1.SyncedMemory
这个类SyncedMemory的作用是数据存储,在CPU和GPU上各有一块内存空间(大小为size_)。
先看初始化
SyncedMemory()
: cpu_ptr_(NULL),
gpu_ptr_(NULL),
size_(0),
head_(UNINITIALIZED),
own_cpu_data_(false),
device_context_(Caffe::GetDefaultDeviceContext()),
cl_gpu_mem_(NULL) {
}
重要的成员变量
enum SyncedHead {
UNINITIALIZED,
HEAD_AT_CPU,
HEAD_AT_GPU,
SYNCED
};//用来表示内存分配状态,未分配、CPU已分配、GPU已分配、两边内存已同步。
private:
void to_cpu();//判断head_,如果CPU未分配就开辟空间或者从GPU上copy
void to_gpu();//<span style="font-family: Arial;">判断head_,如果GPU未分配就开辟空间或者从CPU上copy
</span>
//两个指针
void* cpu_ptr_;
void* gpu_ptr_;
size_t size_;//数据大小
SyncedHead head_;//数据分配状态
bool own_cpu_data_;//CPU指针是自己分配的内存还是引用别人的内存(存疑)
DeviceContext *device_context_;//设备
#ifdef USE_GREENTEA
cl_mem cl_gpu_mem_;//在opencl里用,这时候的<span style="font-family: Arial;">gpu_ptr_是强制类型转换(void*)</span><span style="font-family: Arial;">cl_gpu_mem_。</span><span style="font-family: Arial;">
</span>#endif
还有个比较重要的区分,
const void* cpu_data();//返回CPU指针,注意是const不可改的,返回之前会调用to_cpu()。
void set_cpu_data(void* data);//CPU指针指向data,这时候不再own cpu data,但是size_未设置。
const void* gpu_data();//同上。
void* mutable_cpu_data();//这个返回的指针是mutable可变的,可以改写类的数据,返回前调用to_cpu(),调用之后要设置head_为不同步状态。
void* mutable_gpu_data();//同上。
2.Blob
这个类可以理解成每一层的参数数据(每层有个Blob的向量保存weights、bias等),保存了原始数据和传回来的梯度。
看成员变量,特别说明下,num_axes一般指shape_的size
protected:
shared_ptr<SyncedMemory> data_;//原始数据
shared_ptr<SyncedMemory> diff_;//传回的梯度
shared_ptr<SyncedMemory> shape_data_;//形状信息(就是shape_的一串数)
vector<int> shape_;//形状(也可以叫维度),最常见是4维(num、channel、width、height),最多INT_MAX维,其余维度作用存疑。
int count_;//数据Dtype num,大小是num*channel*width*height。
int capacity_;//按理说同count_,容量,具体作用存疑
DeviceContext *device_context_;//设备
初始化
Blob()
: data_(),
diff_(),
count_(0),
capacity_(0),
device_context_(Caffe::GetDefaultDeviceContext()) {
}
一些有代表性的函数
bool Blob<Dtype>::Reshape(const vector<int>& shape);//看名字很好理解是干啥的,主要会改shape_、shape_data_(内存)、count_、capacity_、两个数据(内存大小)
inline int offset(const int n, const int c = 0, const int h = 0, const int w = 0);//可以看组织形式 ((n * channels() + c) * height() + h) * width() + w
void Update();//data减去diff,调数学库做。
void FromProto(const BlobProto& proto, bool reshape = true);//这就是读protex文件要调的吧(待看)。
//其余一些asum、sumsq、scale、share等等操作,看名字就知道在干嘛了,直接贴上来,也有英文注释
/// @brief Compute the sum of absolute values (L1 norm) of the data.
Dtype asum_data() const;
/// @brief Compute the sum of absolute values (L1 norm) of the diff.
Dtype asum_diff() const;
/// @brief Compute the sum of squares (L2 norm squared) of the data.
Dtype sumsq_data() const;
/// @brief Compute the sum of squares (L2 norm squared) of the diff.
Dtype sumsq_diff() const;
/// @brief Scale the blob data by a constant factor.
void scale_data(Dtype scale_factor);
/// @brief Scale the blob diff by a constant factor.
void scale_diff(Dtype scale_factor);
/// @brief Add diff from other blob by alan
void AddDiffFrom(const Blob& other);
/// @brief Clear diff
void ClearDiff();
/**
* @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.
*/
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.
*/
void ShareDiff(const Blob& other);
bool ShapeEquals(const BlobProto& other);
总结一下吧,写完这些发现,syncedMem其实是保证了数据的存储和拷贝,而Blob封装了对于数据的操作(深度学习要用的操作),而且组织形式(data和diff)也符合网络的结构。