +[route] +layers=-1,-3,-5,-6 +### End SPP ### + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=512 +activation=mish + +[route] +layers = -1, -13 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[upsample] +stride=2 + +[route] +layers = 79 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1, -3 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=256 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=256 +activation=mish + +[route] +layers = -1, -6 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[upsample] +stride=2 + +[route] +layers = 48 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -1, -3 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=128 +activation=mish + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=128 +activation=mish + +[route] +layers = -1, -6 + +[convolutional] +batch_normalize=1 +filters=128 +size=1 +stride=1 +pad=1 +activation=mish + +########################## + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=256 +activation=mish + +[convolutional] +size=1 +stride=1 +pad=1 +filters=21 +activation=logistic + + +[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=2 +num=9 +jitter=.1 +scale_x_y = 2.0 +objectness_smooth=0 +ignore_thresh = .7 +truth_thresh = 1 +#random=1 +resize=1.5 +iou_thresh=0.2 +iou_normalizer=0.05 +cls_normalizer=0.5 +obj_normalizer=4.0 +iou_loss=ciou +nms_kind=diounms +beta_nms=0.6 +new_coords=1 +max_delta=5 + +[route] +layers = -4 + +[convolutional] +batch_normalize=1 +size=3 +stride=2 +pad=1 +filters=256 +activation=mish + +[route] +layers = -1, -20 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=256 +activation=mish + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=256 +activation=mish + +[route] +layers = -1,-6 + +[convolutional] +batch_normalize=1 +filters=256 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=512 +activation=mish + +[convolutional] +size=1 +stride=1 +pad=1 +filters=21 +activation=logistic + + +[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=2 +num=9 +jitter=.1 +scale_x_y = 2.0 +objectness_smooth=1 +ignore_thresh = .7 +truth_thresh = 1 +#random=1 +resize=1.5 +iou_thresh=0.2 +iou_normalizer=0.05 +cls_normalizer=0.5 +obj_normalizer=1.0 +iou_loss=ciou +nms_kind=diounms +beta_nms=0.6 +new_coords=1 +max_delta=5 + +[route] +layers = -4 + +[convolutional] +batch_normalize=1 +size=3 +stride=2 +pad=1 +filters=512 +activation=mish + +[route] +layers = -1, -49 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[route] +layers = -2 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=512 +activation=mish + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=512 +activation=mish + +[route] +layers = -1,-6 + +[convolutional] +batch_normalize=1 +filters=512 +size=1 +stride=1 +pad=1 +activation=mish + +[convolutional] +batch_normalize=1 +size=3 +stride=1 +pad=1 +filters=1024 +activation=mish + +[convolutional] +size=1 +stride=1 +pad=1 +filters=21 +activation=logistic + + +[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=2 +num=9 +jitter=.1 +scale_x_y = 2.0 +objectness_smooth=1 +ignore_thresh = .7 +truth_thresh = 1 +#random=1 +resize=1.5 +iou_thresh=0.2 +iou_normalizer=0.05 +cls_normalizer=0.5 +obj_normalizer=0.4 +iou_loss=ciou +nms_kind=diounms +beta_nms=0.6 +new_coords=1 +max_delta=2 diff --git a/tests/darknet/names/crowdhuman.names b/tests/darknet/names/crowdhuman.names new file mode 100644 index 0000000..5ba1275 --- /dev/null +++ b/tests/darknet/names/crowdhuman.names @@ -0,0 +1,2 @@ +person +head \ No newline at end of file diff --git a/tests/darknet/yolo4-csp_crowd.cpp b/tests/darknet/yolo4-csp_crowd.cpp new file mode 100644 index 0000000..b8f0e6b --- /dev/null +++ b/tests/darknet/yolo4-csp_crowd.cpp @@ -0,0 +1,34 @@ +#include +#include +#include "tkdnn.h" +#include "test.h" +#include "DarknetParser.h" + +int main() { + std::string bin_path = "yolo4-csp_crowd"; + std::vector input_bins = { + bin_path + "/layers/input.bin" + }; + std::vector 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_crowd.cfg"; + std::string name_path = std::string(TKDNN_PATH) + "/tests/darknet/names/crowdhuman.names"; + downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/RKWfWNmWXfJigsK/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; +} \ No newline at end of file diff --git a/tests/shelfnet/shelfnet_coco.cpp b/tests/shelfnet/shelfnet_coco.cpp new file mode 100644 index 0000000..276a09d --- /dev/null +++ b/tests/shelfnet/shelfnet_coco.cpp @@ -0,0 +1,295 @@ +#include +#include +#include + +#include "tkdnn.h" +#include "NetworkViz.h" + + +const char *input_bin = "shelfnet_coco/debug/input.bin"; + +const char *backbone[] = { + "shelfnet_coco/layers/backbone-conv1.bin", + "shelfnet_coco/layers/backbone-layer1-0-conv1.bin", + "shelfnet_coco/layers/backbone-layer1-0-conv2.bin", + "shelfnet_coco/layers/backbone-layer1-1-conv1.bin", + "shelfnet_coco/layers/backbone-layer1-1-conv2.bin", + "shelfnet_coco/layers/backbone-layer2-0-conv1.bin", + "shelfnet_coco/layers/backbone-layer2-0-conv2.bin", + "shelfnet_coco/layers/backbone-layer2-0-downsample-0.bin", + "shelfnet_coco/layers/backbone-layer2-1-conv1.bin", + "shelfnet_coco/layers/backbone-layer2-1-conv2.bin", + "shelfnet_coco/layers/backbone-layer3-0-conv1.bin", + "shelfnet_coco/layers/backbone-layer3-0-conv2.bin", + "shelfnet_coco/layers/backbone-layer3-0-downsample-0.bin", + "shelfnet_coco/layers/backbone-layer3-1-conv1.bin", + "shelfnet_coco/layers/backbone-layer3-1-conv2.bin", + "shelfnet_coco/layers/backbone-layer4-0-conv1.bin", + "shelfnet_coco/layers/backbone-layer4-0-conv2.bin", + "shelfnet_coco/layers/backbone-layer4-0-downsample-0.bin", + "shelfnet_coco/layers/backbone-layer4-1-conv1.bin", + "shelfnet_coco/layers/backbone-layer4-1-conv2.bin"}; + +const char *conv_out[] = { + "shelfnet_coco/layers/conv_out-conv-conv.bin", + "shelfnet_coco/layers/conv_out-conv_out.bin", + "shelfnet_coco/layers/conv_out16-conv-conv.bin", + "shelfnet_coco/layers/conv_out16-conv_out.bin", + "shelfnet_coco/layers/conv_out32-conv-conv.bin", + "shelfnet_coco/layers/conv_out32-conv_out.bin" + }; + +const char *decoder[] = { + "shelfnet_coco/layers/decoder-bottom-conv1.bin", + "shelfnet_coco/layers/decoder-bottom-conv12.bin", + "shelfnet_coco/layers/decoder-up_conv_list-0-conv-conv.bin", + "shelfnet_coco/layers/decoder-up_conv_list-0-conv_atten.bin", + "shelfnet_coco/layers/decoder-up_dense_list-0-conv.bin", + "shelfnet_coco/layers/decoder-up_conv_list-1-conv-conv.bin", + "shelfnet_coco/layers/decoder-up_conv_list-1-conv_atten.bin", + "shelfnet_coco/layers/decoder-up_dense_list-1-conv.bin" + }; + + +const char *ladder[] = { + "shelfnet_coco/layers/ladder-inconv-conv1.bin", + "shelfnet_coco/layers/ladder-inconv-conv12.bin", + "shelfnet_coco/layers/ladder-down_module_list-0-conv1.bin", + "shelfnet_coco/layers/ladder-down_module_list-0-conv12.bin", + "shelfnet_coco/layers/ladder-down_conv_list-0.bin", + + "shelfnet_coco/layers/ladder-down_module_list-1-conv1.bin", + "shelfnet_coco/layers/ladder-down_module_list-1-conv12.bin", + "shelfnet_coco/layers/ladder-down_conv_list-1.bin", + + "shelfnet_coco/layers/ladder-bottom-conv1.bin", + "shelfnet_coco/layers/ladder-bottom-conv12.bin", + + + + "shelfnet_coco/layers/ladder-up_conv_list-0-conv-conv.bin", + "shelfnet_coco/layers/ladder-up_conv_list-0-conv_atten.bin", + "shelfnet_coco/layers/ladder-up_dense_list-0-conv.bin", + + + "shelfnet_coco/layers/ladder-up_conv_list-1-conv-conv.bin", + "shelfnet_coco/layers/ladder-up_conv_list-1-conv_atten.bin", + "shelfnet_coco/layers/ladder-up_dense_list-1-conv.bin"}; + +const char *trans[] = { + "shelfnet_coco/layers/trans1-conv.bin", + "shelfnet_coco/layers/trans2-conv.bin", + "shelfnet_coco/layers/trans3-conv.bin"}; +int main() +{ + + downloadWeightsifDoNotExist(input_bin, "shelfnet_coco", "https://cloud.hipert.unimore.it/s/KfQ9fGJQsgzNbiW/download"); + + int classes = 183; + + // 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 features; + for(int i=0;i<3;++i){ + int out_channel = pow(2,7+i); + std::cout< 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<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 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_coco/debug/softmax.bin"; + + // Load input + dnnType *data; + dnnType *input_h; + readBinaryFile(input_bin, dim.tot(), &input_h, &data); + std::cout<<"Input:"<