扒扒Tiny YoloV3的.weights文件经过TensorFlow处理后得到的Frozen Graph里面有什么

最新推荐文章于 2025-05-19 12:30:28 发布
最新推荐文章于 2025-05-19 12:30:28 发布
阅读量1.4k 收藏 3
点赞数
CC 4.0 BY-SA版权
分类专栏： 硬件平台 文章标签： YoloV3 OpenVINO TensorFlow 神经计算棒 树莓派
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.youkuaiyun.com/weixin_41665225/article/details/88669915

最近要把tiny yolov3的模型移植到NCS2，从.weights文件到IE model的转化相当繁琐（见另外一篇博文《在树莓派+Intel NCS2上跑YoloV3 Tiny》），并且TensorFlow的版本不对还报错，实在是讨厌。
一直没怎么用tensorflow，将来打算在darknet里面直接输出.pb文件（而不是.weights文件），今天先扒了扒tensorflow 1.9环境下生成的tiny_yolov3.pb文件。
附生成命令（Window10 + Python 3.6.7+ TensorFlow 1.9.0）：

python ../tensorflow-yolo-v3/convert_weights_pb.py --weights_file yolov3-tiny.weights \
       --class_names coco.names  --tiny True --size 416 --data_format NHWC \
       --output_graph yolov3-tiny.pb

使用ProtoBuf 3.0.0生成*.pb.h和*.pb.cc，如图
在这里插入图片描述
解析程序代码main.cpp（Compiled under Visual Studio 2015）：


/*****************************************************************
main.cpp
******************************************************************/
#include <iostream>
#include <fstream>
#include <google/protobuf/util/time_util.h>
#include "tensorflow/core/framework/graph.pb.h"

using namespace std;

using google::protobuf::util::TimeUtil;

using namespace tensorflow;

constexpr char* pb_path = "E:\\AI\\OpenVINO\\COCO\\tiny_yolov3.pb";

static const char* type_to_str(tensorflow::DataType t) {
   
   

	switch (t)
	{
   
   
	case tensorflow::DT_INVALID:
		return "Invalid";
	case tensorflow::DT_FLOAT:
		return "Float";
	case tensorflow::DT_DOUBLE:
		return "Double";
	case tensorflow::DT_INT32:
		return "Int32";
	case tensorflow::DT_UINT8:
		return "Uint8";
	case tensorflow::DT_INT16:
		return "Uint16";
	case tensorflow::DT_INT8:
		return "Int8";
	case tensorflow::DT_STRING:
		return "String";
	case tensorflow::DT_COMPLEX64:
		return "Complex64";
	case tensorflow::DT_INT64:
		return "Int64";
	case tensorflow::DT_BOOL:
		return "Bool";
	case tensorflow::DT_QINT8:
		return "Qint8";
	case tensorflow::DT_QUINT8:
		return "QUint8";
	case tensorflow::DT_QINT32:
		return "Qint32";
	case tensorflow::DT_BFLOAT16:
		return "BFloat16";
	case tensorflow::DT_QINT16:
		return "Qint16";
	case tensorflow::DT_QUINT16:
		return "Quint16";
	case tensorflow::DT_UINT16:
		return "Uint16";
	case tensorflow::DT_COMPLEX128:
		return "Complex128";
	case tensorflow::DT_HALF:
		return "Half";
	case tensorflow::DT_RESOURCE:
		return "Resource";
	case tensorflow::DT_VARIANT:
		return "Variant";
	case tensorflow::DT_UINT32:
		return "Uint32";
	case tensorflow::DT_UINT64:
		return "Uint64";
	case tensorflow::DT_FLOAT_REF:
		return "FloatRef";
	case tensorflow::DT_DOUBLE_REF:
		return "DoubleRef";
	case tensorflow::DT_INT32_REF:
		return "Int32Ref";
	case tensorflow::DT_UINT8_REF:
		return "Uint8Ref";
	case tensorflow::DT_INT16_REF:
		return "Int16Ref";
	case tensorflow::DT_INT8_REF:
		return "Int8Ref";
	case tensorflow::DT_STRING_REF:
		return "StringRef";
	case tensorflow::DT_COMPLEX64_REF:
		return "Complex64Ref";
	case tensorflow::DT_INT64_REF:
		return "Int64Ref";
	case tensorflow::DT_BOOL_REF:
		return "BoolRef";
	case tensorflow::DT_QINT8_REF:
		return "Qint8Ref";
	case tensorflow::DT_QUINT8_REF:
		return "QUint8Ref";
	case tensorflow::DT_QINT32_REF:
		return "QUint32Ref";
	case tensorflow::DT_BFLOAT16_REF:
		return "DT_BFLOAT16_REF";
	case tensorflow::DT_QINT16_REF:
		return "DT_QINT16_REF";
	case tensorflow::DT_QUINT16_REF:
		return "DT_QUINT16_REF";
	case tensorflow::DT_UINT16_REF:
		return "DT_UINT16_REF";
	case tensorflow::DT_COMPLEX128_REF:
		return "DT_COMPLEX128_REF";
	case tensorflow::DT_HALF_REF:
		return "DT_HALF_REF";
	case tensorflow::DT_RESOURCE_REF:
		return "DT_RESOURCE_REF";
	case tensorflow::DT_VARIANT_REF:
		return "DT_VARIANT_REF";
	case tensorflow::DT_UINT32_REF:
		return "DT_UINT32_REF";
	case tensorflow::DT_UINT64_REF:
		return "DT_UINT64_REF"; 
	default:
		return "Illegal";
	}
	return "";
}
void dump_attr(ostream& of, const AttrValue& v) {
   
   
	switch (v.value_case())
	{
   
   
	case tensorflow::AttrValue::kS:
		of << " type=\"String\" value=\"" << v.s() << "\" />" <<endl;
		break;
	case tensorflow::AttrValue::kI:
		of << " type=\"Integer\" value=\"" << v.i() << "\" />" << endl;
		break;
	case tensorflow::AttrValue::kF:
		of << " type=\"Float\" value=\"" << v.f() << "\" />" << endl;
		break;
	case tensorflow::AttrValue::kB:
		of << " type=\"Bool\" value=\"" << (v.b() ? "True": "False" ) << "\" />" << endl;
		break;
	case tensorflow::AttrValue::kType:
		of << " type=\"Type\" value=\"" << type_to_str(v.type()) << "\" />" << endl;
		break;
	case tensorflow::AttrValue::kShape:
	{
   
   
		const tensorflow::TensorShapeProto& sh = v.shape();
		of << " type=\"Shape\" dim-size=\"" << sh.dim_size() << "\" dims=\"[";
		for (int i = 0; i < sh.dim_size(); i++) {
   
   
			const tensorflow::TensorShapeProto_Dim& d = sh.dim(i);
			if ((i + 1) < sh.dim_size())
				of << d.size()<< ",";
			else
				of << d.size() << "]\" />" << endl;
		}
			 
		break;
	}
	case tensorflow::AttrValue::kTensor:
	{
   
   
		const tensorflow::TensorProto& t = v.tensor(); 
		const tensorflow::TensorShapeProto& sh = t.tensor_shape();
		size_t content_size = t.tensor_content().size();
		int dim_size = sh.dim_size();
		of << " type=\"Tensor\" type=\"" << type_to_str(t.dtype()) << "\"  dims=\"[";
		if(0 == dim_size)
			of << "]\"  size = \"" << content_size << "\" />" << endl;
		for (int i = 0; i < dim_size; i++) {
   
   
			const tensorflow::TensorShapeProto_Dim& d = sh.dim(i);
			if ((i + 1) < sh.dim_size())
				of << d.size() << ",";
			else
				of << d.size() << "]\"  bytes=\"" << content_size << "\" />" << endl;
		}
		 
		
		break;
	}
	case tensorflow::AttrValue::kList:
	{
   
   
		const tensorflow::AttrValue_ListValue& lv = v.list();
		of << " type=\"List\" ";
		if (lv.i_size() > 0) {
   
   
			of << "int-size=\"" << lv.i_size() << "\" int-vals=\"";
			for (int i = 0; i < lv.i_size(); i++) {
   
   
				if ((i + 1) < lv.i_size())
					of << lv.i(i) << ",";
				else
					of << lv.i(i) << "\" ";
			}
		}
		if (lv.s_size() > 0) {
   
   
			of << "string-size=\"" << lv.s_size() << "\" strings=\"";
			for (int i = 0; i < lv.s_size(); i++) {
   
   
				if ((i + 1) < lv.s_size())
					of << lv.s(i) << ",";
				else
					of << lv.s(i) << "\" "; 
			}
		} 
		if (lv.f_size() > 0) {
   
   
			of << "float-size=\"" << lv.s_size() << "\" float-vals=\"";
			for (int i = 0; i < lv.f_size(); i++) {
   
   
				if ((i + 1) < lv.f_size())
					of << lv.f(i) << ",";
				else
					of << lv.f(i) << "\" ";
			}
		}
		of << "/>" << endl; 
		break;
	}
	case tensorflow::AttrValue::kFunc:
		of << " type=\"Function\" name=\"" << v.func().name() << "\" />" << endl;
		break;
	case tensorflow::AttrValue::kPlaceholder:
		of << " type=\"PlaceHolder\" name=\"" << v.placeholder() << "\" />" << endl;
		break; 		 
	default:
		of << " type-value=\"" << (int)v.value_case() << "\" />" << endl;
		break;
	}
}

int main(int argc, char* argv[]) {
   
   

	// Verify that the version of the library that we linked against is
	// compatible with the version of the headers we compiled against.
	GOOGLE_PROTOBUF_VERIFY_VERSION;

	GraphDef graph;
	
	fstream input(pb_path, ios::in | ios::binary);
	if (!graph.ParseFromIstream(&input)) {
   
   
		cerr << "Failed to parse graph." << endl;
		return -1;
	}
	
	cout << "parse graph OK. node size:" << graph.node_size() << endl;
	fstream of("output.xml", ios::out | ios::trunc);
	if (!of.is_open()) {
   
   
		cerr << "open output.xml failed! " << endl;
		return -1;
	} 
	of << "<graphdef from=\"" << pb_path << "\" >" << endl ; 

	for (int i = 0; i < graph.node_size(); i++) {
   
   
		const NodeDef& d = graph.node(i);
		
		of << "\t<nodedef index=\"" << i << "\" name=\"" << d.name() << "\" op=\"" << d.op() <<  "\">" << endl;
		if (d.input_size() == 0)
			of << "\t\t<inputs />" << endl;
		else {
   
   
			of << "\t\t<inputs>" << endl;
			for (int j = 0; j < d.input_size(); j++) {
   
   
				of << "\t\t\t<input node=\"" << d.input(j) << "\" />" << endl;
			}
			of << "\t\t</inputs>" << endl;
		}
		const google::protobuf::Map<string, AttrValue>& m = d.attr();
		of << "\t\t<attributes>" << endl;
		for (google::protobuf::Map<string, AttrValue>::const_iterator it = m.begin(); it != m.end(); it++) {
   
   
			const AttrValue& v = it->second;
			of << "\t\t\t<attr name=\"" << it->first << "\" ";
			dump_attr(of,v); 
			
		}
		
		of << "\t\t</attributes>" << endl;
		of << "\t</nodedef>" << endl;
		
			
	}
	of << "</graphdef>" << endl;
	of.close();

	cout << "translated into .xml.\n";
	// Optional:  Delete all global objects allocated by libprotobuf.
	google::protobuf::ShutdownProtobufLibrary();

	return 0;
}

output.xml：

<graphdef from="E:\AI\OpenVINO\COCO\yolov3-tiny.pb" >
	<nodedef index="0" name="inputs" op="Placeholder">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="shape"  type="Shape" dim-size="4" dims="[-1,416,416,3]" />
		</attributes>
	</nodedef>
	<nodedef index="1" name="detector/truediv/y" op="Const">
		<inputs />
		<attributes>
			<attr name="value"  type="Tensor" type="Float"  dims="[]"  size = "0" />
			<attr name="dtype"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="2" name="detector/truediv" op="RealDiv">
		<inputs>
			<input node="inputs" />
			<input node="detector/truediv/y" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="3" name="detector/yolo-v3-tiny/Conv/weights" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[3,3,3,16]"  bytes="1728" />
		</attributes>
	</nodedef>
	<nodedef index="4" name="detector/yolo-v3-tiny/Conv/weights/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/weights" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv/weights" />
		</attributes>
	</nodedef>
	<nodedef index="5" name="detector/yolo-v3-tiny/Conv/Conv2D" op="Conv2D">
		<inputs>
			<input node="detector/truediv" />
			<input node="detector/yolo-v3-tiny/Conv/weights/read" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="strides"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="padding"  type="String" value="SAME" />
			<attr name="dilations"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="use_cudnn_on_gpu"  type="Bool" value="True" />
		</attributes>
	</nodedef>
	<nodedef index="6" name="detector/yolo-v3-tiny/Conv/BatchNorm/gamma" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[16]"  bytes="64" />
		</attributes>
	</nodedef>
	<nodedef index="7" name="detector/yolo-v3-tiny/Conv/BatchNorm/gamma/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/gamma" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv/BatchNorm/gamma" />
		</attributes>
	</nodedef>
	<nodedef index="8" name="detector/yolo-v3-tiny/Conv/BatchNorm/beta" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[16]"  bytes="64" />
		</attributes>
	</nodedef>
	<nodedef index="9" name="detector/yolo-v3-tiny/Conv/BatchNorm/beta/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/beta" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv/BatchNorm/beta" />
		</attributes>
	</nodedef>
	<nodedef index="10" name="detector/yolo-v3-tiny/Conv/BatchNorm/moving_mean" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[16]"  bytes="64" />
		</attributes>
	</nodedef>
	<nodedef index="11" name="detector/yolo-v3-tiny/Conv/BatchNorm/moving_mean/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/moving_mean" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv/BatchNorm/moving_mean" />
		</attributes>
	</nodedef>
	<nodedef index="12" name="detector/yolo-v3-tiny/Conv/BatchNorm/moving_variance" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[16]"  bytes="64" />
		</attributes>
	</nodedef>
	<nodedef index="13" name="detector/yolo-v3-tiny/Conv/BatchNorm/moving_variance/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/moving_variance" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv/BatchNorm/moving_variance" />
		</attributes>
	</nodedef>
	<nodedef index="14" name="detector/yolo-v3-tiny/Conv/BatchNorm/FusedBatchNorm" op="FusedBatchNorm">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/Conv2D" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/gamma/read" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/beta/read" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/moving_mean/read" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/moving_variance/read" />
		</inputs>
		<attributes>
			<attr name="is_training"  type="Bool" value="False" />
			<attr name="T"  type="Type" value="Float" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="epsilon"  type="Float" value="1.001e-05" />
		</attributes>
	</nodedef>
	<nodedef index="15" name="detector/yolo-v3-tiny/Conv/LeakyRelu/alpha" op="Const">
		<inputs />
		<attributes>
			<attr name="value"  type="Tensor" type="Float"  dims="[]"  size = "0" />
			<attr name="dtype"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="16" name="detector/yolo-v3-tiny/Conv/LeakyRelu/mul" op="Mul">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/LeakyRelu/alpha" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/FusedBatchNorm" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="17" name="detector/yolo-v3-tiny/Conv/LeakyRelu" op="Maximum">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/LeakyRelu/mul" />
			<input node="detector/yolo-v3-tiny/Conv/BatchNorm/FusedBatchNorm" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="18" name="detector/yolo-v3-tiny/pool2/MaxPool" op="MaxPool">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv/LeakyRelu" />
		</inputs>
		<attributes>
			<attr name="strides"  type="List" int-size="4" int-vals="1,2,2,1" />
			<attr name="T"  type="Type" value="Float" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="padding"  type="String" value="VALID" />
			<attr name="ksize"  type="List" int-size="4" int-vals="1,2,2,1" />
		</attributes>
	</nodedef>
	<nodedef index="19" name="detector/yolo-v3-tiny/Conv_1/weights" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[3,3,16,32]"  bytes="18432" />
		</attributes>
	</nodedef>
	<nodedef index="20" name="detector/yolo-v3-tiny/Conv_1/weights/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/weights" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_1/weights" />
		</attributes>
	</nodedef>
	<nodedef index="21" name="detector/yolo-v3-tiny/Conv_1/Conv2D" op="Conv2D">
		<inputs>
			<input node="detector/yolo-v3-tiny/pool2/MaxPool" />
			<input node="detector/yolo-v3-tiny/Conv_1/weights/read" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="strides"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="padding"  type="String" value="SAME" />
			<attr name="dilations"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="use_cudnn_on_gpu"  type="Bool" value="True" />
		</attributes>
	</nodedef>
	<nodedef index="22" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/gamma" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[32]"  bytes="128" />
		</attributes>
	</nodedef>
	<nodedef index="23" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/gamma/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/gamma" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_1/BatchNorm/gamma" />
		</attributes>
	</nodedef>
	<nodedef index="24" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/beta" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[32]"  bytes="128" />
		</attributes>
	</nodedef>
	<nodedef index="25" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/beta/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/beta" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_1/BatchNorm/beta" />
		</attributes>
	</nodedef>
	<nodedef index="26" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_mean" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[32]"  bytes="128" />
		</attributes>
	</nodedef>
	<nodedef index="27" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_mean/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_mean" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_mean" />
		</attributes>
	</nodedef>
	<nodedef index="28" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_variance" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[32]"  bytes="128" />
		</attributes>
	</nodedef>
	<nodedef index="29" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_variance/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_variance" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_variance" />
		</attributes>
	</nodedef>
	<nodedef index="30" name="detector/yolo-v3-tiny/Conv_1/BatchNorm/FusedBatchNorm" op="FusedBatchNorm">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/Conv2D" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/gamma/read" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/beta/read" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_mean/read" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/moving_variance/read" />
		</inputs>
		<attributes>
			<attr name="is_training"  type="Bool" value="False" />
			<attr name="T"  type="Type" value="Float" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="epsilon"  type="Float" value="1.001e-05" />
		</attributes>
	</nodedef>
	<nodedef index="31" name="detector/yolo-v3-tiny/Conv_1/LeakyRelu/alpha" op="Const">
		<inputs />
		<attributes>
			<attr name="value"  type="Tensor" type="Float"  dims="[]"  size = "0" />
			<attr name="dtype"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="32" name="detector/yolo-v3-tiny/Conv_1/LeakyRelu/mul" op="Mul">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/LeakyRelu/alpha" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/FusedBatchNorm" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="33" name="detector/yolo-v3-tiny/Conv_1/LeakyRelu" op="Maximum">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/LeakyRelu/mul" />
			<input node="detector/yolo-v3-tiny/Conv_1/BatchNorm/FusedBatchNorm" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="34" name="detector/yolo-v3-tiny/pool2_1/MaxPool" op="MaxPool">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_1/LeakyRelu" />
		</inputs>
		<attributes>
			<attr name="strides"  type="List" int-size="4" int-vals="1,2,2,1" />
			<attr name="T"  type="Type" value="Float" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="padding"  type="String" value="VALID" />
			<attr name="ksize"  type="List" int-size="4" int-vals="1,2,2,1" />
		</attributes>
	</nodedef>
	<nodedef index="35" name="detector/yolo-v3-tiny/Conv_2/weights" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[3,3,32,64]"  bytes="73728" />
		</attributes>
	</nodedef>
	<nodedef index="36" name="detector/yolo-v3-tiny/Conv_2/weights/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/weights" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_2/weights" />
		</attributes>
	</nodedef>
	<nodedef index="37" name="detector/yolo-v3-tiny/Conv_2/Conv2D" op="Conv2D">
		<inputs>
			<input node="detector/yolo-v3-tiny/pool2_1/MaxPool" />
			<input node="detector/yolo-v3-tiny/Conv_2/weights/read" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="strides"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="padding"  type="String" value="SAME" />
			<attr name="dilations"  type="List" int-size="4" int-vals="1,1,1,1" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="use_cudnn_on_gpu"  type="Bool" value="True" />
		</attributes>
	</nodedef>
	<nodedef index="38" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/gamma" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[64]"  bytes="256" />
		</attributes>
	</nodedef>
	<nodedef index="39" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/gamma/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/gamma" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_2/BatchNorm/gamma" />
		</attributes>
	</nodedef>
	<nodedef index="40" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/beta" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[64]"  bytes="256" />
		</attributes>
	</nodedef>
	<nodedef index="41" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/beta/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/beta" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_2/BatchNorm/beta" />
		</attributes>
	</nodedef>
	<nodedef index="42" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_mean" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[64]"  bytes="256" />
		</attributes>
	</nodedef>
	<nodedef index="43" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_mean/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_mean" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_mean" />
		</attributes>
	</nodedef>
	<nodedef index="44" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_variance" op="Const">
		<inputs />
		<attributes>
			<attr name="dtype"  type="Type" value="Float" />
			<attr name="value"  type="Tensor" type="Float"  dims="[64]"  bytes="256" />
		</attributes>
	</nodedef>
	<nodedef index="45" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_variance/read" op="Identity">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_variance" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
			<attr name="_class"  type="List" string-size="1" strings="loc:@detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_variance" />
		</attributes>
	</nodedef>
	<nodedef index="46" name="detector/yolo-v3-tiny/Conv_2/BatchNorm/FusedBatchNorm" op="FusedBatchNorm">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/Conv2D" />
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/gamma/read" />
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/beta/read" />
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_mean/read" />
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/moving_variance/read" />
		</inputs>
		<attributes>
			<attr name="is_training"  type="Bool" value="False" />
			<attr name="T"  type="Type" value="Float" />
			<attr name="data_format"  type="String" value="NHWC" />
			<attr name="epsilon"  type="Float" value="1.001e-05" />
		</attributes>
	</nodedef>
	<nodedef index="47" name="detector/yolo-v3-tiny/Conv_2/LeakyRelu/alpha" op="Const">
		<inputs />
		<attributes>
			<attr name="value"  type="Tensor" type="Float"  dims="[]"  size = "0" />
			<attr name="dtype"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="48" name="detector/yolo-v3-tiny/Conv_2/LeakyRelu/mul" op="Mul">
		<inputs>
			<input node="detector/yolo-v3-tiny/Conv_2/LeakyRelu/alpha" />
			<input node="detector/yolo-v3-tiny/Conv_2/BatchNorm/FusedBatchNorm" />
		</inputs>
		<attributes>
			<attr name="T"  type="Type" value="Float" />
		</attributes>
	</nodedef>
	<nodedef index="49" name="detector/yolo-v3-tiny/Conv_2/LeakyRelu" op=
最低0.47元/天 解锁文章

200万优质内容无限畅学
TF：基于tensorflow框架利用python脚本下将YoloV3训练好的.ckpt模型文件转换为推理时采用的.pb文件
头部AI社区如有邀博主AI主题演讲请私信—心比天高，仗剑走天涯，保持热爱，奔赴向梦想！低调，专注，谦虚，自律，反思，成长，还算比较正能量的博主，公益免费传播…内心特别想在AI界做出一些可以推进历史进程影响力的技术(兴趣使然，有点小情怀，也有点使命感呀
01-09 1万+
TF：基于tensorflow框架利用python脚本下将YoloV3训练好的.ckpt模型文件转换为推理时采用的.pb文件 目录 准备.ckpt模型文件→最后生成的.pb文件 脚本代码转换命令 脚本代码转换过程记录 脚本代码 1、convert_weight.py 2、freeze_graph.py 相关文章Tensorflow：TF模型文件(checkp...
Github 项目 - YOLOV3 的 TensorFlow 复现
热门推荐
长风破浪会有时，直挂云帆济沧海
01-30 1万+
原文：Github 项目 - YOLOV3 的 TensorFlow 复现 - AIUAI Github 项目 - tensorflow-yolov3 作者：YunYang1994 论文：yolov3 最近 YunYang1994开源的基于 TensorFlow(TF-Slim) 复现的 YOLOv3 复现，并支持自定义数据集的训练. 该开源项目组成： YOLO v3 网络结构 权重转换We...
yolov3.weights
12-03
yolov3.weights文件，yolov3的训练参数，用于预训练。
【免费下载】 yolov3.weights - YOLOv3预训练权重文件
gitblog_09716的博客
09-07 1557
yolov3.weights - YOLOv3预训练权重文件 【下载地址】yolov3.weights-YOLOv3预训练权重文件 yolov3.weights - YOLOv3预训练权重文件欢迎来到YOLOv3预训练权重文件的存储库！本仓库提供的是YOLO（You Only Look Once）第三版本的预训练模型权重...
YOLOv3-Tiny预训练模型应用与实战解析
最新发布
weixin_36304957的博客
05-19 864
YOLOv3-Tiny是YOLO系列中一个轻量级的目标检测模型，它以简洁而有效的设计受到了许多行业从业者的青睐。这一章节将介绍YOLOv3-Tiny的发展历史，以及它在设计理念上的独到之处。我们将探讨其诞生的背景，一个模型如何在保持高检测精度的同时降低计算需求，以及YOLOv3-Tiny相比其它目标检测模型具有哪些独特优势。通过本章，读者将对YOLOv3-Tiny有一个全面的初步了解，为后续深入分析其网络架构、预训练模型和实际应用打下坚实基础。
yolov3weights.zip
07-04
1、darknet深度学习框架使用所用文件 2、可直接下载使用，官网较慢
yolov3-tiny.weights
12-16
基于YOLOv3-Tiny训练的人脸检测数据集，在darknet中迭代7000次，可以达到简单的演示效果
Tensorflow 读取frozen graph（.pb格式）模型并运行
PianGe_zyl的博客
08-21 4461
软件版本： tensorflow == 1.12.0 python == 3.6.1 关于cpkt模型的加载与使用可以看这个 关于.pb模型 不同于cpkt的是.pb模型将模型参数和网络结构固化在同一个文件中，使用中无需分别读取参数和结构，通过tensorflow自带函数可一并读取至graph，之后操作graph即可完成运算。 模型保存 https://blog.csdn.net/u014568072/article/details/85281769 模型读取 首先将模型文件存放至代码的同级目录，我的模型文
树莓派4B+NCS2代测试yolov3和yolov3_tiny实现object_detection
GankinTao的博客
01-10 6898
使用因特尔神经计算棒主要有两个原因，一、单独在树莓派4B上测试yolo会出现“段错误”，修改过后又会出现“总线错误”，这个问题没有得到解决，只好添加神经计算棒（有出现并解决此类问题的小伙伴可以讨论一下）二、使用神经计算棒可以明显提高模型在树莓派这种计算量较小的设备上的速度。 避免踩坑，总的步骤如下： 一、在树莓派上安装NCS环境 二、在windows上安装NCS环境（格式转换用到） 三、格式转换 ...
YOLOV3的配置文件，yolov3.weights等
11-20
YOLO v3的配置文件，模型文件等，包括yolov3.weights，yolov3.cfg，coco.name三个主要文件。
yolov3-tiny.weights.tar.gz
05-09
yolov3(pytorch)训练自己的数据集可参看本人blog。要使用的预训练权重：yolov3-tiny.weights
yolov3-tiny.weight
04-04
基于pytorch复现的yolov3网络，使用coco2017数据集得到的训练权重，注意，该权重为【weight】版本，在torch中调用需要程序转换。
yolov3_tiny.weights 生成的 tiny_yolo_weights.h5 下载
12-29
由 yolov3_tiny.weights生成的tiny_yolo_weights.h5 ,适用与keras-yolov3 版本
yolov3-tiny.conv.15
05-06
./darknet partial cfg/yolov3-tiny.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15 结果
yolov3-tiny权重和配置文件
02-16
yolov3-tiny 相关的配置文件包括训练好的权重，可以直接拿来用
yolov3-tiny原始weights模型转onnx模型并进行推理
weixin_38106878的博客
05-26 5744
时隔不知道多少天，我记起来我还有部分博客没写完(偷懒)，所以不能偷懒把它完成！！ 这篇博客的主要内容 将yolov3-tiny.weights模型转换到.onnx模型； 使用onnnxruntime-gpu（加速）模型推理；（不知道这里用加速形容是否合适） 照片推理测试； 视频文件推理测试； （注：本文用的模型为本人日常测试的时候训练出来的一个小模型，能够检测车辆、行人、人脸，但是效果有限，仅限用于本文的博客内容测试使用） 接下来，我们就按照博客内容一步一步的来进行我们的学习（ctrl+c、ctrl+v）
pycharm下python使用yolov3/yolov3-tiny训练好的权重文件.weights进行行人检测，批量测试自定义文件夹下的图片并输出至指定文件夹
CPZ742954007的博客
04-11 1万+
python使用yolov3/yolov3-tiny训练好的权重文件.weights进行行人检测，批量测试自定义文件夹下的图片并输出至指定文件夹 一、写在开头 最近在做毕业设计，研究的是目标识别与追踪，前段时间打算只用opencv识别个简单的目标就算了，但参考着论文硬着头皮撸了一星期的图片预处理，到了后面的识别部分，实在做不下去了，太南了！也不知道怎么具体去改其中的一些参数(数...
Python-纯tensorflow实现YOLOv3支持训练您自己的数据集
08-11
纯tensorflow实现YOLOv3，支持训练您自己的数据集
yolov3.cfg和yolov3.weights必须都安装吗
03-29
嗯，用户问的是“yolov3.cfg和yolov3.weights必须都安装吗”。首先，我需要确认用户对YOLO模型的理解程度。可能用户在使用OpenCV加载YOLOv3时遇到了问题，之前的对话已经涉及到了文件路径的问题，现在他们可能对这两...
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值