Caffe:Develop
lastmod 2018-08-17created 2016-01-0816 edits
Contains source code analysis for Caffe development.
Blob
A wrapper class around #SyncedMemory that interacts with #Layer, #Net, and #Solver objects.
- Header:
include/caffe/blob.hpp - Source:
src/caffe/blob.cpp
To understand this object, you first need to understand #SyncedMemory. Its characteristics are as follows.
- A 4-dimensional array.
- Stores data and is used for communication between modules.
- Provides a memory interface so you don't have to worry about CPU/GPU synchronization and heterogeneous computation.
- Models are stored with Google Protocol Buffers.
- Large datasets are stored in LevelDB.
Member variables
protected:
shared_ptr<SyncedMemory> data_;
shared_ptr<SyncedMemory> diff_;
shared_ptr<SyncedMemory> shape_data_;
vector<int> shape_;
int count_;
int capacity_;
Layer
Its characteristics are as follows.
- The building block of a network (#Net), taking one or more #Blobs as input and producing one or more #Blobs as output.
- Has a Forward Pass and a Backward Pass, computes the gradients of the Layer parameters, and backpropagates (Back probagation) the error of the final output.
- Provides Layers of types such as Convolution, Pooling, Inner products, Rectified linear, Logistic, Local response normalization, Element-wise operations, Losses like softmax and hinge.
Net
Solver
SyncedMemory
Manages memory allocation and synchronization between the host (CPU) and device (GPU).
- Header:
include/caffe/syncedmem.hpp - Source:
src/caffe/syncedmem.cpp
Defines allocation/free operations for CPU memory and GPU memory.
Member variables
-
void* cpu_ptr_;: memory pointer for the CPU. -
void* gpu_ptr_;: memory pointer for the GPU. -
size_t size_;: allocated memory size.
util/math_functions
Mostly contains API wrappers for the GPU (CUDA), like the following.
inline void caffe_gpu_memset(const size_t N, const int alpha, void* X) {
#ifndef CPU_ONLY
CUDA_CHECK(cudaMemset(X, alpha, N)); // NOLINT(caffe/alt_fn)
#else
NO_GPU;
#endif
}