主要定义了一个模板类net
首先看一下数据成员:
先看一下AppendTop AppendBottom AppendParam:
然后看一下Init函数
首先看一下数据成员:
/// @brief The network name
string name_;
/// @brief The phase: TRAIN or TEST
Phase phase_;
/// @brief Individual layers in the net
vector<shared_ptr<Layer<Dtype> > > layers_;
vector<string> layer_names_;
map<string, int> layer_names_index_;
vector<bool> layer_need_backward_;
/// @brief the blobs storing intermediate results between the layer.
vector<shared_ptr<Blob<Dtype> > > blobs_;//blobs_存储的是中间结果,是针对整个网络中所有非参数blob而设计的一个变量。我觉得params_存储的也是中间结果
vector<string> blob_names_;//整个网络中,所有非参数blob的name。
map<string, int> blob_names_index_;
vector<bool> blob_need_backward_;//整个网络中,所有非参数blob,是否需要backward。注意,这里所说的所有非参数blob其实指的是AppendTop函数中遍历的所有top blob,并不是每一层的top+bottom,因为这一层的top就是下一层的bottom,网络是一层一层堆起来的。
/// bottom_vecs stores the vectors containing the input for each layer.
/// They don't actually host the blobs (blobs_ does), so we simply store
/// pointers.
vector<vector<Blob<Dtype>*> > bottom_vecs_;//存储整个网络所有网络层的bottom blob指针,实际上存储的是前一层的top,因为网络是一层一层堆起来的
vector<vector<int> > bottom_id_vecs_;//存储整个网络所有网络层的bottom blob的ID
vector<vector<bool> > bottom_need_backward_;//整个网络所有网络层的bottom blob是否需要backward
/// top_vecs stores the vectors containing the output for each layer
vector<vector<Blob<Dtype>*> > top_vecs_;//存储整个网络所有网络层的top blob指针.
vector<vector<int> > top_id_vecs_;//存储整个网络所有网络层的top blob的ID.top_id_vecs_中存储的最基本元素是blob_id ——> 每一个新的blob都会赋予其一个blob_id,top_vecs_则与之对应,但是这个blob_id可能是会有重复的(因为in-place)
/// Vector of weight in the loss (or objective) function of each net blob,
/// indexed by blob_id.
vector<Dtype> blob_loss_weights_;//每次遍历一个layer的时候,都会resize blob_loss_weights_, 然后调用模板类layer的loss函数返回loss_weight
vector<vector<int> > param_id_vecs_;//存储的基本元素是net_param_id,每遍历一个参数blob,net_param_id和param_id_vecs_都会更新
vector<int> param_owners_;//param_owners_ 是一个存储parameter "onwer"的一个向量 ——> -1 表示当前Layer就是该parameter的"owner"
vector<string> param_display_names_;
vector<pair<int, int> > param_layer_indices_;//其元素为当layer_id 与当前param_id 组成的pair.vector<pair<int, int> > param_layer_indices_
map<string, int> param_names_index_;//是整个网络的参数non-empty name与index的映射。注意,这个name是ParamSpec 类型中的name。
/// blob indices for the input and the output of the net整个网络的输入输出blob以及ID
vector<int> net_input_blob_indices_;
vector<int> net_output_blob_indices_;
vector<Blob<Dtype>*> net_input_blobs_;
vector<Blob<Dtype>*> net_output_blobs_;
/// The parameters in the network.
vector<shared_ptr<Blob<Dtype> > > params_;//整个网络的参数blob。 !!!不管这个参数有没有non-emty name,是否参与share!!!
vector<Blob<Dtype>*> learnable_params_;//
/**
* The mapping from params_ -> learnable_params_: we have
* learnable_param_ids_.size() == params_.size(),
* and learnable_params_[learnable_param_ids_[i]] == params_[i].get()
* if and only if params_[i] is an "owner"; otherwise, params_[i] is a sharer
* and learnable_params_[learnable_param_ids_[i]] gives its owner.
*/
vector<int> learnable_param_ids_;
/// the learning rate multipliers for learnable_params_
vector<float> params_lr_;
vector<bool> has_params_lr_;
/// the weight decay multipliers for learnable_params_
vector<float> params_weight_decay_;
vector<bool> has_params_decay_;
/// The bytes of memory used by this net
size_t memory_used_;
/// Whether to compute and display debug info for the net.
bool debug_info_;
/// The root net that actually holds the shared layers in data parallelism
const Net* const root_net_;
先看一下AppendTop AppendBottom AppendParam:
================================AppendTop函数================================================
// Helper for Net::Init: add a new input or top blob to the net. (Inputs have
// layer_id == -1, tops have layer_id >= 0.)
template <typename Dtype>
void Net<Dtype>::AppendTop(const NetParameter& param, const int layer_id,
const int top_id, set<string>* available_blobs,
map<string, int>* blob_name_to_idx) {
shared_ptr<LayerParameter> layer_param((layer_id >= 0) ?
(new LayerParameter(param.layer(layer_id))) : NULL);
const string& blob_name = layer_param ?
(layer_param->top_size() > top_id ?
layer_param->top(top_id) : "(automatic)") : param.input(top_id);
// Check if we are doing in-place computation
if (blob_name_to_idx && layer_param && layer_param->bottom_size() > top_id &&
blob_name == layer_param->bottom(top_id)) {
// In-place computation
LOG_IF(INFO, Caffe::root_solver())
<< layer_param->name() << " -> " << blob_name << " (in-place)";
top_vecs_[layer_id].push_back(blobs_[(*blob_name_to_idx)[blob_name]].get());
top_id_vecs_[layer_id].push_back((*blob_name_to_idx)[blob_name]);
} else if (blob_name_to_idx &&
blob_name_to_idx->find(blob_name) != blob_name_to_idx->end()) {
// If we are not doing in-place computation but have duplicated blobs,
// raise an error.
LOG(FATAL) << "Top blob '" << blob_name
<< "' produced by multiple sources.";
} else {
// Normal output.
if (Caffe::root_solver()) {
if (layer_param) {
LOG(INFO) << layer_param->name() << " -> " << blob_name;
} else {
LOG(INFO) << "Input " << top_id << " -> " << blob_name;
}
}
shared_ptr<Blob<Dtype> > blob_pointer(new Blob<Dtype>());
const int blob_id = blobs_.size();//blobs只是存储中间结果;每次遍历到一个top blob都会更新blob_id
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(false);
//blob_name_to_idx是一个局部变量,其实它是在当前layer的top blob 和下一层的bottom blob间起着一个桥梁作用。
//blob_name_to_idx中元素的pair是从网络最开始一层一层搭建的过程中压入map的,其中的name和id都是不重复的。name是关键字——不重复是map数据结构的必然要求,id也是不重复的——0,1,2...
//blob_name_to_idx和blobs_一样,在"Normal output"的情形下,每次遍历到一个top blob的时候都会更新
if (blob_name_to_idx) { (*blob_name_to_idx)[blob_name] = blob_id; }//添加新元素-->map可以通过下标访问符为(关联)容器添加新元素 blob_name_to_idx是指针,并不是空指针,所以可以执行if之后的代码。
if (layer_id == -1) {
if (layer_id == -1) {
// Set the (explicitly specified) dimensions of the input blob.
if (param.input_dim_size() > 0) {
blob_pointer->Reshape(param.input_dim(top_id * 4),
param.input_dim(top_id * 4 + 1),
param.input_dim(top_id * 4 + 2),
param.input_dim(top_id * 4 + 3));
} else {
blob_pointer->Reshape(param.input_shape(top_id));
}
net_input_blob_indices_.push_back(blob_id);
net_input_blobs_.push_back(blob_pointer.get());//当layer_id==-1时,即当前层为输入层的时候,会向net_input_blob_indices_里添加新元素,即add new input blob
} else {
top_id_vecs_[layer_id].push_back(blob_id);
top_vecs_[layer_id].push_back(blob_pointer.get());//当layer_id !=-1时,即当前层不是输入层的时候,会向net_input_blob_indices_里添加新元素,即add new top blob
}
}
if (available_blobs) { available_blobs->insert(blob_name); }
}
================================AppendBottom函数================================================
// Helper for Net::Init: add a new bottom blob to the net.
template <typename Dtype>
int Net<Dtype>::AppendBottom(const NetParameter& param, const int layer_id,
const int bottom_id, set<string>* available_blobs,
map<string, int>* blob_name_to_idx) {
const LayerParameter& layer_param = param.layer(layer_id);
const string& blob_name = layer_param.bottom(bottom_id);
if (available_blobs->find(blob_name) == available_blobs->end()) {
LOG(FATAL) << "Unknown bottom blob '" << blob_name << "' (layer '"
<< layer_param.name() << "', bottom index " << bottom_id << ")";
}
const int blob_id = (*blob_name_to_idx)[blob_name];//blob_name_to_idx是一个map,其关键字是不重复的。blob_name_to_idx在输入层初始化过了-->*blob_name_to_idx)[blob_name] = blob_id
LOG_IF(INFO, Caffe::root_solver())
<< layer_names_[layer_id] << " <- " << blob_name;
bottom_vecs_[layer_id].push_back(blobs_[blob_id].get());//调用shared_ptr类的get()方法提取存储在blobs_中的中间变量
bottom_id_vecs_[layer_id].push_back(blob_id);
available_blobs->erase(blob_name);
bool propagate_down = true; // propagate_down默认为true
// Check if the backpropagation on bottom_id should be skipped
if (layer_param.propagate_down_size() > 0)
propagate_down = layer_param.propagate_down(bottom_id);
const bool need_backward = blob_need_backward_[blob_id] &&
propagate_down;//propagate_down为true,则表示参与BP;否则,skip bp
bottom_need_backward_[layer_id].push_back(need_backward);
return blob_id;
}
================================AppendParam函数================================================
template <typename Dtype>
void Net<Dtype>::AppendParam(const NetParameter& param, const int layer_id,
const int param_id) {
const LayerParameter& layer_param = layers_[layer_id]->layer_param();//模板类Layer的layer_param方法,返回Layerparameter类型成员
const int param_size = layer_param.param_size();
string param_name =
(param_size > param_id) ? layer_param.param(param_id).name() : "";
if (param_name.size()) {
param_display_names_.push_back(param_name);//vector<string> param_display_names_ 这里param_name获取的是PaParamSpec类型中的name成员,如果有name且非空,就把name压入该向量,否则就压入param_id
} else {
ostringstream param_display_name;
param_display_name << param_id;
param_display_names_.push_back(param_display_name.str());
}
//Append 参数blob 每一次循环,net_param_id和param_id_vecs_都会更新
const int net_param_id = params_.size();//vector<shared_ptr<Blob<Dtype> > > params_--->The parameters in the network,整个网络的参数的id,!!!不管这个参数有没有non-emty name,是否参与share!!!
params_.push_back(layers_[layer_id]->blobs()[param_id]);//将当前layer当前"参数blob"压入params_ --->vector<shared_ptr<Blob<Dtype> > > params_
param_id_vecs_[layer_id].push_back(net_param_id);//将整个网络的参数按层的形式来存储,存储的元素可以理解为params_这个向量的下标值(类型为整型)
param_layer_indices_.push_back(make_pair(layer_id, param_id));//param_layer_indices_是向量,其元素为当layer_id 与当前param_id 组成的pair.vector<pair<int, int> > param_layer_indices_
//获取每个param_id所对应的Paramspec类型成员,如果param_id >= param_size 则返回default_param_spec。注意param_size <= num_param_blobs
ParamSpec default_param_spec;
const ParamSpec* param_spec = (layer_param.param_size() > param_id) ?
&layer_param.param(param_id) : &default_param_spec;
if (!param_size || !param_name.size() || (param_name.size() &&
param_names_index_.find(param_name) == param_names_index_.end())) {
// This layer "owns" this parameter blob -- it is either anonymous
// (i.e., not given a param_name) or explicitly given a name that we
// haven't already seen.
// 相反,如果param_name不为空,而且能够在param_names_index_中找到,说明这个parameter已经存在于之前的某个或者某些网络层里,说明这个parameter是共享于多个layer
// 在caffe.proto的message ParamSpec里关于name的注释——>To share a parameter between two layers, give it a (non-empty) name, 可见,如果一个parameter是共享与多个网络层,那么它会有一个非空的name
param_owners_.push_back(-1);//vector<int> param_owners_ 是一个存储parameter "onwer"的一个向量 ——> -1 表示当前Layer就是该parameter的"owner"
//添加param_name
if (param_name.size()) {
//map<string, int> param_names_index_是整个网络的参数non-empty name与index的映射。
//注意,这个name是ParamSpec 类型中的name,而且,""To share a parameter between two layers, give it a (non-empty) name"",所以说这个map中存储的pair是<会被share的parameter_name, 其对应index>
param_names_index_[param_name] = net_param_id;//map<string, int> param_names_index_ 。虽然每一次循环,net_param_id都会更新,但是net_param_id只有当param_name.size()>0时才会被压入向量param_names_index_
}
//添加learnable_param
const int learnable_param_id = learnable_params_.size();//vector<Blob<Dtype>*> learnable_params_
learnable_params_.push_back(params_[net_param_id].get());//压入learnable parameter ---> 在模板类layer中,定义了一个blobs_成员,其存储的就是learnable parameter。随后压入learnable_param_id
learnable_param_ids_.push_back(learnable_param_id);//vector<int> learnable_param_ids_
has_params_lr_.push_back(param_spec->has_lr_mult());//vector<bool> has_params_lr_
has_params_decay_.push_back(param_spec->has_decay_mult());
params_lr_.push_back(param_spec->lr_mult());//vector<float> params_lr_
params_weight_decay_.push_back(param_spec->decay_mult());
} else {
// Named param blob with name we've seen before: share params
const int owner_net_param_id = param_names_index_[param_name];//因为"To share a parameter between two layers, give it a (non-empty) name",所以这句代码就是获取shared parameter的"owner" net_param_id
param_owners_.push_back(owner_net_param_id);//vector<int> param_owners_
const pair<int, int>& owner_index =
param_layer_indices_[owner_net_param_id];//只获取了那些shared的parameter,即具有non-empty name的parameter的pair<layer_id, param_id>
const int owner_layer_id = owner_index.first;
const int owner_param_id = owner_index.second;
LOG_IF(INFO, Caffe::root_solver()) << "Sharing parameters '" << param_name
<< "' owned by "
<< "layer '" << layer_names_[owner_layer_id] << "', param "
<< "index " << owner_param_id;
Blob<Dtype>* this_blob = layers_[layer_id]->blobs()[param_id].get();//获取当前层的当前参数Blob
Blob<Dtype>* owner_blob =
layers_[owner_layer_id]->blobs()[owner_param_id].get();//获取owner layer的对应的参数blob
const int param_size = layer_param.param_size();
if (param_size > param_id && (layer_param.param(param_id).share_mode() ==
ParamSpec_DimCheckMode_PERMISSIVE)) {
// Permissive dimension checking -- only check counts are the same.
CHECK_EQ(this_blob->count(), owner_blob->count())
<< "Cannot share param '" << param_name << "' owned by layer '"
<< layer_names_[owner_layer_id] << "' with layer '"
<< layer_names_[layer_id] << "'; count mismatch. Owner layer param "
<< "shape is " << owner_blob->shape_string() << "; sharing layer "
<< "shape is " << this_blob->shape_string();
} else {
// Strict dimension checking -- all dims must be the same.
CHECK(this_blob->shape() == owner_blob->shape())
<< "Cannot share param '" << param_name << "' owned by layer '"
<< layer_names_[owner_layer_id] << "' with layer '"
<< layer_names_[layer_id] << "'; shape mismatch. Owner layer param "
<< "shape is " << owner_blob->shape_string() << "; sharing layer "
<< "expects shape " << this_blob->shape_string();
}
//获取owner layer的learnable_param_id,并且压入当前layer的向量learnable_param_ids_。
//而且在这里也没有把参数blob压入learnable_params_向量(只是将id压入learnable_param_ids_),从而避免当前layer与sharing layer之间关于shared parameter blob 的重复
const int learnable_param_id = learnable_param_ids_[owner_net_param_id];//vector<int> learnable_param_ids_ ; vector<float> params_lr_;
learnable_param_ids_.push_back(learnable_param_id);
if (param_spec->has_lr_mult()) {
if (has_params_lr_[learnable_param_id]) {
CHECK_EQ(param_spec->lr_mult(), params_lr_[learnable_param_id])
<< "Shared param '" << param_name << "' has mismatched lr_mult.";
} else {
has_params_lr_[learnable_param_id] = true;
params_lr_[learnable_param_id] = param_spec->lr_mult();
}
}
if (param_spec->has_decay_mult()) {
if (has_params_decay_[learnable_param_id]) {
CHECK_EQ(param_spec->decay_mult(),
params_weight_decay_[learnable_param_id])
<< "Shared param '" << param_name << "' has mismatched decay_mult.";
} else {
has_params_decay_[learnable_param_id] = true;
params_weight_decay_[learnable_param_id] = param_spec->decay_mult();
}
}
}
}
然后看一下Init函数
====================================Init函数================================================
template <typename Dtype>
void Net<Dtype>::Init(const NetParameter& in_param) {
CHECK(Caffe::root_solver() || root_net_)
<< "root_net_ needs to be set for all non-root solvers";
// Set phase from the state.
phase_ = in_param.state().phase();
// Filter layers based on their include/exclude rules and
// the current NetState.
NetParameter filtered_param;
FilterNet(in_param, &filtered_param);
LOG_IF(INFO, Caffe::root_solver())
<< "Initializing net from parameters: " << std::endl
<< filtered_param.DebugString();
// Create a copy of filtered_param with splits added where necessary.
NetParameter param;
InsertSplits(filtered_param, ¶m);
// Basically, build all the layers and set up their connections.
name_ = param.name();
map<string, int> blob_name_to_idx;//blob_name_to_idx是一个map,其关键字是不重复的
set<string> available_blobs;//available_blobs是一个set,其关键字是不重复的
CHECK(param.input_dim_size() == 0 || param.input_shape_size() == 0)
<< "Must specify either input_shape OR deprecated input_dim, not both.";
if (param.input_dim_size() > 0) {
// Deprecated 4D dimensions.
CHECK_EQ(param.input_size() * 4, param.input_dim_size())
<< "Incorrect input blob dimension specifications.";
} else {
CHECK_EQ(param.input_size(), param.input_shape_size())
<< "Exactly one input_shape must be specified per input.";
}
memory_used_ = 0;
// set the input blobs
for (int input_id = 0; input_id < param.input_size(); ++input_id) {
const int layer_id = -1; // inputs have fake layer ID -1
AppendTop(param, layer_id, input_id, &available_blobs, &blob_name_to_idx);//available_blobs存储blob name.blob_name_to_idx是如何初始化的?????
}
// For each layer, set up its input and output
bottom_vecs_.resize(param.layer_size());
top_vecs_.resize(param.layer_size());
bottom_id_vecs_.resize(param.layer_size());
param_id_vecs_.resize(param.layer_size());
top_id_vecs_.resize(param.layer_size());
bottom_need_backward_.resize(param.layer_size());
for (int layer_id = 0; layer_id < param.layer_size(); ++layer_id) {
// For non-root solvers, whether this layer is shared from root_net_.
bool share_from_root = !Caffe::root_solver()
&& root_net_->layers_[layer_id]->ShareInParallel();
// Inherit phase from net if unset.
if (!param.layer(layer_id).has_phase()) {
param.mutable_layer(layer_id)->set_phase(phase_);//实参phase_是网络的phase,为模板类layer设置shape_属性
}
// Setup layer.
const LayerParameter& layer_param = param.layer(layer_id);
//检查LayerParameter类型propagate_down成员的个数师傅达标
if (layer_param.propagate_down_size() > 0) {
CHECK_EQ(layer_param.propagate_down_size(),
layer_param.bottom_size())
<< "propagate_down param must be specified "
<< "either 0 or bottom_size times ";
}
//Creating Layer
if (share_from_root) {
LOG(INFO) << "Sharing layer " << layer_param.name() << " from root net";
layers_.push_back(root_net_->layers_[layer_id]);
layers_[layer_id]->SetShared(true);//调用的是模板类Layer的SetShared方法
} else {
layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param));
}
layer_names_.push_back(layer_param.name());//为layer_names_添加新元素
LOG_IF(INFO, Caffe::root_solver())
<< "Creating Layer " << layer_param.name();
bool need_backward = false;
// Figure out this layer's input and output
for (int bottom_id = 0; bottom_id < layer_param.bottom_size();
++bottom_id) {
const int blob_id = AppendBottom(param, layer_id, bottom_id,
&available_blobs, &blob_name_to_idx);
// If a blob needs backward, this layer should provide it.
need_backward |= blob_need_backward_[blob_id];在遍历所有的bottom_id的过程中,只要有一次使得need_backward为真,则这个for循环结束后,need_backward也为真。也就是说该层前一层的top blob中只要有一个blob在blob_need_backward_中为true,则backward就为true,后面的layer_need_backward_也就push_back(true)
}
int num_top = layer_param.top_size();
for (int top_id = 0; top_id < num_top; ++top_id) {
AppendTop(param, layer_id, top_id, &available_blobs, &blob_name_to_idx);//在AppendTop函数中,会为向量blob_need_backward_添加新元素
}
// If the layer specifies that AutoTopBlobs() -> true and the LayerParameter
// specified fewer than the required number (as specified by
// ExactNumTopBlobs() or MinTopBlobs()), allocate them here.
Layer<Dtype>* layer = layers_[layer_id].get();
if (layer->AutoTopBlobs()) {
const int needed_num_top =
std::max(layer->MinTopBlobs(), layer->ExactNumTopBlobs());
for (; num_top < needed_num_top; ++num_top) {
// Add "anonymous" top blobs -- do not modify available_blobs or
// blob_name_to_idx as we don't want these blobs to be usable as input
// to other layers.
AppendTop(param, layer_id, num_top, NULL, NULL);
}
}
// After this layer is connected, set it up.
if (share_from_root) {
// Set up size of top blobs using root_net_
const vector<Blob<Dtype>*>& base_top = root_net_->top_vecs_[layer_id];
const vector<Blob<Dtype>*>& this_top = this->top_vecs_[layer_id];
for (int top_id = 0; top_id < base_top.size(); ++top_id) {
this_top[top_id]->ReshapeLike(*base_top[top_id]);
LOG(INFO) << "Created top blob " << top_id << " (shape: "
<< this_top[top_id]->shape_string() << ") for shared layer "
<< layer_param.name();
}
} else {
layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]); //调用模板类layer的SetUp方法,如果在网络的定义文件里没有设置loss_weight,那么loss layer的LayerSetup函数里会设置loww_weght, 且默认值
}
LOG_IF(INFO, Caffe::root_solver())
<< "Setting up " << layer_names_[layer_id];
//每次循环,都会更新向量blob_loss_weights
for (int top_id = 0; top_id < top_vecs_[layer_id].size(); ++top_id) {
if (blob_loss_weights_.size() <= top_id_vecs_[layer_id][top_id]) {
blob_loss_weights_.resize(top_id_vecs_[layer_id][top_id] + 1, Dtype(0));
}
//top_id_vecs_中存储的最基本元素是blob_id ——> 每一个新的blob都会赋予其一个blob_id,但是这个blob_id可能是会有重复的
blob_loss_weights_[top_id_vecs_[layer_id][top_id]] = layer->loss(top_id);//loss函数返回loss_weight ——> 在模板类的SetUp方法中会调用SetLossWeights来设置其私有数据成员loss_,里面存储的其实是loss_weight
LOG_IF(INFO, Caffe::root_solver())
<< "Top shape: " << top_vecs_[layer_id][top_id]->shape_string();
if (layer->loss(top_id)) {
LOG_IF(INFO, Caffe::root_solver())
<< " with loss weight " << layer->loss(top_id);
}
memory_used_ += top_vecs_[layer_id][top_id]->count();
}
LOG_IF(INFO, Caffe::root_solver())
<< "Memory required for data: " << memory_used_ * sizeof(Dtype);
const int param_size = layer_param.param_size();
const int num_param_blobs = layers_[layer_id]->blobs().size();
//param_size是Layermeter类型对象layer_param中ParamSpec param成员的个数, num_param_blobs是一个Layer中learnable parameter blob的个数,param_size <= num_param_blobs
CHECK_LE(param_size, num_param_blobs)
<< "Too many params specified for layer " << layer_param.name();
ParamSpec default_param_spec;
for (int param_id = 0; param_id < num_param_blobs; ++param_id) {
const ParamSpec* param_spec = (param_id < param_size) ?
&layer_param.param(param_id) : &default_param_spec;
const bool param_need_backward = param_spec->lr_mult() != 0;//need backward 则为真。
need_backward |= param_need_backward;//由 param_need_backward 来决定need_backward是否为真(网络定义文件中的lr_mult很重要),并且,只要有一次遍历使得need_backward为真,则这个for循环结束后,need_backward也为真
layers_[layer_id]->set_param_propagate_down(param_id,
param_need_backward);//设定一个Layer的parameter blob 是否需要计算diff backward--->set_param_propagate_down是模板类Layer的方法。
}
for (int param_id = 0; param_id < num_param_blobs; ++param_id) {
//添加parameter blob,如果当前layer没有parameter blob(num_param_blobs==0),比如RELU,那么就不进入循环,不添加parameter blob
//AppendParam只是执行为当前layer添加parameter blob的相关工作,并不会修改与backward的相关属性
AppendParam(param, layer_id, param_id);
}
// Finally, set the backward flag
// 在这里初始化向量layer_need_backward_
layer_need_backward_.push_back(need_backward);
//在上述的AppendTop函数中,在遍历当前层的每一个top blob的时候都会将一个false(默认值)压入向量blob_need_backward_。在下面的代码中,如果这个layer need backward,则会更新blob_need_backward_
if (need_backward) {
for (int top_id = 0; top_id < top_id_vecs_[layer_id].size(); ++top_id) {
blob_need_backward_[top_id_vecs_[layer_id][top_id]] = true; //重新设置每一层的 blob_need_backward_ ,一开始是在AppendTop里将各 top blob 默认设置为false,这里根据need_backward来重新设置
}
}
}
// Go through the net backwards to determine which blobs contribute to the
// loss. We can skip backward computation for blobs that don't contribute
// to the loss. 不仅仅是确定某个layer是否需要BP,还需要确定layer的某些blob是否需要BP
// Also checks if all bottom blobs don't need backward computation (possible
// because the skip_propagate_down param) and so we can skip bacward
// computation for the entire layer
// 需要注意的是,上述代码中关于backward设置的部分,是按照前向的顺序设置的,而下面的代码是按后向顺序修正前向设置的结果。
// 一个layer是否需要backward computation,主要依据两个方面:(1)该layer的top blob 是否参与loss的计算;(2):该layer是否至少有一个 bottom blob 需要backward computation,比如Data层一般就不需要backward computation
set<string> blobs_under_loss;
set<string> blobs_skip_backp;
for (int layer_id = layers_.size() - 1; layer_id >= 0; --layer_id) {
bool layer_contributes_loss = false;
bool layer_skip_propagate_down = true;
//为true,则表示当前layer的bottom blob不需要backward computation,即该层不需要backward computation。
//这个局部变量所表示的意义与caffe.proto里message Layerparameter的propagate_down的定义恰好相反。
for (int top_id = 0; top_id < top_vecs_[layer_id].size(); ++top_id) {
const string& blob_name = blob_names_[top_id_vecs_[layer_id][top_id]];
<span style="white-space:pre"> </span>//如果在网络的定义文件文件中,没有设置loss_weight, 那么loss layer 的 LayerSetUp 方法会设置loss_weight,且默认值为1
if (layers_[layer_id]->loss(top_id) ||
(blobs_under_loss.find(blob_name) != blobs_under_loss.end())) {
layer_contributes_loss = true;
}
if (blobs_skip_backp.find(blob_name) == blobs_skip_backp.end()) {
layer_skip_propagate_down = false; //只要有一个top blob不在 blobs_skip_backp 里面,layer_skip_propagate_down 就为false,即该层不会跳过BP
}
if (layer_contributes_loss && !layer_skip_propagate_down)
break;//只是跳出当前if语句
}
// If this layer can skip backward computation, also all his bottom blobs
// don't need backpropagation
if (layer_need_backward_[layer_id] && layer_skip_propagate_down) {
layer_need_backward_[layer_id] = false;
for (int bottom_id = 0; bottom_id < bottom_vecs_[layer_id].size();
++bottom_id) {
bottom_need_backward_[layer_id][bottom_id] = false;
}
}
if (!layer_contributes_loss) { layer_need_backward_[layer_id] = false; }
if (Caffe::root_solver()) {
if (layer_need_backward_[layer_id]) {
LOG(INFO) << layer_names_[layer_id] << " needs backward computation.";
} else {
LOG(INFO) << layer_names_[layer_id]
<< " does not need backward computation.";
}
}
for (int bottom_id = 0; bottom_id < bottom_vecs_[layer_id].size();//修正前向设置的结果
++bottom_id) {
if (layer_contributes_loss) {
const string& blob_name =
blob_names_[bottom_id_vecs_[layer_id][bottom_id]];
blobs_under_loss.insert(blob_name);//为blobs_under_loss添加新元素
} else {
bottom_need_backward_[layer_id][bottom_id] = false;
}
if (!bottom_need_backward_[layer_id][bottom_id]) {
const string& blob_name =
blob_names_[bottom_id_vecs_[layer_id][bottom_id]];
blobs_skip_backp.insert(blob_name);//为blobs_skip_backp添加新元素
}
}
}
// Handle force_backward if needed.Netparameter类型的force_backward方法
if (param.force_backward()) {
for (int layer_id = 0; layer_id < layers_.size(); ++layer_id) {
layer_need_backward_[layer_id] = true;
for (int bottom_id = 0;
bottom_id < bottom_need_backward_[layer_id].size(); ++bottom_id) {
bottom_need_backward_[layer_id][bottom_id] =
bottom_need_backward_[layer_id][bottom_id] ||
layers_[layer_id]->AllowForceBackward(bottom_id);
blob_need_backward_[bottom_id_vecs_[layer_id][bottom_id]] =
blob_need_backward_[bottom_id_vecs_[layer_id][bottom_id]] ||
bottom_need_backward_[layer_id][bottom_id];
}
for (int param_id = 0; param_id < layers_[layer_id]->blobs().size();
++param_id) {
layers_[layer_id]->set_param_propagate_down(param_id, true);
}
}
}
// In the end, all remaining blobs are considered output blobs.
for (set<string>::iterator it = available_blobs.begin();
it != available_blobs.end(); ++it) {
LOG_IF(INFO, Caffe::root_solver())
<< "This network produces output " << *it;
net_output_blobs_.push_back(blobs_[blob_name_to_idx[*it]].get());
net_output_blob_indices_.push_back(blob_name_to_idx[*it]);
}
for (size_t blob_id = 0; blob_id < blob_names_.size(); ++blob_id) {
blob_names_index_[blob_names_[blob_id]] = blob_id;//第一次使用向量blob_names_index_,逐一添加元素,是一个map
}
for (size_t layer_id = 0; layer_id < layer_names_.size(); ++layer_id) {
layer_names_index_[layer_names_[layer_id]] = layer_id;//第一次使用向量layer_names_index_,逐一添加元素,是一个map
}
ShareWeights();
debug_info_ = param.debug_info();
LOG_IF(INFO, Caffe::root_solver()) << "Network initialization done.";
}
template <typename Dtype>
void Net<Dtype>::BackwardFromTo(int start, int end) {
CHECK_GE(end, 0);
CHECK_LT(start, layers_.size());
// 一般情况下,第一个卷积层conv1的propagatedown为false,即bottom_need_backward_[0]为false, 也就是说不需要求关于conv1的bottom blob的梯度,因为这些bottom blob是data,label, 它们毕竟是死的,不变的,不会随着模型的学习而改变
for (int i = start; i >= end; --i) {
if (layer_need_backward_[i]) {
layers_[i]->Backward(
top_vecs_[i], bottom_need_backward_[i], bottom_vecs_[i]);
if (debug_info_) { BackwardDebugInfo(i); }
}
}
}
以上几个函数最为主要。
然后看开其他的一些函数:
================================FilterNet函数========================================
FilterNet()给定当前phase/level/stage,移除指定层
template <typename Dtype>
void Net<Dtype>::FilterNet(const NetParameter& param,
NetParameter* param_filtered) {
NetState net_state(param.state());
param_filtered->CopyFrom(param);
param_filtered->clear_layer();
for (int i = 0; i < param.layer_size(); ++i) {
const LayerParameter& layer_param = param.layer(i);
const string& layer_name = layer_param.name();
CHECK(layer_param.include_size() == 0 || layer_param.exclude_size() == 0)
<< "Specify either include rules or exclude rules; not both.";
// If no include rules are specified, the layer is included by default and
// only excluded if it meets one of the exclude rules.
bool layer_included = (layer_param.include_size() == 0);
for (int j = 0; layer_included && j < layer_param.exclude_size(); ++j) {
if (StateMeetsRule(net_state, layer_param.exclude(j), layer_name)) {
layer_included = false;//如果不包含include,只要meet一个include_size(idx)即可
}
}
for (int j = 0; !layer_included && j < layer_param.include_size(); ++j) {
if (StateMeetsRule(net_state, layer_param.include(j), layer_name)) {
layer_included = true;//如果包含include,只要符合一个include_size(idx)即可
}
}
if (layer_included) {
param_filtered->add_layer()->CopyFrom(layer_param);
}
}
}
StateMeetsRule()中net的state是否满足NetStaterule
ForwardPrefilled()用于前馈预先填满,即预先进行一次前馈。
Forward()把网络输入层的blob读到net_input_blobs_,然后进行前馈,计算出loss。Forward的重载,只是输入层的blob以string的格式传入。
Backward()对整个网络进行反向传播。
Reshape()用于改变每层的尺寸,比如输出的feature map的size
Update()更新params_中blob的值。
ShareTrainedLayersWith(Net* other)从Other网络复制某些层 。
CopyTrainedLayersFrom()调用FromProto函数把源层的blob赋给目标层的blob。
ToProto()把网络的参数存入prototxt中。
params_lr()和params_weight_decay()学习速率和权重衰减;
blob_by_name()判断是否存在名字为blob_name的blob;