From 2151aab63f0b7661f75f0b2dda532d23068fc33e Mon Sep 17 00:00:00 2001
From: enazoe <hzr921120@gmail.com>
Date: Mon, 31 Aug 2020 13:27:06 +0800
Subject: [PATCH] yolov5-2.0

---
 configs/yolov5-3.0/yolov5s.cfg | 4 ++--
 modules/hardswish.cu           | 9 +++++++--
 modules/trt_utils.h            | 4 ++--
 samples/sample_detector.cpp    | 9 ++++-----
 4 files changed, 15 insertions(+), 11 deletions(-)

diff --git a/configs/yolov5-3.0/yolov5s.cfg b/configs/yolov5-3.0/yolov5s.cfg
index 999635a..bd47f57 100644
--- a/configs/yolov5-3.0/yolov5s.cfg
+++ b/configs/yolov5-3.0/yolov5s.cfg
@@ -1,7 +1,7 @@
 
 [net]
-width=672
-height=672
+width=320
+height=320
 channels=3
 nc=80
 depth_multiple=0.33
diff --git a/modules/hardswish.cu b/modules/hardswish.cu
index 088942f..fe5dc31 100644
--- a/modules/hardswish.cu
+++ b/modules/hardswish.cu
@@ -27,6 +27,7 @@ namespace nvinfer1
 		cudaDeviceProp prop;
 		cudaGetDeviceProperties(&prop, 0);
 		_n_max_thread_pre_block = prop.maxThreadsPerBlock;
+	//	printf("Hardswish():%d\n", _n_max_thread_pre_block);
 	}
 
 	Hardswish::Hardswish(const void* data, size_t length)
@@ -34,6 +35,7 @@ namespace nvinfer1
 		const char *d = reinterpret_cast<const char*>(data), *a = d;
 		r(d, _n_max_thread_pre_block);
 		r(d, _n_output_size);
+//		printf("r:threads:%d,size:%d\n", _n_max_thread_pre_block, _n_output_size);
 		assert(d == a + length);
 	}
 
@@ -66,6 +68,7 @@ namespace nvinfer1
 		cudaStream_t stream_)
 	{
 		int n_data_size = n_batch_size_ * n_output_size_;
+//		printf("cuda_hardswish_layer:%d,size:%d\n", n_batch_size_, n_output_size_);
 		kernel_hardswish << <(n_data_size + threads_ -1)/threads_, threads_ >> >(
 				reinterpret_cast<const float*>(input_),
 				reinterpret_cast<float*>(output_),
@@ -76,7 +79,7 @@ namespace nvinfer1
 	int Hardswish::enqueue(int batchSize, const void* const* inputs, void** outputs, void* workspace,
 		cudaStream_t stream)
 	{
-		//printf("batch_size:%d,output_size:%d,threads:%d\n", batchSize, _n_output_size, _n_max_thread_pre_block);
+//		printf("batch_size:%d,output_size:%d,threads:%d\n", batchSize, _n_output_size, _n_max_thread_pre_block);
 		NV_CUDA_CHECK(cuda_hardswish_layer(inputs[0], outputs[0], batchSize, _n_output_size , _n_max_thread_pre_block,stream));
 		return 0;
 	}
@@ -91,6 +94,7 @@ namespace nvinfer1
 		char *d = static_cast<char*>(buffer), *a = d;
 		w(d, _n_max_thread_pre_block);
 		w(d, _n_output_size);
+//		printf("serialize:%d,%d\n", _n_max_thread_pre_block, _n_output_size);
 		assert(d == a + getSerializationSize());
 	}
 
@@ -98,7 +102,7 @@ namespace nvinfer1
 	{
 		
 		_n_output_size = in->dims.d[0] * in->dims.d[1] * in->dims.d[2];
-	//	printf("output_size:%d,threads:%d\n", _n_output_size, _n_max_thread_pre_block);
+//		printf("configurePlugin:%d,%d,%d\n", in->dims.d[0], in->dims.d[1], in->dims.d[2]);
 	}
 	IPluginV2IOExt* Hardswish::clone() const
 	{
@@ -106,6 +110,7 @@ namespace nvinfer1
 		p->setPluginNamespace(_s_plugin_namespace.c_str());
 		p->_n_max_thread_pre_block = _n_max_thread_pre_block;
 		p->_n_output_size = _n_output_size;
+//		printf("clone:%d,%d\n", _n_max_thread_pre_block, _n_output_size);
 		return p;
 	}
 
diff --git a/modules/trt_utils.h b/modules/trt_utils.h
index 7fd0959..14d22c4 100644
--- a/modules/trt_utils.h
+++ b/modules/trt_utils.h
@@ -215,11 +215,11 @@ nvinfer1::ILayer * layer_conv_bn_act(
 	const int group_ =1,
 	const bool b_padding_ = true,
 	const bool b_bn_ = true,
-	const std::string s_act_ = "hardswish");
+	const std::string s_act_ = "leaky");
 
 nvinfer1::ILayer * layer_act(nvinfer1::ITensor* input_,
 	nvinfer1::INetworkDefinition* network_,
-	const std::string s_act_ = "hardswish");
+	const std::string s_act_ = "leaky");
 
 nvinfer1::ILayer * layer_bottleneck_csp(
 	std::string s_model_name_,
diff --git a/samples/sample_detector.cpp b/samples/sample_detector.cpp
index 362937a..3e2f332 100644
--- a/samples/sample_detector.cpp
+++ b/samples/sample_detector.cpp
@@ -39,11 +39,10 @@ int main()
 	Config config_v5;
 	config_v5.net_type = YOLOV5;
 	config_v5.detect_thresh = 0.5;
-	config_v5.file_model_cfg = "../configs/yolov5-3.0/yolov5s.cfg";
-	config_v5.file_model_weights = "../configs/yolov5-3.0/yolov5s.weights";
-	config_v5.calibration_image_list_file_txt = "../configs/calibration_images.txt";
+	config_v5.file_model_cfg = "../configs/yolov5-2.0/yolov5s.cfg";
+	config_v5.file_model_weights = "../configs/yolov5-2.0/yolov5s.weights";
 	config_v5.inference_precison = FP32;
-	config_v5.n_max_batch = 1;
+	//config_v5.n_max_batch = 2;
 
 	cv::Mat image0 = cv::imread("../configs/dog.jpg", cv::IMREAD_UNCHANGED);
 	cv::Mat image1 = cv::imread("../configs/person.jpg", cv::IMREAD_UNCHANGED);
@@ -58,7 +57,7 @@ int main()
 		cv::Mat temp0 = image0.clone();
 		cv::Mat temp1 = image1.clone();
 		batch_img.push_back(temp0);
-		//batch_img.push_back(temp1);
+		batch_img.push_back(temp1);
 
 		//detect
 		timer.reset();