Caffe源码中blob文件分析

Caffe源码(caffe version commit: 09868ac , date: 2015.08.15)中有一些重要的头文件，这里介绍下include/caffe/blob.hpp文件的内容：

1.      Include文件：

(1)、<caffe/common.hpp>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/54955236

(2)、<caffe/proto/caffe.pb.h>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/55267162

(3)、<caffe/syncedmem.hpp>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/56665919

(4)、<caffe/util/math_functinons.hpp>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/56280708

2.        全局常量kMaxBlobAxes：

由const声明的全局常量kMaxBlobAxes表示Blob可以支持的最高维数，目前设置的支持的最高维数为32。

3.        类Blob：

Blob是Caffe中处理和传递实际数据的数据封装包，并且在CPU与GPU之间具有数据同步处理能力。从数学意义上说，blob是按C风格连续存储的N维数组，即在内部所存储的数据是一块连续的内存。

Blob是用来存储图像数据、网络参数(包括权值、偏置以及它们的梯度)、模型参数、学习到的参数、网络传输过程中产生的数据、网络中间的处理结果、优化过程的偏导数等各种数据。

Blob可以动态改变数组的尺寸，当拓展数组导致原有内存空间不足以存放下数据时(count_>capacity_)，就会通过Reshape函数实现重新确定空间大小。

Blob数据可以通过Protobuf来做相应的序列化操作，ToProto和FromProto两个函数完成相应的序列化、反序列化(数据解析)操作。

Caffe基于blobs存储和交换数据。网络各层之间的数据都是通过Blob来传递的。为了便于优化，blobs提供统一的内存接口来存储某种类型的数据，例如批量图像数据、模型参数以及用来进行优化的导数。

blobs可根据CPU主机与GPU设备的同步需要，屏蔽CPU/GPU混合运算在计算上的开销。主机和设备上的内存按需分配，以提高内存的使用效率。

对于批量图像数据来说，blob常规的维数为图像数量N*通道数K*图像高度H*图像宽度W。blob按行为主(row-major)进行存储，所以一个4维blob中，坐标为(n,k,h,w)的值的物理位置为((n*K+k)*H+h)*W+w，这也使得最后面/最右边的维度更新最快，其中：

(1)、Number/N是每个批次处理的数据量。批量处理信息有利于提供设备处理和交换的数据的吞吐率。在ImageNet上每个训练批量为256张图像，则N=256；

(2)、Channel/K是特征维度，例如对RGB图像来说，可以理解为通道数量，K=3；如果是网络中间结果，就是feature map的数量；

(3)、H、W：如果是图像数据，可以理解为图像的高度和宽度；如果是参数数据，可以理解为滤波核的高度和宽度。

虽然Caffe的图像应用例子中很多blobs都是4维坐标，但是对于非图像应用任务，blobs也完全可以照常使用。

参数Blob的维度是根据层的类型和配置而变化的。

对于blob中的数据，我们关心的是values(值)和gradients(梯度)，所以一个blob单元存储了两块数据------data_和diff_。前者是我们在网络中传送的普通数据，后者是通过网络计算得到的梯度。而且，由于数据既可存储在CPU上，又可存储在GPU上，因而有两种数据访问方式，如在CPU上的data_：静态方式，不改变数值(const Dtype* cpu_data() const;)；动态方式，改变数值(Dtype*mutable_cpu_data();)。GPU和diff_的操作与在CPU上的data_类似。

之所以这么设计是因为blob使用了一个SyncedMemory类来同步CPU和GPU上的数据，以隐藏同步的细节和最小化传送数据。一个经验准则是，如果不想改变数据，就一直使用常量调用，而且决不要在自定义类中存储指针。每次操作blob时，调用相应的函数来获取它的指针，因为SyncedMemory需要用这种方式来确定何时需要复制数据。

实际上，使用GPU时，Caffe中CPU代码先从磁盘中加载数据到blob，同时请求分配一个GPU设备核(devicekernel)以使用GPU进行计算，再将计算好的blob数据送入下一层，这样既实现了高效运算，又忽略了底层细节。只要所有layers均有GPU实现，这种情况下所有的中间数据和梯度都会保留在GPU 上。

注：以上关于Blob内容的介绍主要摘自由CaffeCN社区翻译的《Caffe官方教程中译本》。

<caffe/blob.hpp>文件的详细介绍如下：

#ifndef CAFFE_BLOB_HPP_
#define CAFFE_BLOB_HPP_#include <algorithm>
#include <string>
#include <vector>#include "caffe/common.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/syncedmem.hpp"
#include "caffe/util/math_functions.hpp"// 全局常量，用来表示Blob可以支持的最高维数
const int kMaxBlobAxes = 32;namespace caffe {/*** @brief A wrapper around SyncedMemory holders serving as the basic*        computational unit through which Layer%s, Net%s, and Solver%s*        interact.** TODO(dox): more thorough description.*/
template <typename Dtype>
class Blob {public:
// 默认不带参数的构造函数,初始化count_=0,capacity_=0Blob() : data_(), diff_(), count_(0), capacity_(0) {}/// @brief Deprecated; use <code>Blob(const vector<int>& shape)</code>.
// 带参数的显示构造函数，推荐使用带vector<int>参数的构造函数
// 这两个构造函数内部会均会调用Reshape(const vector<int>)函数
// 注：执行这两个构造函数后，并不会真正分配内存空间，只是用来设置当前blob的shape_、count_和capacity_大小explicit Blob(const int num, const int channels, const int height, const int width);explicit Blob(const vector<int>& shape);/// @brief Deprecated; use <code>Reshape(const vector<int>& shape)</code>.
// Reshape系列函数通过输入参数用来设置或重新设置当前blob的shape_、count_和capacity_大小
// 推荐使用带vector<int>参数的Reshape函数
// 内部会调用SyncedMemory的构造函数，但不会真正分配内存空间
// 通过num/channes/height/width参数设置shape_、count_和capacity_大小void Reshape(const int num, const int channels, const int height, const int width);/*** @brief Change the dimensions of the blob, allocating new memory if*        necessary.** This function can be called both to create an initial allocation* of memory, and to adjust the dimensions of a top blob during Layer::Reshape* or Layer::Forward. When changing the size of blob, memory will only be* reallocated if sufficient memory does not already exist, and excess memory* will never be freed.** Note that reshaping an input blob and immediately calling Net::Backward is* an error; either Net::Forward or Net::Reshape need to be called to* propagate the new input shape to higher layers.*/
// 通过vector<int>参数设置shape_、count_和capacity_大小void Reshape(const vector<int>& shape);
// 通过类BlobShape参数设置shape_、count_和capacity_大小
// BlobShape是定义在caffe.proto中的一个message，其字段有dimvoid Reshape(const BlobShape& shape);
// 通过外部的blob参数来设置shape_、count_和capacity_大小void ReshapeLike(const Blob& other);// 以string类型获得当前blob的shape_和count_值inline string shape_string() const {ostringstream stream;for (int i = 0; i < shape_.size(); ++i) {stream << shape_[i] << " ";}stream << "(" << count_ << ")";return stream.str();}// 获得当前Blob的所有维度值inline const vector<int>& shape() const { return shape_; }/*** @brief Returns the dimension of the index-th axis (or the negative index-th*        axis from the end, if index is negative).** @param index the axis index, which may be negative as it will be*        "canonicalized" using CanonicalAxisIndex.*        Dies on out of range index.*/
// 获得当前Blob指定索引的维度值inline int shape(int index) const {return shape_[CanonicalAxisIndex(index)];}// 获得当前Blob的维数inline int num_axes() const { return shape_.size(); }// 获得当前Blob的元素个数inline int count() const { return count_; }/*** @brief Compute the volume of a slice; i.e., the product of dimensions*        among a range of axes.** @param start_axis The first axis to include in the slice.** @param end_axis The first axis to exclude from the slice.*/
// 根据指定的start axis和end axis(部分blob)计算blob元素个数inline int count(int start_axis, int end_axis) const {CHECK_LE(start_axis, end_axis);CHECK_GE(start_axis, 0);CHECK_GE(end_axis, 0);CHECK_LE(start_axis, num_axes());CHECK_LE(end_axis, num_axes());int count = 1;for (int i = start_axis; i < end_axis; ++i) {count *= shape(i);}return count;}/*** @brief Compute the volume of a slice spanning from a particular first*        axis to the final axis.** @param start_axis The first axis to include in the slice.*/
// 根据指定的start axis(部分blob)计算blob元素个数inline int count(int start_axis) const {return count(start_axis, num_axes());}/*** @brief Returns the 'canonical' version of a (usually) user-specified axis,*        allowing for negative indexing (e.g., -1 for the last axis).** @param index the axis index.*        If 0 <= index < num_axes(), return index.*        If -num_axes <= index <= -1, return (num_axes() - (-index)),*        e.g., the last axis index (num_axes() - 1) if index == -1,*        the second to last if index == -2, etc.*        Dies on out of range index.*/
// Blob的index可以是负值,对参数axis_index进行判断,结果返回一个正的索引值
// 如果axis_index是负值，则要求axis_index>=-shape_.size(),则返回axis_index+shape_.size()
// 如果axis_index是正值，则要求axis_index<shape_.size(),则直接返回axis_indexinline int CanonicalAxisIndex(int axis_index) const {CHECK_GE(axis_index, -num_axes())<< "axis " << axis_index << " out of range for " << num_axes()<< "-D Blob with shape " << shape_string();CHECK_LT(axis_index, num_axes())<< "axis " << axis_index << " out of range for " << num_axes()<< "-D Blob with shape " << shape_string();if (axis_index < 0) {return axis_index + num_axes();}return axis_index;}/// @brief Deprecated legacy shape accessor num: use shape(0) instead.
// 获得当前blob的num，推荐调用shape(0)函数inline int num() const { return LegacyShape(0); }/// @brief Deprecated legacy shape accessor channels: use shape(1) instead.
// 获得当前blob的channels，推荐调用shape(1)函数inline int channels() const { return LegacyShape(1); }/// @brief Deprecated legacy shape accessor height: use shape(2) instead.
// 获得当前blob的height，推荐调用shape(2)函数inline int height() const { return LegacyShape(2); }/// @brief Deprecated legacy shape accessor width: use shape(3) instead.
// 获得当前blob的width，推荐调用shape(3)函数inline int width() const { return LegacyShape(3); }
// 获得当前blob的某一维度值inline int LegacyShape(int index) const {CHECK_LE(num_axes(), 4)<< "Cannot use legacy accessors on Blobs with > 4 axes.";CHECK_LT(index, 4);CHECK_GE(index, -4);if (index >= num_axes() || index < -num_axes()) {// Axis is out of range, but still in [0, 3] (or [-4, -1] for reverse// indexing) -- this special case simulates the one-padding used to fill// extraneous axes of legacy blobs.return 1;}return shape(index);}// 根据num、channels、height、width计算偏移量:((n*K+k)*H+h)*W+winline int offset(const int n, const int c = 0, const int h = 0, const int w = 0) const {CHECK_GE(n, 0);CHECK_LE(n, num());CHECK_GE(channels(), 0);CHECK_LE(c, channels());CHECK_GE(height(), 0);CHECK_LE(h, height());CHECK_GE(width(), 0);CHECK_LE(w, width());return ((n * channels() + c) * height() + h) * width() + w;}
// 根据vector<int> index计算偏移量：((n*K+k)*H+h)*W+winline int offset(const vector<int>& indices) const {CHECK_LE(indices.size(), num_axes());int offset = 0;for (int i = 0; i < num_axes(); ++i) {offset *= shape(i);if (indices.size() > i) {CHECK_GE(indices[i], 0);CHECK_LT(indices[i], shape(i));offset += indices[i];}}return offset;}/*** @brief Copy from a source Blob.** @param source the Blob to copy from* @param copy_diff if false, copy the data; if true, copy the diff* @param reshape if false, require this Blob to be pre-shaped to the shape*        of other (and die otherwise); if true, Reshape this Blob to other's*        shape if necessary*/
// 从外部blob拷贝数据到当前的blob
// 若reshape参数为true，如果两边blob的reshape不相同，则会重新reshape
// 若copy_diff为false,则拷贝data_数据；若copy_diff为true，则拷贝diff_数据void CopyFrom(const Blob<Dtype>& source, bool copy_diff = false, bool reshape = false);// 根据给定的位置访问数据
// 根据指定的偏移量获得前向传播数据data_的一个元素的值inline Dtype data_at(const int n, const int c, const int h, const int w) const {return cpu_data()[offset(n, c, h, w)];}
// 根据指定的偏移量获得反向传播梯度diff_的一个元素的值inline Dtype diff_at(const int n, const int c, const int h, const int w) const {return cpu_diff()[offset(n, c, h, w)];}
// 根据指定的偏移量获得前向传播数据data_的一个元素的值inline Dtype data_at(const vector<int>& index) const {return cpu_data()[offset(index)];}
// 根据指定的偏移量获得反向传播梯度diff_的一个元素的值inline Dtype diff_at(const vector<int>& index) const {return cpu_diff()[offset(index)];}// 获得前向传播数据data_的指针inline const shared_ptr<SyncedMemory>& data() const {CHECK(data_);return data_;}
// 获得反向传播梯度diff_的指针inline const shared_ptr<SyncedMemory>& diff() const {CHECK(diff_);return diff_;}// Blob的数据访问函数，包括CPU和GPU
// 带mutable_前缀的函数是可以对Blob数据进行改写的；其它不带的是只读的，不允许改写数据const Dtype* cpu_data() const; // 调用SyncedMemory::cpu_data()函数const Dtype* gpu_data() const; // 调用SyncedMemory::gpu_data()函数const Dtype* cpu_diff() const; // 调用SyncedMemory::cpu_data()函数const Dtype* gpu_diff() const; // 调用SyncedMemory::gpu_data()函数Dtype* mutable_cpu_data(); // 调用SyncedMemory::mutable_cpu_data()函数Dtype* mutable_gpu_data(); // 调用SyncedMemory::mutable_gpu_data()函数Dtype* mutable_cpu_diff(); // 调用SyncedMemory::mutable_cpu_data()函数Dtype* mutable_gpu_diff(); // 调用SyncedMemory::mutable_gpu_data()函数void set_cpu_data(Dtype* data); // 调用SyncedMemory::set_cpu_data(void*)函数// 它会被网络中存储参数的Blob调用，完成梯度下降过程中的参数更新
// 调用caffe_axpy函数重新计算data_(weight，bias 等减去对应的导数): data_ = -1 * diff_ + data_void Update();// Blob的数据持久化函数，通过Protobuf来做相应的序列化/反序列化操作
// BlobProto是定义在caffe.proto中的一个message，其字段有shape(BlobShape)、data、diff、num、channels、height、width
// 将BlobProto的shape/data/diff分别copy给当前blob的shape_/data_/diff_完成数据解析(反序列化)
// 若reshape参数为true，则会对当前的blob重新进行reshapevoid FromProto(const BlobProto& proto, bool reshape = true);
// 将Blob的shape_/data_/diff_(如果write_diff为true)分别copy给BlobProto的shape/data/diff完成序列化void ToProto(BlobProto* proto, bool write_diff = false) const;/// @brief Compute the sum of absolute values (L1 norm) of the data.
// 计算data_的L1范式：向量中各个元素绝对值之和Dtype asum_data() const;/// @brief Compute the sum of absolute values (L1 norm) of the diff.
// 计算diff_的L1范式：向量中各个元素绝对值之和Dtype asum_diff() const;/// @brief Compute the sum of squares (L2 norm squared) of the data.
// 计算data_的L2范式平方：向量中各元素的平方和Dtype sumsq_data() const;/// @brief Compute the sum of squares (L2 norm squared) of the diff.
// // 计算diff_的L2范式平方：向量中各元素的平方和Dtype sumsq_diff() const;/// @brief Scale the blob data by a constant factor.
// 将data_数据乘以一个因子：X = alpha*Xvoid scale_data(Dtype scale_factor);/// @brief Scale the blob diff by a constant factor.
// 将diff_数据乘以一个因子：X = alpha*Xvoid scale_diff(Dtype scale_factor);/*** @brief Set the data_ shared_ptr to point to the SyncedMemory holding the*        data_ of Blob other -- useful in Layer%s which simply perform a copy*        in their Forward pass.** This deallocates the SyncedMemory holding this Blob's data_, as* shared_ptr calls its destructor when reset with the "=" operator.*/
// 将外部指定的blob的data_指针指向给当前blob的data_,以实现共享data_void ShareData(const Blob& other);/*** @brief Set the diff_ shared_ptr to point to the SyncedMemory holding the*        diff_ of Blob other -- useful in Layer%s which simply perform a copy*        in their Forward pass.** This deallocates the SyncedMemory holding this Blob's diff_, as* shared_ptr calls its destructor when reset with the "=" operator.*/
// 将外部指定的blob的diff_指针指向给当前blob的diff_,以实现共享diff_void ShareDiff(const Blob& other);
// 比较两个blob的shape是否相同
// BlobProto是定义在caffe.proto中的一个message，其字段有shape(BlobShape)、data、diff、num、channels、height、widthbool ShapeEquals(const BlobProto& other);protected:
// Caffe中类的成员变量名都带有后缀"_"，这样就容易区分临时变量和类成员变量shared_ptr<SyncedMemory> data_; // 存储前向传播的数据shared_ptr<SyncedMemory> diff_; // 存储反向传播的导数、梯度、偏差vector<int> shape_; // Blob的维度值，通过Reshape函数的shape参数获得相应值,若为4维，则依次为num、channels、height、widthint count_; // 表示Blob中的元素个数，shape_所有元素的乘积int capacity_; // 表示当前Blob的元素个数(控制动态分配)，因为Blob可能会reshape// 禁止使用Blob类的拷贝和赋值操作DISABLE_COPY_AND_ASSIGN(Blob);
};  // class Blob}  // namespace caffe#endif  // CAFFE_BLOB_HPP_

在caffe.proto文件中，有3个message是与blob有关的，如下：

// 可选package声明符，用来防止不同的消息类型有命名冲突
package caffe; // 以下所有生成的信息将会在命名空间caffe内： namespace caffe { ... }// 说明：带"deprecated"关键字的信息可以不用看，过时的，后面有可能是被废弃的// 以下三个是关于blob的三个类:BlobShape、BlobProto、BlobProtoVector
// Specifies the shape (dimensions) of a Blob.
message BlobShape { // 数据块形状(Blob的维度)，若为4维，则为num、channel、height、widthrepeated int64 dim = 1 [packed = true]; // blob shape数组
}message BlobProto { // blob属性类optional BlobShape shape = 7; // BlobShappe类对象repeated float data = 5 [packed = true]; // float类型的data，前向repeated float diff = 6 [packed = true]; // float类型的diff，后向repeated double double_data = 8 [packed = true]; // double类型的data，前向repeated double double_diff = 9 [packed = true]; // double类型的diff，后向// 4D dimensions -- deprecated.  Use "shape" instead.// 已使用BlobShape shape替代optional int32 num = 1 [default = 0];optional int32 channels = 2 [default = 0];optional int32 height = 3 [default = 0];optional int32 width = 4 [default = 0];
}// The BlobProtoVector is simply a way to pass multiple blobproto instances
// around.
message BlobProtoVector { // 存放多个BlobProto实例repeated BlobProto blobs = 1;
}

blob的测试代码如下：

#include "funset.hpp"
#include <string>
#include <vector>
#include "common.hpp"int test_caffe_blob()
{caffe::Blob<float> blob1;std::vector<int> shape{ 2, 3, 4, 5 };caffe::Blob<float> blob2(shape);std::vector<int> blob_shape = blob2.shape();fprintf(stderr, "blob shape: ");for (auto index : blob_shape) {fprintf(stderr, "%d    ", index);}std::vector<int> shape_{ 6, 7, 8, 9 };blob2.Reshape(shape_);std::vector<int> blob_shape_ = blob2.shape();fprintf(stderr, "\nnew blob shape: ");for (auto index : blob_shape_) {fprintf(stderr, "%d    ", index);}fprintf(stderr, "\n");int value = blob2.shape(-1);fprintf(stdout, "blob index -1: %d\n", value);int num_axes = blob2.num_axes();fprintf(stderr, "blob num axes(dimension): %d\n", num_axes);int count = blob2.count();fprintf(stderr, "blob count sum: %d\n", count);count = blob2.count(2, 4);fprintf(stderr, "blob count(start_axis(2), end_axis(4)): %d\n", count);count = blob2.count(1);fprintf(stderr, "blob count(start_axis(1)): %d\n", count);int canonical_axis_index = blob2.CanonicalAxisIndex(-2);fprintf(stderr, "blob canonical axis index: %d\n", canonical_axis_index);int num = blob2.num();int channels = blob2.channels();int height = blob2.height();int width = blob2.width();int legacy_shape = blob2.LegacyShape(-2);fprintf(stderr, "blob num: %d, channels: %d, height: %d, width: %d, legacy shape(-2): %d\n",num, channels, height, width, legacy_shape);std::vector<int> indices{ 2, 3, 7, 6 };int offset1 = blob2.offset(indices);int offset2 = blob2.offset(indices[0], indices[1], indices[2], indices[3]);fprintf(stderr, "blob offset1: %d, offset2: %d\n", offset1, offset2);std::string shape_string = blob2.shape_string();fprintf(stderr, "shape string: %s\n", shape_string.c_str());caffe::BlobProto blob_proto;blob_proto.set_num(6);blob_proto.set_channels(7);blob_proto.set_height(8);blob_proto.set_width(9);bool flag = blob2.ShapeEquals(blob_proto);fprintf(stderr, "blob2's shape and blob_proto's shape are equal: %d\n", flag);int blob_proto_data_size_float = blob_proto.data_size();int blob_proto_data_size_double = blob_proto.double_data_size();int blob_proto_diff_size_float = blob_proto.diff_size();int blob_proto_diff_size_double = blob_proto.double_diff_size();fprintf(stderr, "blob_proto data/diff size: %d, %d, %d, %d\n", blob_proto_data_size_float,blob_proto_data_size_double, blob_proto_diff_size_float, blob_proto_diff_size_double);caffe::BlobShape blob_proto_shape;for (int i = 0; i < 4; ++i) {blob_proto_shape.add_dim(i + 10);}blob2.Reshape(blob_proto_shape);blob_shape_ = blob2.shape();fprintf(stderr, "new blob shape: ");for (auto index : blob_shape_) {fprintf(stderr, "%d    ", index);}fprintf(stderr, "\n");fprintf(stderr, "blob proto shape: ");for (int i = 0; i < blob_proto_shape.dim_size(); ++i) {fprintf(stderr, "%d    ", blob_proto_shape.dim(i));}fprintf(stderr, "\n");// 注：以上进行的所有操作均不会申请分配任何内存// cv::Mat -> Blobstd::string image_name = "E:/GitCode/Caffe_Test/test_data/images/a.jpg";cv::Mat mat = cv::imread(image_name, 1);if (!mat.data) {fprintf(stderr, "read image fail: %s\n", image_name.c_str());return -1;}cv::Mat mat2;mat.convertTo(mat2, CV_32FC3);std::vector<int> mat_reshape{ 1, mat2.channels(), mat2.rows, mat2.cols };blob2.Reshape(mat_reshape);float sum1 = blob2.asum_data();blob2.set_cpu_data((float*)mat2.data);float sum2 = blob2.asum_data();blob2.scale_data(0.5);float sum3 = blob2.asum_data();float sum4 = blob2.sumsq_data();fprintf(stderr, "sum1: %f, sum2: %f, sum3: %f, sum4: %f\n", sum1, sum2, sum3, sum4);float value2 = blob2.data_at(0, 2, 100, 200);fprintf(stderr, "data at value: %f\n", value2);const float* data = blob2.cpu_data();fprintf(stderr, "data at 0: %f\n", data[0]);cv::Mat mat3;mat2.convertTo(mat3, CV_8UC3);image_name = "E:/GitCode/Caffe_Test/test_data/images/a_ret.jpg";cv::imwrite(image_name, mat3);return 0;
}

测试结果如下：

GitHub：https://github.com/fengbingchun/Caffe_Test