Caffe解读（三）
Caffe学习（三）
Caffe 梳理
- layer.hpp中：Forward和Backward对应前向计算和反向更新，输入统一都是bottom，输出为top，其中Backward里面有个propagate_down参数，用来表示该Layer是否反向传播梯度。在Forward和Backward的具体实现里，会根据Caffe::mode()进行对应的操作，即使用cpu或者gpu进行计算，两个都实现了对应的接口Forward_cpu、Forward_gpu和Backward_cpu、Backward_gpu，这些接口都是virtual，具体还是要根据layer的类型进行对应的计算（注意：有些layer并没有GPU计算的实现，所以封装时加入了CPU的计算作为后备）。另外，还实现了ToProto的接口，将Layer的参数写入到protocol buffer文件中。
- 在梳理Net.cpp代码后，我们将详细说明propagate_down的具体使用方法，以及在finetune中决定是否反传的各个参数的区别
Net-Init
Net.cpp-Init函数
- 代码梳理
#include ::Net(const NetParameter& param) {
Init(param);
}
template <typename Dtype>
Net::Net(const string& param_file, Phase phase,
const int level, const vector<string>* stages) {
NetParameter param;
ReadNetParamsFromTextFileOrDie(param_file, ¶m);
// Set phase, stages and level
param.mutable_state()->set_phase(phase);
if (stages != NULL) {
for (int i = 0; i < stages->size(); i++) {
param.mutable_state()->add_stage((*stages)[i]);
}
}
param.mutable_state()->set_level(level);
Init(param);
}
template <typename Dtype>
void Net::Init(const NetParameter& in_param) {
// Set phase from the state.
phase_ = in_param.state().phase();
// Filter layers based on their include/exclude rules and
// the current NetState.
// FilterNet
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.
// 第二次处理Net 的param，作用：对于底层一个输出blob对应多个上层的情况，要再加入Split 层，
// 形成新的网络，这么做的主要原因是多个层反传给该blob的梯度需要累加
// 如Label需传入accuracy 层和loss 层，就会在这里插入一层，也就是caffe 的log 中打印出来的
// 一个新的层label_mnist_1_split 层，为该层创建两个top blob 分别为：
// Label_mnist_1_split_0 和Label_mnist_1_split_1 分别传入accuracy 层和loss
NetParameter param;
InsertSplits(filtered_param, ¶m);
// Basically, build all the layers and set up their connections.
// 下面是层及层间blob的创建
name_ = param.name();
map<string, int> blob_name_to_idx;
set<string> available_blobs;
memory_used_ = 0;
// For each layer, set up its input and output
bottom_vecs_.resize(param.layer_size());// 存每一层的输入(bottom)blob指针
top_vecs_.resize(param.layer_size());// 存每一层输出(top)的blob指针
bottom_id_vecs_.resize(param.layer_size());// 存每一层输入(bottom)blob的id
param_id_vecs_.resize(param.layer_size());// 存每一层参数blob的id
top_id_vecs_.resize(param.layer_size());// 存每一层输出(top)的blob的id
bottom_need_backward_.resize(param.layer_size());//该blob是需要返回的bool值
//for循环对每一层处理
for (int layer_id = 0; layer_id < param.layer_size(); ++layer_id) {
// Inherit phase from net if unset.
// 如果当前层没有设置phase，则将当前层phase设置为网络net的phase
if (!param.layer(layer_id).has_phase()) {
param.mutable_layer(layer_id)->set_phase(phase_);
}
// Setup layer.
const LayerParameter& layer_param = param.layer(layer_id);//当前层的参数
// 检查LayerParameter 类型propagate_down 成员的个数是否达标
// 也就是我们在prototxt 中设置的“propagate_down: ××”，他的个数要么是0（也就是不写）
// 要么等于bottom blob 的个数，代表当前layer的梯度是否反向传播（后续有详细解释）
// indicating whether to compute the diff of each param blob.
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 ";
}
// 创建一个具体的层，并压入到layers_ 中
layers_.push_back(LayerRegistry::CreateLayer(layer_param));
layer_names_.push_back(layer_param.name());
LOG_IF(INFO, Caffe::root_solver())
<< "Creating Layer " << layer_param.name();
bool need_backward = false;
// Figure out this layer's input and output
// 分别处理输入（bottom blob）和输出（top blob）
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);
// AppendBottom() 函数为该层创建bottom blob
// If a blob needs backward, this layer should provide it.
need_backward |= blob_need_backward_[blob_id];
}
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);
// 创建top blob压入top_vecs_
// Collect Input layer tops as Net inputs.
if (layer_param.type() == "Input") {
const int blob_id = blobs_.size() - 1;
net_input_blob_indices_.push_back(blob_id);
net_input_blobs_.push_back(blobs_[blob_id].get());
}
}
// 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* 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.
// Set_Up中为AppendTop() 中创建的Blob 分配内存空间
layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]);
LOG_IF(INFO, Caffe::root_solver())
<< "Setting up " << layer_names_[layer_id];
// 每次循环都会更新
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));
}
//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
blob_loss_weights_[top_id_vecs_[layer_id][top_id]] = layer->loss(top_id);
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();// 计算所需内存
}
LOG_IF(INFO, Caffe::root_solver())
<< "Memory required for data: " << memory_used_ * sizeof(Dtype);
// 对每层的param blob的处理
const int param_size = layer_param.param_size();
// param_size 是Layermeter 类型对象layer_param 中ParamSpec param 成员的个数,
// 是层内blob_ 的数量，即该层有几个权重参数（每个blob内有一个参数）
// 例如；cov层和IP层都有两个参数对应w和b
const int num_param_blobs = layers_[layer_id]->blobs().size();
// 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;
// 这里说明了如果在prototxt 中将lr置为0，即关掉，该层参数便不再更新
need_backward |= param_need_backward;
// 由param_need_backward 来决定need_backward 是否为真，
// 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);
}
// 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
for (int param_id = 0; param_id < num_param_blobs; ++param_id) {
AppendParam(param, layer_id, param_id);
// 将param blob以及blob的id添加到params_,param_id_vecs_等
}
// Finally, set the backward flag
// 之前在AppendTop 函数中，在遍历当前层的每一个top blob的时候
// 都已将一个false（默认值）压入向量blob_need_backward_中
// 下面如果这个layer need backward，则会更新blob_need_backward_
layer_need_backward_.push_back(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;
}
}
}
// 至此，各个层都已被创建并启动，下面部分是按照反向顺序修正backward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.
// 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，
// 而下面的代码是按反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的bottom blob 是否需要backward computation，
// 比如Data层一般就不需要backward computation
set<string> blobs_under_loss;
set<string> blobs_skip_backp;
// 反向
for (int layer_id = layers_.size() - 1; layer_id >= 0; --layer_id) {
bool layer_contributes_loss = false;// 这就是上面所说的标志layer是否参与loss计算
bool layer_skip_propagate_down = true;//标志layer是否需要 “跳过backward compute”
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]];
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;
}
if (layer_contributes_loss && !layer_skip_propagate_down)
break;
}
// 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;
// 所含项标志整个网络所有网络层的bottom blob 是否需要backward
}
}
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.";
}
}
// 修正前向设置的结果
for (int bottom_id = 0; bottom_id < bottom_vecs_[layer_id].size();
++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);// 添加到blobs_under_loss
} 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);// 添加到blobs_skip_backp
}
}
}
// Handle force_backward if needed.
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;
}
for (size_t layer_id = 0; layer_id < layer_names_.size(); ++layer_id) {
layer_names_index_[layer_names_[layer_id]] = layer_id;
}
ShareWeights();
debug_info_ = param.debug_info();
LOG_IF(INFO, Caffe::root_solver()) << "Network initialization done.";
}
template <typename Dtype>
void Net::FilterNet(const NetParameter& param,
NetParameter* param_filtered) {
// 作用：把模型参数文件中不符合当前阶段规则的层去掉
// 如Test阶段只用网络的前向，需要将设置为phase：Train的层去掉
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();
// include 和 exclude 分别代表什么？
// Rules controlling whether and when a layer is included in the network,
// based on the current NetState. You may specify a non-zero number of rules
// to include OR exclude, but not both. If no include or exclude rules are
// specified, the layer is always included. If the current NetState meets
// ANY (i.e., one or more) of the specified rules, the layer is
// included/excluded.
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);// 默认true
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（exclude），只要meet一个exclude rules就false
}
}
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meet一个include rules就true
}
}
if (layer_included) { //如果是true（包含），就添加进去
param_filtered->add_layer()->CopyFrom(layer_param);
}
}
}
template <typename Dtype>
bool Net::StateMeetsRule(const NetState& state,
const NetStateRule& rule, const string& layer_name) {
// Check whether the rule is broken due to phase.
if (rule.has_phase()) {
if (rule.phase() != state.phase()) {
LOG_IF(INFO, Caffe::root_solver())
<< "The NetState phase (" << state.phase()
<< ") differed from the phase (" << rule.phase()
<< ") specified by a rule in layer " << layer_name;
return false;
}
}
// Check whether the rule is broken due to min level.
if (rule.has_min_level()) {
if (state.level() < rule.min_level()) {
LOG_IF(INFO, Caffe::root_solver())
<< "The NetState level (" << state.level()
<< ") is above the min_level (" << rule.min_level()
<< ") specified by a rule in layer " << layer_name;
return false;
}
}
// Check whether the rule is broken due to max level.
if (rule.has_max_level()) {
if (state.level() > rule.max_level()) {
LOG_IF(INFO, Caffe::root_solver())
<< "The NetState level (" << state.level()
<< ") is above the max_level (" << rule.max_level()
<< ") specified by a rule in layer " << layer_name;
return false;
}
}
// Check whether the rule is broken due to stage. The NetState must
// contain ALL of the rule's stages to meet it.
for (int i = 0; i < rule.stage_size(); ++i) {
// Check that the NetState contains the rule's ith stage.
bool has_stage = false;
for (int j = 0; !has_stage && j < state.stage_size(); ++j) {
if (rule.stage(i) == state.stage(j)) { has_stage = true; }
}
if (!has_stage) {
LOG_IF(INFO, Caffe::root_solver())
<< "The NetState did not contain stage '" << rule.stage(i)
<< "' specified by a rule in layer " << layer_name;
return false;
}
}
// Check whether the rule is broken due to not_stage. The NetState must
// contain NONE of the rule's not_stages to meet it.
for (int i = 0; i < rule.not_stage_size(); ++i) {
// Check that the NetState contains the rule's ith not_stage.
bool has_stage = false;
for (int j = 0; !has_stage && j < state.stage_size(); ++j) {
if (rule.not_stage(i) == state.stage(j)) { has_stage = true; }
}
if (has_stage) {
LOG_IF(INFO, Caffe::root_solver())
<< "The NetState contained a not_stage '" << rule.not_stage(i)
<< "' specified by a rule in layer " << layer_name;
return false;
}
}
return true;
}
// Helper for Net::Init: add a new top blob to the net.
// 此函数为该层创建top blob，该函数真正的new的一个blob的对象
// 并将top blob 的指针压入到top_vecs_中
template <typename Dtype>
void Net::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 layer_param(
new LayerParameter(param.layer(layer_id)));
const string& blob_name = (layer_param->top_size() > top_id) ?
layer_param->top(top_id) : "(automatic)";
// Check if we are doing in-place computation
if (blob_name_to_idx && 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()) {
LOG(INFO) << layer_param->name() << " -> " << blob_name;
}
shared_ptr > blob_pointer(new Blob());
const int blob_id = blobs_.size();
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(false);
if (blob_name_to_idx) { (*blob_name_to_idx)[blob_name] = blob_id; }
top_id_vecs_[layer_id].push_back(blob_id);
top_vecs_[layer_id].push_back(blob_pointer.get());
}
if (available_blobs) { available_blobs->insert(blob_name); }
}
// Helper for Net::Init: add a new bottom blob to the net.
// AppendBottom()函数为该层创建bottom blob，由于网络是堆叠而成，
// 即：当前层的输入bottom blob是前一层的输出top blob，因此此函数并没没有真正的创建blob，
// 只是在将前一层的指针压入到了bottom_vecs_中
template <typename Dtype>
int Net::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];
LOG_IF(INFO, Caffe::root_solver())
<< layer_names_[layer_id] << " <- " << blob_name;
bottom_vecs_[layer_id].push_back(blobs_[blob_id].get());
bottom_id_vecs_[layer_id].push_back(blob_id);
available_blobs->erase(blob_name);
bool need_backward = blob_need_backward_[blob_id];
// Check if the backpropagation on bottom_id should be skipped
if (layer_param.propagate_down_size() > 0) {
need_backward = layer_param.propagate_down(bottom_id);
}
bottom_need_backward_[layer_id].push_back(need_backward);
return blob_id;
}
template <typename Dtype>
void Net::AppendParam(const NetParameter& param, const int layer_id,
const int param_id) {
const LayerParameter& layer_param = layers_[layer_id]->layer_param();
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);
} else {
ostringstream param_display_name;
param_display_name << param_id;
param_display_names_.push_back(param_display_name.str());
}
const int net_param_id = params_.size();
params_.push_back(layers_[layer_id]->blobs()[param_id]);
param_id_vecs_[layer_id].push_back(net_param_id);
param_layer_indices_.push_back(make_pair(layer_id, param_id));
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.
param_owners_.push_back(-1);
if (param_name.size()) {
param_names_index_[param_name] = net_param_id;
}
const int learnable_param_id = learnable_params_.size();
learnable_params_.push_back(params_[net_param_id].get());
learnable_param_ids_.push_back(learnable_param_id);
has_params_lr_.push_back(param_spec->has_lr_mult());
has_params_decay_.push_back(param_spec->has_decay_mult());
params_lr_.push_back(param_spec->lr_mult());
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];
param_owners_.push_back(owner_net_param_id);
const pair<int, int>& owner_index =
param_layer_indices_[owner_net_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* this_blob = layers_[layer_id]->blobs()[param_id].get();
Blob* owner_blob =
layers_[owner_layer_id]->blobs()[owner_param_id].get();
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();
}
const int learnable_param_id = learnable_param_ids_[owner_net_param_id];
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();
}
}
}
}
template <typename Dtype>
Dtype Net::ForwardFromTo(int start, int end) {
CHECK_GE(start, 0);
CHECK_LT(end, layers_.size());
Dtype loss = 0;
for (int i = start; i <= end; ++i) {
for (int c = 0; c < before_forward_.size(); ++c) {
before_forward_[c]->run(i);
}
Dtype layer_loss = layers_[i]->Forward(bottom_vecs_[i], top_vecs_[i]);
loss += layer_loss;
if (debug_info_) { ForwardDebugInfo(i); }
for (int c = 0; c < after_forward_.size(); ++c) {
after_forward_[c]->run(i);
}
}
return loss;
}
template <typename Dtype>
Dtype Net::ForwardFrom(int start) {
return ForwardFromTo(start, layers_.size() - 1);
}
template <typename Dtype>
Dtype Net::ForwardTo(int end) {
return ForwardFromTo(0, end);
}
template <typename Dtype>
const vector::Forward(Dtype* loss) {
if (loss != NULL) {
*loss = ForwardFromTo(0, layers_.size() - 1);
} else {
ForwardFromTo(0, layers_.size() - 1);
}
return net_output_blobs_;
}
template <typename Dtype>
const vector::Forward(
const vector::BackwardFromTo(int start, int end) {
CHECK_GE(end, 0);
CHECK_LT(start, layers_.size());
for (int i = start; i >= end; --i) {
for (int c = 0; c < before_backward_.size(); ++c) {
before_backward_[c]->run(i);
}
if (layer_need_backward_[i]) {
layers_[i]->Backward(
top_vecs_[i], bottom_need_backward_[i], bottom_vecs_[i]);
if (debug_info_) { BackwardDebugInfo(i); }
}
for (int c = 0; c < after_backward_.size(); ++c) {
after_backward_[c]->run(i);
}
}
}
template <typename Dtype>
void Net::ForwardDebugInfo(const int layer_id) {
for (int top_id = 0; top_id < top_vecs_[layer_id].size(); ++top_id) {
const Blob& blob = *top_vecs_[layer_id][top_id];
const string& blob_name = blob_names_[top_id_vecs_[layer_id][top_id]];
const Dtype data_abs_val_mean = blob.asum_data() / blob.count();
LOG_IF(INFO, Caffe::root_solver())
<< " [Forward] "
<< "Layer " << layer_names_[layer_id]
<< ", top blob " << blob_name
<< " data: " << data_abs_val_mean;
}
for (int param_id = 0; param_id < layers_[layer_id]->blobs().size();
++param_id) {
const Blob& blob = *layers_[layer_id]->blobs()[param_id];
const int net_param_id = param_id_vecs_[layer_id][param_id];
const string& blob_name = param_display_names_[net_param_id];
const Dtype data_abs_val_mean = blob.asum_data() / blob.count();
LOG_IF(INFO, Caffe::root_solver())
<< " [Forward] "
<< "Layer " << layer_names_[layer_id]
<< ", param blob " << blob_name
<< " data: " << data_abs_val_mean;
}
}
template <typename Dtype>
void Net::BackwardDebugInfo(const int layer_id) {
const vector& blob = *bottom_vec[bottom_id];
const string& blob_name = blob_names_[bottom_id_vecs_[layer_id][bottom_id]];
const Dtype diff_abs_val_mean = blob.asum_diff() / blob.count();
LOG_IF(INFO, Caffe::root_solver())
<< " [Backward] "
<< "Layer " << layer_names_[layer_id]
<< ", bottom blob " << blob_name
<< " diff: " << diff_abs_val_mean;
}
for (int param_id = 0; param_id < layers_[layer_id]->blobs().size();
++param_id) {
if (!layers_[layer_id]->param_propagate_down(param_id)) { continue; }
const Blob& blob = *layers_[layer_id]->blobs()[param_id];
const Dtype diff_abs_val_mean = blob.asum_diff() / blob.count();
LOG_IF(INFO, Caffe::root_solver())
<< " [Backward] "
<< "Layer " << layer_names_[layer_id]
<< ", param blob " << param_id
<< " diff: " << diff_abs_val_mean;
}
}
template <typename Dtype>
void Net::UpdateDebugInfo(const int param_id) {
const Blob& blob = *params_[param_id];
const int param_owner = param_owners_[param_id];
const string& layer_name = layer_names_[param_layer_indices_[param_id].first];
const string& param_display_name = param_display_names_[param_id];
const Dtype diff_abs_val_mean = blob.asum_diff() / blob.count();
if (param_owner < 0) {
const Dtype data_abs_val_mean = blob.asum_data() / blob.count();
LOG_IF(INFO, Caffe::root_solver())
<< " [Update] Layer " << layer_name
<< ", param " << param_display_name
<< " data: " << data_abs_val_mean
<< "; diff: " << diff_abs_val_mean;
} else {
const string& owner_layer_name =
layer_names_[param_layer_indices_[param_owner].first];
LOG_IF(INFO, Caffe::root_solver())
<< " [Update] Layer " << layer_name
<< ", param blob " << param_display_name
<< " (owned by layer " << owner_layer_name << ", " << "param "
<< param_display_names_[param_owners_[param_id]] << ")"
<< " diff: " << diff_abs_val_mean;
}
}
// 功能：从Other网络复制某些层
// 步骤：对Other网络的第i层（源层）：
// 1. 定义一个Layer的指针指向第i层
// 2. 读取第i层（源层）的名字
// 3. 找通过名字来找目标层如果没找到，即target_layer_id == layer_names_.size()则忽略Other的第i层，
// 即Other的第i层不需要share给网络
// 4. 如果找到了，即other的第i层需要share给网络，则把目标层的所有blob读到target_blobs中
// 1判断目标层和源层的blob数量是否相等
// 2判断每个blob大小是否相等
// 3调用ShareData函数把源层的blob赋给目标层的blob
template <typename Dtype>
void Net::ShareTrainedLayersWith(const Net* other) {
int num_source_layers = other->layers().size();
for (int i = 0; i < num_source_layers; ++i) {
Layer* source_layer = other->layers()[i].get();
const string& source_layer_name = other->layer_names()[i];
int target_layer_id = 0;
while (target_layer_id != layer_names_.size() &&
layer_names_[target_layer_id] != source_layer_name) {
++target_layer_id;
}
if (target_layer_id == layer_names_.size()) {
LOG(INFO) << "Ignoring source layer " << source_layer_name;
continue;
}
DLOG(INFO) << "Copying source layer " << source_layer_name;
vector<shared_ptr* source_blob = source_layer->blobs()[j].get();
CHECK(target_blobs[j]->shape() == source_blob->shape())
<< "Cannot share param " << j << " weights from layer '"
<< source_layer_name << "'; shape mismatch. Source param shape is "
<< source_blob->shape_string() << "; target param shape is "
<< target_blobs[j]->shape_string();
target_blobs[j]->ShareData(*source_blob);
}
}
}
template <typename Dtype>
void Net::BackwardFrom(int start) {
BackwardFromTo(start, 0);
}
template <typename Dtype>
void Net::BackwardTo(int end) {
BackwardFromTo(layers_.size() - 1, end);
}
// 对整个网络进行反向传播
template <typename Dtype>
void Net::Backward() {
BackwardFromTo(layers_.size() - 1, 0);
if (debug_info_) {
Dtype asum_data = 0, asum_diff = 0, sumsq_data = 0, sumsq_diff = 0;
for (int i = 0; i < learnable_params_.size(); ++i) {
asum_data += learnable_params_[i]->asum_data();
asum_diff += learnable_params_[i]->asum_diff();
sumsq_data += learnable_params_[i]->sumsq_data();
sumsq_diff += learnable_params_[i]->sumsq_diff();
}
const Dtype l2norm_data = std::sqrt(sumsq_data);
const Dtype l2norm_diff = std::sqrt(sumsq_diff);
LOG(ERROR) << " [Backward] All net params (data, diff): "
<< "L1 norm = (" << asum_data << ", " << asum_diff << "); "
<< "L2 norm = (" << l2norm_data << ", " << l2norm_diff << ")";
}
}
template <typename Dtype>
void Net::Reshape() {
for (int i = 0; i < layers_.size(); ++i) {
layers_[i]->Reshape(bottom_vecs_[i], top_vecs_[i]);
}
}
// 功能：和ShareTrainedLayersWith一样
// 步骤：不同的是调用FromProto函数把源层的blob赋给目标层的blob
template <typename Dtype>
void Net::CopyTrainedLayersFrom(const NetParameter& param) {
int num_source_layers = param.layer_size();
for (int i = 0; i < num_source_layers; ++i) {
const LayerParameter& source_layer = param.layer(i);
const string& source_layer_name = source_layer.name();
int target_layer_id = 0;
while (target_layer_id != layer_names_.size() &&
layer_names_[target_layer_id] != source_layer_name) {
++target_layer_id;
}
if (target_layer_id == layer_names_.size()) {
LOG(INFO) << "Ignoring source layer " << source_layer_name;
continue;
}
DLOG(INFO) << "Copying source layer " << source_layer_name;
vector<shared_ptr source_blob;
const bool kReshape = true;
source_blob.FromProto(source_layer.blobs(j), kReshape);
LOG(FATAL) << "Cannot copy param " << j << " weights from layer '"
<< source_layer_name << "'; shape mismatch. Source param shape is "
<< source_blob.shape_string() << "; target param shape is "
<< target_blobs[j]->shape_string() << ". "
<< "To learn this layer's parameters from scratch rather than "
<< "copying from a saved net, rename the layer.";
}
const bool kReshape = false;
target_blobs[j]->FromProto(source_layer.blobs(j), kReshape);
}
}
}
template <typename Dtype>
void Net::CopyTrainedLayersFrom(const string trained_filename) {
if (trained_filename.size() >= 3 &&
trained_filename.compare(trained_filename.size() - 3, 3, ".h5") == 0) {
CopyTrainedLayersFromHDF5(trained_filename);
} else {
CopyTrainedLayersFromBinaryProto(trained_filename);
}
}
template <typename Dtype>
void Net::CopyTrainedLayersFromBinaryProto(
const string trained_filename) {
NetParameter param;
ReadNetParamsFromBinaryFileOrDie(trained_filename, ¶m);
CopyTrainedLayersFrom(param);
}
template <typename Dtype>
void Net::CopyTrainedLayersFromHDF5(const string trained_filename) {
hid_t file_hid = H5Fopen(trained_filename.c_str(), H5F_ACC_RDONLY,
H5P_DEFAULT);
CHECK_GE(file_hid, 0) << "Couldn't open " << trained_filename;
hid_t data_hid = H5Gopen2(file_hid, "data", H5P_DEFAULT);
CHECK_GE(data_hid, 0) << "Error reading weights from " << trained_filename;
int num_layers = hdf5_get_num_links(data_hid);
for (int i = 0; i < num_layers; ++i) {
string source_layer_name = hdf5_get_name_by_idx(data_hid, i);
if (!layer_names_index_.count(source_layer_name)) {
LOG(INFO) << "Ignoring source layer " << source_layer_name;
continue;
}
int target_layer_id = layer_names_index_[source_layer_name];
DLOG(INFO) << "Copying source layer " << source_layer_name;
vector<shared_ptr::ToProto(NetParameter* param, bool write_diff) const {
param->Clear();
param->set_name(name_);
// Add bottom and top
DLOG(INFO) << "Serializing " << layers_.size() << " layers";
for (int i = 0; i < layers_.size(); ++i) {
LayerParameter* layer_param = param->add_layer();
layers_[i]->ToProto(layer_param, write_diff);
}
}
template <typename Dtype>
void Net::ToHDF5(const string& filename, bool write_diff) const {
hid_t file_hid = H5Fcreate(filename.c_str(), H5F_ACC_TRUNC, H5P_DEFAULT,
H5P_DEFAULT);
CHECK_GE(file_hid, 0)
<< "Couldn't open " << filename << " to save weights.";
hid_t data_hid = H5Gcreate2(file_hid, "data", H5P_DEFAULT, H5P_DEFAULT,
H5P_DEFAULT);
CHECK_GE(data_hid, 0) << "Error saving weights to " << filename << ".";
hid_t diff_hid = -1;
if (write_diff) {
diff_hid = H5Gcreate2(file_hid, "diff", H5P_DEFAULT, H5P_DEFAULT,
H5P_DEFAULT);
CHECK_GE(diff_hid, 0) << "Error saving weights to " << filename << ".";
}
for (int layer_id = 0; layer_id < layers_.size(); ++layer_id) {
const LayerParameter& layer_param = layers_[layer_id]->layer_param();
string layer_name = layer_param.name();
hid_t layer_data_hid = H5Gcreate2(data_hid, layer_name.c_str(),
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
CHECK_GE(layer_data_hid, 0)
<< "Error saving weights to " << filename << ".";
hid_t layer_diff_hid = -1;
if (write_diff) {
layer_diff_hid = H5Gcreate2(diff_hid, layer_name.c_str(),
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
CHECK_GE(layer_diff_hid, 0)
<< "Error saving weights to " << filename << ".";
}
int num_params = layers_[layer_id]->blobs().size();
for (int param_id = 0; param_id < num_params; ++param_id) {
ostringstream dataset_name;
dataset_name << param_id;
const int net_param_id = param_id_vecs_[layer_id][param_id];
if (param_owners_[net_param_id] == -1) {
// Only save params that own themselves
hdf5_save_nd_dataset(layer_data_hid, dataset_name.str(),
*params_[net_param_id]);
}
if (write_diff) {
// Write diffs regardless of weight-sharing
hdf5_save_nd_dataset(layer_diff_hid, dataset_name.str(),
*params_[net_param_id], true);
}
}
H5Gclose(layer_data_hid);
if (write_diff) {
H5Gclose(layer_diff_hid);
}
}
H5Gclose(data_hid);
if (write_diff) {
H5Gclose(diff_hid);
}
H5Fclose(file_hid);
}
// 更新params_中blob的值
template <typename Dtype>
void Net::Update() {
for (int i = 0; i < learnable_params_.size(); ++i) {
learnable_params_[i]->Update();
}
}
template <typename Dtype>
void Net::ClearParamDiffs() {
for (int i = 0; i < learnable_params_.size(); ++i) {
Blob* blob = learnable_params_[i];
switch (Caffe::mode()) {
case Caffe::CPU:
caffe_set(blob->count(), static_cast(0),
blob->mutable_cpu_diff());
break;
case Caffe::GPU:
#ifndef CPU_ONLY
caffe_gpu_set(blob->count(), static_cast(0),
blob->mutable_gpu_diff());
#else
NO_GPU;
#endif
break;
}
}
}
template <typename Dtype>
void Net::ShareWeights() {
for (int i = 0; i < params_.size(); ++i) {
if (param_owners_[i] < 0) { continue; }
params_[i]->ShareData(*params_[param_owners_[i]]);
params_[i]->ShareDiff(*params_[param_owners_[i]]);
}
}
template <typename Dtype>
bool Net::has_blob(const string& blob_name) const {
return blob_names_index_.find(blob_name) != blob_names_index_.end();
}
template <typename Dtype>
const shared_ptr > Net::blob_by_name(
const string& blob_name) const {
shared_ptr > blob_ptr;
if (has_blob(blob_name)) {
blob_ptr = blobs_[blob_names_index_.find(blob_name)->second];
} else {
blob_ptr.reset((Blob*)(NULL));
LOG(WARNING) << "Unknown blob name " << blob_name;
}
return blob_ptr;
}
template <typename Dtype>
bool Net::has_layer(const string& layer_name) const {
return layer_names_index_.find(layer_name) != layer_names_index_.end();
}
template <typename Dtype>
const shared_ptr > Net::layer_by_name(
const string& layer_name) const {
shared_ptr > layer_ptr;
if (has_layer(layer_name)) {
layer_ptr = layers_[layer_names_index_.find(layer_name)->second];
} else {
layer_ptr.reset((Layer*)(NULL));
LOG(WARNING) << "Unknown layer name " << layer_name;
}
return layer_ptr;
}
INSTANTIATE_CLASS(Net);
} // namespace caffe
propagate_down具体使用方法及决定是否反传的各个参数的区别
假设我们有4个卷积层A->B->C->D
propagate_down
- 我们希望C层的参数不改变，以及C前面的A、B层的参数也不改变，这种情况也就是D层的梯度不往前反向传播到D层的输入blob（也就是C层的输出blob 没有得到梯度），我们可以通过设置D层的propagate_down为false来做到 propagate_down的数量与输入blob的数量相同，假如你某个层有2个输入blob，那么你应该在该layer的Param里面写上两行：
propagate_down : 0 # 第1个输入blob不会得到反向传播的梯度 propagate_down : 0 # 第2个输入blob不会得到反向传播的梯度
这样，你这个layer的梯度就不会反向传播，前面的所有layer的参数也就不会改变了 propagate_down的设置使得前面的所有层都不能得到反传梯度，即前面所有层的参数都不更新
具体使用方法
layer{
name: "conv2_3x3"
type: "ConvolutionData"
bottom: "pool1_3x3"
top: "conv2_3x3"
param{
lr_mult: 1
decay_mult: 1
}
propagate_down: 0
convolution_param {
num_output: 64
bias_term: false
pad: 1
kernel_size: 3
stride: 1
weight_filler {
type: "xavier"
}
}
}lr_mult
- 我们希望C层的参数不会改变，但是C前面的A、B层的参数需要改变，这种情况，只是固定了C层的参数，C层得到的梯度依然会反向传播给前面的B层。只需要将对应的参数blob的学习率调整为0：在layer里面加上 “ param { lr_mult: 0 }” 就可以了
- lr_mult的设置只是使得该层的学习率为零该层参数不更新，梯度依然会反向传播回去
- 具体使用方法
layer{
name: "ip1"
type: "InnerProduct"
bottom: "pool3"
top: "ip1"
param{ # 对应第1个参数blob的配置，也就是全连接层的参数矩阵的配置
lr_mult: 0 # 学习率为0
decay_mult: 0
}
param{ # 对应第2个参数blob的配置，也就是全连接层的偏置项的配置
lr_mult: 0 # 学习率为0
decay_mult: 0
}
inner_product_param {
num_output: 2
weight_filler {
type: "gaussian"
std: 0.01
}
bias_filler {
type: "constant"
value: 0
}
}
}参考
https://www.zhihu.com/question/35754716
http://blog.csdn.net/langb2014/article/details/50987593
http://blog.csdn.net/qq_16055159/article/details/45057297