void Net<Dtype>::CopyTrainedLayersFrom(const NetParameter& param) {
int num_source_layers = param.layers_size();
for (int i = 0; i < num_source_layers; ++i) {
const LayerParameter& source_layer = param.layers(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()) {
DLOG(INFO) << "Ignoring source layer " << source_layer_name;
continue;
}
DLOG(INFO) << "Copying source layer " << source_layer_name;
vector<shared_ptr<Blob<Dtype> > >& target_blobs =
layers_[target_layer_id]->blobs();
// blob 0# weights, 1# bias term
CHECK_EQ(target_blobs.size(), source_layer.blobs_size())
<< "Incompatible number of blobs for layer " << source_layer_name;
for (int j = 0; j < target_blobs.size(); ++j) {
CHECK_EQ(target_blobs[j]->num(), source_layer.blobs(j).num());
CHECK_EQ(target_blobs[j]->channels(), source_layer.blobs(j).channels());
CHECK_EQ(target_blobs[j]->height(), source_layer.blobs(j).height());
CHECK_EQ(target_blobs[j]->width(), source_layer.blobs(j).width());
target_blobs[j]->FromProto(source_layer.blobs(j));
}
}
}
void Net<Dtype>::filterNet(const NetParameter& param, NetParameter* filtered_param){
NetState state(param.state());
filtered_param->CopyFrom(param);
// remove all layer params and then filter
filtered_param->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();
// usually a layer has not any include/exclude rules
CHECK(layer_param.include_size() == 0 || layer_param.exclude_size() == 0)
<< "Specify either include or exclude rules.";
bool layer_included = (layer_param.include_size() == 0);
// assume 'included' and check if meet any excluded rules
for (int j = 0; layer_included&&j < layer_param.exclude_size(); j++){
if (stateMeetRule(state, layer_param.exclude(j), layer_name)){
// cancel 'included'
layer_included = false;
}
}
// assume 'excluded' and check if meet any included rules
for (int j = 0; !layer_included&&j < layer_param.include_size(); j++){
if (stateMeetRule(state, layer_param.include(j), layer_name)){
// cancel 'excluded'
layer_included = true;
}
}
// copy the included layer to filtered_param
if (layer_included) filtered_param->add_layer()->CopyFrom(layer_param);
}
}
SolverParameter ModelServer<Dtype>::prepare_model() {
NetParameter net;
solver->net()->ToProto(&net);
for (int i = 0; i < net.layer_size(); ++i) {
LayerParameter& layer = *net.mutable_layer(i);
layer.clear_blobs();
if ((layer.type().find("Data") != std::string::npos)
&& (layer.has_remote_data_param())) {
layer.set_type("RemoteData");
for (int j = 0; j < layer.top_size(); ++j) {
*layer.mutable_remote_data_param()->add_shape()
= blob_shape_by_name(layer.top(j));
}
}
}
SolverParameter ret = solver->param();
ret.clear_net();
ret.clear_net_param();
ret.clear_test_net();
ret.clear_test_net_param();
ret.clear_train_net();
*ret.mutable_train_net_param() = net;
return ret;
}
void InputProducer::to_proto(NetParameter& net) const {
*net.add_input() = name_;
net.add_input_dim(dims_[0]);
net.add_input_dim(dims_[1]);
net.add_input_dim(dims_[2]);
net.add_input_dim(dims_[3]);
}
Net<Dtype>::Net(const string& param_file, Phase phase, const Net* root_net = NULL):
root_net(root_net){
NetParameter param;
readNetParamsFromTextFileOrDie(param_file, ¶m);
param.mutable_state()->set_phase(phase);
Init(param);
}
void NGNet::Init( ) {
input_layer_top_idx_ = 0;
output_layer_top_idx_ = 0;
/* Load the network. */
net_.reset(new Net<float>(model_file_, TEST));
NetParameter param;
CHECK(ReadProtoFromTextFile(model_file_, ¶m))
<< "Failed to parse NetParameter file: " << model_file_;
for (int ip = 0; ip < param.layer_size(); ip++) {
LayerParameter layer_param = param.layer(ip);
if (layer_param.has_inner_product_param()) {
InnerProductParameter* inner_product_param = layer_param.mutable_inner_product_param();
int num_output = inner_product_param->num_output();
if (num_output > 0) {
inner_product_param->set_num_output(num_output * 2);
}
}
}
// //param.mutable_state()->set_phase(phase);
Net<float> * new_net = new Net<float>(param);
net_->CopyTrainedLayersFrom(trained_file_);
int input_layer_idx = -1;
for (size_t layer_id = 0; layer_id < net_->layer_names().size(); ++layer_id) {
if (net_->layer_names()[layer_id] == input_layer_name_) {
input_layer_idx = layer_id;
break;
}
}
if (input_layer_idx == -1) {
LOG(FATAL) << "Unknown layer name " << input_layer_name_;
}
input_layer_idx_ = input_layer_idx;
input_layer_top_idx_ = 0;
Blob<float>* input_layer = net_->top_vecs()[input_layer_idx_][input_layer_top_idx_];
input_layer_dim_ = input_layer->shape(1);
int output_layer_idx = -1;
for (size_t layer_id = 0; layer_id < net_->layer_names().size(); ++layer_id) {
if (net_->layer_names()[layer_id] == output_layer_name_) {
output_layer_idx = layer_id;
break;
}
}
if (output_layer_idx == -1) {
LOG(FATAL) << "Unknown layer name " << output_layer_name_;
}
output_layer_idx_ = output_layer_idx;
}
bool NetNeedsV0ToV1Upgrade(const NetParameter& net_param) {
for (int i = 0; i < net_param.layers_size(); ++i) {
if (net_param.layers(i).has_layer()) {
return true;
}
}
return false;
}
void UpgradeV0PaddingLayers(const NetParameter& param,
NetParameter* param_upgraded_pad) {
// Copy everything other than the layers from the original param.
param_upgraded_pad->Clear();
param_upgraded_pad->CopyFrom(param);
param_upgraded_pad->clear_layers();
// Figure out which layer each bottom blob comes from.
map<string, int> blob_name_to_last_top_idx;
for (int i = 0; i < param.input_size(); ++i) {
const string& blob_name = param.input(i);
blob_name_to_last_top_idx[blob_name] = -1;
}
for (int i = 0; i < param.layers_size(); ++i) {
const V1LayerParameter& layer_connection = param.layers(i);
const V0LayerParameter& layer_param = layer_connection.layer();
// Add the layer to the new net, unless it's a padding layer.
if (layer_param.type() != "padding") {
param_upgraded_pad->add_layers()->CopyFrom(layer_connection);
}
for (int j = 0; j < layer_connection.bottom_size(); ++j) {
const string& blob_name = layer_connection.bottom(j);
if (blob_name_to_last_top_idx.find(blob_name) ==
blob_name_to_last_top_idx.end()) {
LOG(FATAL) << "Unknown blob input " << blob_name << " to layer " << j;
}
const int top_idx = blob_name_to_last_top_idx[blob_name];
if (top_idx == -1) {
continue;
}
const V1LayerParameter& source_layer = param.layers(top_idx);
if (source_layer.layer().type() == "padding") {
// This layer has a padding layer as input -- check that it is a conv
// layer or a pooling layer and takes only one input. Also check that
// the padding layer input has only one input and one output. Other
// cases have undefined behavior in Caffe.
CHECK((layer_param.type() == "conv") || (layer_param.type() == "pool"))
<< "Padding layer input to "
"non-convolutional / non-pooling layer type "
<< layer_param.type();
CHECK_EQ(layer_connection.bottom_size(), 1)
<< "Conv Layer takes a single blob as input.";
CHECK_EQ(source_layer.bottom_size(), 1)
<< "Padding Layer takes a single blob as input.";
CHECK_EQ(source_layer.top_size(), 1)
<< "Padding Layer produces a single blob as output.";
int layer_index = param_upgraded_pad->layers_size() - 1;
param_upgraded_pad->mutable_layers(layer_index)->mutable_layer()
->set_pad(source_layer.layer().pad());
param_upgraded_pad->mutable_layers(layer_index)
->set_bottom(j, source_layer.bottom(0));
}
}
for (int j = 0; j < layer_connection.top_size(); ++j) {
const string& blob_name = layer_connection.top(j);
blob_name_to_last_top_idx[blob_name] = i;
}
}
}
bool NetNeedsDataUpgrade(const NetParameter& net_param) {
for (int i = 0; i < net_param.layers_size(); ++i) {
if (net_param.layers(i).type() == LayerParameter_LayerType_DATA) {
DataParameter layer_param = net_param.layers(i).data_param();
if (layer_param.has_scale()) { return true; }
if (layer_param.has_mean_file()) { return true; }
if (layer_param.has_crop_size()) { return true; }
if (layer_param.has_mirror()) { return true; }
}
if (net_param.layers(i).type() == LayerParameter_LayerType_IMAGE_DATA) {
ImageDataParameter layer_param = net_param.layers(i).image_data_param();
if (layer_param.has_scale()) { return true; }
if (layer_param.has_mean_file()) { return true; }
if (layer_param.has_crop_size()) { return true; }
if (layer_param.has_mirror()) { return true; }
}
if (net_param.layers(i).type() == LayerParameter_LayerType_WINDOW_DATA) {
WindowDataParameter layer_param = net_param.layers(i).window_data_param();
if (layer_param.has_scale()) { return true; }
if (layer_param.has_mean_file()) { return true; }
if (layer_param.has_crop_size()) { return true; }
if (layer_param.has_mirror()) { return true; }
}
if (net_param.layers(i).type() == LayerParameter_LayerType_QDATA) {
QDataParameter layer_param = net_param.layers(i).qdata_param();
if (layer_param.has_scale()) { return true; }
if (layer_param.has_mean_file()) { return true; }
if (layer_param.has_crop_size()) { return true; }
if (layer_param.has_mirror()) { return true; }
}
}
return false;
}
bool NetNeedsBatchNormUpgrade(const NetParameter& net_param) {
for (int i = 0; i < net_param.layer_size(); ++i) {
// Check if BatchNorm layers declare three parameters, as required by
// the previous BatchNorm layer definition.
if (net_param.layer(i).type() == "BatchNorm"
&& net_param.layer(i).param_size() == 3) {
return true;
}
}
return false;
}
void Net<Dtype>::ReInit( NetParameter& param, const int batch_size )
{
layers_.clear();
layer_names_.clear();
layer_need_backward_.clear();
// blobs stores the blobs that store intermediate results between the
// layers.
blobs_.clear();
blob_names_.clear();
blob_need_backward_.clear();
// 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.
bottom_vecs_.clear();
bottom_id_vecs_.clear();
// top_vecs stores the vectors containing the output for each layer
top_vecs_.clear();
top_id_vecs_.clear();
// blob indices for the input and the output of the net
net_input_blob_indices_.clear();
net_input_blobs_.clear();
net_output_blobs_.clear();
// The parameters in the network.
params_.clear();
// the learning rate multipliers
params_lr_.clear();
// the weight decay multipliers
params_weight_decay_.clear();
param.mutable_layers(0)->mutable_layer()->set_batchsize(batch_size);
Init( param );
}
bool UpgradeV1Net(const NetParameter& v1_net_param, NetParameter* net_param) {
bool is_fully_compatible = true;
if (v1_net_param.layer_size() > 0) {
LOG(ERROR) << "Input NetParameter to be upgraded already specifies 'layer' "
<< "fields; these will be ignored for the upgrade.";
is_fully_compatible = false;
}
net_param->CopyFrom(v1_net_param);
net_param->clear_layers();
net_param->clear_layer();
for (int i = 0; i < v1_net_param.layers_size(); ++i) {
if (!UpgradeV1LayerParameter(v1_net_param.layers(i),
net_param->add_layer())) {
LOG(ERROR) << "Upgrade of input layer " << i << " failed.";
is_fully_compatible = false;
}
}
return is_fully_compatible;
}
void NetParameterToPrettyPrint(const NetParameter& param,
NetParameterPrettyPrint* pretty_param) {
pretty_param->Clear();
if (param.has_name()) {
pretty_param->set_name(param.name());
}
if (param.has_force_backward()) {
pretty_param->set_force_backward(param.force_backward());
}
for (int i = 0; i < param.input_size(); ++i) {
pretty_param->add_input(param.input(i));
}
for (int i = 0; i < param.input_dim_size(); ++i) {
pretty_param->add_input_dim(param.input_dim(i));
}
for (int i = 0; i < param.layers_size(); ++i) {
pretty_param->add_layers()->CopyFrom(param.layers(i));
}
}
bool UpgradeV1Net(const NetParameter& v1_net_param, NetParameter* net_param) {
if (v1_net_param.layer_size() > 0) {
LOG(FATAL) << "Refusing to upgrade inconsistent NetParameter input; "
<< "the definition includes both 'layer' and 'layers' fields. "
<< "The current format defines 'layer' fields with string type like "
<< "layer { type: 'Layer' ... } and not layers { type: LAYER ... }. "
<< "Manually switch the definition to 'layer' format to continue.";
}
bool is_fully_compatible = true;
net_param->CopyFrom(v1_net_param);
net_param->clear_layers();
net_param->clear_layer();
for (int i = 0; i < v1_net_param.layers_size(); ++i) {
if (!UpgradeV1LayerParameter(v1_net_param.layers(i),
net_param->add_layer())) {
LOG(ERROR) << "Upgrade of input layer " << i << " failed.";
is_fully_compatible = false;
}
}
return is_fully_compatible;
}
void Net<Dtype>::copyTrainedLayerFrom(const NetParameter& param){
int num_layers = param.layer_size();
for (int i = 0; i < num_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()) continue;
const vector < boost::shared_ptr<Blob<Dtype>>>& target_blobs = layers[target_layer_id]->getBlobs();
for (int j = 0; j < target_blobs.size(); j++){
Blob<Dtype> source_blob;
source_blob.FromProto(source_layer.blobs(j));
Blob<Dtype>* target_blob = target_blobs[j].get();
CHECK(source_blob.shape() == target_blob->shape())
<< "Incompatible shape when sharing trained params.";
target_blob->FromProto(source_layer.blobs(j), false);
}
}
}
void ApolloNet<Dtype>::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();
if (layers_map_.find(source_layer_name) == layers_map_.end()) {
LOG(INFO) << "Ignoring source layer " << source_layer_name;
continue;
}
LOG(INFO) << "Copying source layer " << source_layer_name;
vector<shared_ptr<Blob<Dtype> > >& target_blobs =
layers_map_[source_layer_name]->blobs();
ASSERT(target_blobs.size() == source_layer.blobs_size(),
"Incompatible number of blobs for layer " << source_layer_name);
for (int j = 0; j < target_blobs.size(); ++j) {
const bool kReshape = false;
target_blobs[j]->FromProto(source_layer.blobs(j), kReshape);
}
}
}
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->count(blob_name))
LOG(FATAL) << "Unknown bottom blob: " << blob_name<< " at layer: " << layer_param.name() << ".";
// a bottom blob must share a top blob
const int blob_id = (*blob_name_to_idx)[blob_name];
LOG_IF(INFO, Dragon::get_root_solver())
<< layer_param.name() << "[Layer-Accept] <- " << blob_name << " [Blob-Name]";
bottom_vecs[layer_id].push_back(blobs[blob_id].get());
bottom_id_vecs[layer_id].push_back(blob_id);
// ensure that a top blob must specify only one bottom blob
// SplitLayer can be used to shadow a top blob into several top blobs
available_blobs->erase(blob_name);
bool need_bp = true;
// default(TEST) is false
bottoms_need_backward[layer_id].push_back(need_bp & blobs_need_backward[blob_id]);
return blob_id;
}
void Net<Dtype>::Init(const NetParameter& in_param){
CHECK(Dragon::get_root_solver() || root_net)
<< "Root net need to be set for all non-root solvers.";
phase = in_param.state().phase();
NetParameter filtered_param, param;
// filter for unqualified LayerParameters(e.g Test DataLayer)
filterNet(in_param, &filtered_param);
insertSplits(filtered_param, ¶m);
name = param.name();
LOG_IF(INFO, Dragon::get_root_solver())
<< "Initialize net from parameters: ";/*<< endl << param.DebugString();*/
map<string, int> blob_name_to_idx;
set<string> available_blobs;
CHECK_EQ(param.input_size(), param.input_shape_size())<< "input blob_shape must specify a blob.";
memory_used = 0;
// check and stuff virtual input blobs firstly [Viewing Mode Only]
for (int input_id=0; input_id < param.input_size(); input_id++){
const int layer_id = -1;
// net_input.push_back(.....virtual blob.....)
appendTop(param, layer_id, input_id, &available_blobs, &blob_name_to_idx);
}
// stuff real blobs for each layer then [Traning/Testing/Viewing Mode]
bottom_vecs.resize(param.layer_size());
bottom_id_vecs.resize(param.layer_size());
bottoms_need_backward.resize(param.layer_size());
top_vecs.resize(param.layer_size());
top_id_vecs.resize(param.layer_size());
param_id_vecs.resize(param.layer_size());
for (int layer_id = 0; layer_id < param.layer_size(); layer_id++){
bool share_from_root = !Dragon::get_root_solver()
&& root_net->layers[layer_id]->shareInParallel();
// copy net phase to layer if not set
if (!param.layer(layer_id).has_phase())
param.mutable_layer(layer_id)->set_phase(phase);
const LayerParameter& layer_param = param.layer(layer_id);
if (share_from_root){
LOG(INFO) << "Share Layer: " << layer_param.name() << " from the root net.";
// share layer by pointer
layers.push_back(root_net->layers[layer_id]);
layers[layer_id]->setShared(true);
}
else{
// use layer factory to create a pointer
// layer type is referred by layer_param->type()
// see more in layer_factory.hpp
layers.push_back(LayerFactory<Dtype>::createLayer(layer_param));
}
layer_names.push_back(layer_param.name());
LOG_IF(INFO, Dragon::get_root_solver()) << "Create Layer: " << layer_param.name();
bool need_bp = false;
// stuff bottom blobs
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);
// check whether a bottom need back propogation
need_bp |= blobs_need_backward[blob_id];
}
// stuff top blobs
for (int top_id = 0; top_id < layer_param.top_size(); top_id++)
appendTop(param, layer_id, top_id, &available_blobs, &blob_name_to_idx);
// auto top blobs
// NOT_IMPLEMENTED;
Layer<Dtype>* layer = layers[layer_id].get();
// setup for layer
if (share_from_root){
const vector<Blob<Dtype>*> base_top = root_net->top_vecs[layer_id];
const vector<Blob<Dtype>*> this_top = this->top_vecs[layer_id];
// reshape solely after root_net finishing
for (int top_id = 0; top_id < base_top.size(); top_id++){
this_top[top_id]->reshapeLike(*base_top[top_id]);
}
}
else layer->setup(bottom_vecs[layer_id], top_vecs[layer_id]);
LOG_IF(INFO, Dragon::get_root_solver()) << "Setup Layer: " << layer_param.name();
for (int top_id = 0; top_id < top_vecs[layer_id].size(); top_id++){
// extend size to max number of blobs if necessary
if (blobs_loss_weight.size() <= top_id_vecs[layer_id][top_id])
blobs_loss_weight.resize(top_id_vecs[layer_id][top_id] + 1, Dtype(0));
// store global loss weights from each layer each blob
blobs_loss_weight[top_id_vecs[layer_id][top_id]] = layer->getLoss(top_id);
LOG_IF(INFO, Dragon::get_root_solver())
<< "Top shape: " << top_vecs[layer_id][top_id]->shape_string();
if (layer->getLoss(top_id)) LOG_IF(INFO, Dragon::get_root_solver())
<< " with loss weight " << layer->getLoss(top_id);
// sum up for training parameter statistic
memory_used += top_vecs[layer_id][top_id]->count();
}
LOG_IF(INFO, Dragon::get_root_solver())
<< "Memory required for Data: " << memory_used*sizeof(Dtype);
const int param_size = layer_param.param_size();
// blobs_size will be set after layer->setup()
const int param_blobs_size = layer->getBlobs().size();
CHECK_LE(param_size, param_blobs_size)<< "Too many params specify for layer.";
// use if do not specify hyperparameter
// lr_mult=decay_mult=1.0
ParamSpec default_hyperparameter;
for (int param_id = 0; param_id < param_blobs_size; param_id++){
const ParamSpec* hyperparameter = param_id < param_size ?
&layer_param.param(param_id) : &default_hyperparameter;
const bool param_need_bp = hyperparameter->lr_mult() != 0;
// check whether a param blob need back propogation [default=true]
need_bp |= param_need_bp;
layer->setParamNeedBp(param_id, param_need_bp);
}
// stuff param blobs
for (int param_id = 0; param_id < param_blobs_size; param_id++)
appendParam(param, layer_id, param_id);
// update param blobs if share others
shareWeights();
layer_need_backward.push_back(need_bp);
// after checking all bottom blobs and param blobs
if (need_bp)
for (int top_id = 0; top_id < top_id_vecs[layer_id].size(); top_id++)
blobs_need_backward[top_id_vecs[layer_id][top_id]] = true;
} // end layer_id
set<string> blobs_under_loss, blobs_skip_bp;
for (int layer_id = layers.size()-1; layer_id >= 0; layer_id--){
bool layer_contributes_loss = false;
bool layer_skip_bp = true;
Layer<Dtype>* layer = layers[layer_id].get();
for (int top_id = 0; top_id < top_vecs[layer_id].size(); top_id++){
const string& blob_name = blobs_name[top_id_vecs[layer_id][top_id]];
if (layer->getLoss(top_id) || blobs_under_loss.count(blob_name))
layer_contributes_loss = true;
if (!blobs_skip_bp.count(blob_name)) layer_skip_bp = false;
// find any top blobs if affected by loss and do not force to skip bp
if (layer_contributes_loss&&!layer_skip_bp) break;
}
// optimization trick: set lr_mult but is not affected by loss
if (layer_need_backward[layer_id] && layer_skip_bp){
// cancel layer
layer_need_backward[layer_id] = false;
// cancel bottom
for (int bottom_id = 0; bottom_id < bottom_vecs[layer_id].size(); bottom_id++){
bottoms_need_backward[layer_id][bottom_id] = false;
}
}
// cancel directly if layer is not affected by loss
if (!layer_contributes_loss) layer_need_backward[layer_id] = false;
// debug info
if (Dragon::get_root_solver()){
if (layer_need_backward[layer_id])
LOG(INFO) << "Layer: " << layer_names[layer_id] << " need back-propogation.";
else LOG(INFO) << "Layer: " << layer_names[layer_id] << " does not need back-propogation.";
}
// if one top blob affected by loss
// all bottom blobs will be affected
// regard it as "loss back-affected"
for (int bottom_id = 0; bottom_id < bottom_vecs[layer_id].size(); bottom_id++){
const string& blob_name = blobs_name[bottom_id_vecs[layer_id][bottom_id]];
if (layer_contributes_loss) blobs_under_loss.insert(blob_name);
else bottoms_need_backward[layer_id][bottom_id] = false;
// use for optimization trick : skip all bottom blobs
if (!bottoms_need_backward[layer_id][bottom_id]) blobs_skip_bp.insert(blob_name);
}
} // end layer id
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_vecs[layer_id].size(); bottom_id++){
// set for bottoms
bottoms_need_backward[layer_id][bottom_id] =
bottoms_need_backward[layer_id][bottom_id]||layers[layer_id]->allowForceBackward(bottom_id);
// set for blobs
blobs_need_backward[bottom_id_vecs[layer_id][bottom_id]] =
blobs_need_backward[bottom_id_vecs[layer_id][bottom_id]]||bottoms_need_backward[layer_id][bottom_id];
}
// set for params
for (int param_id = 0; param_id < layers[layer_id]->getBlobs().size(); param_id++){
layers[layer_id]->setParamNeedBp(param_id, true);
}
}
}
// move un-used(declare top but not use as bottom) blobs into output blobs
// usually contain loss blobs
for (set<string>::iterator i = available_blobs.begin(); i != available_blobs.end(); i++){
LOG_IF(INFO, Dragon::get_root_solver())
<< "Network produces output: " << *i;
net_output_blobs.push_back(blobs[blob_name_to_idx[*i]].get());
net_output_blob_indices.push_back(blob_name_to_idx[*i]);
}
// store blob_name -> blob_ids
blobs_name_idx = blob_name_to_idx;
// store layer_name -> layer_id
for (size_t layer_id = 0; layer_id < layer_names.size(); layer_id++)
layers_name_idx[layer_names[layer_id]] = layer_id;
debug_info = param.debug_info();
LOG_IF(INFO, Dragon::get_root_solver()) << "Network Initializion done.";
}
void Net<Dtype>::Init(const NetParameter& param) {
// Basically, build all the layers and set up its connections.
name_ = param.name();
map<string, int> blob_name_to_idx;
set<string> available_blobs;
int num_layers = param.layers_size();
CHECK_EQ(param.input_size() * 4, param.input_dim_size())
<< "Incorrect bottom blob dimension specifications.";
// set the input blobs
for (int i = 0; i < param.input_size(); ++i) {
const string& blob_name = param.input(i);
shared_ptr<Blob<Dtype> > blob_pointer(
new Blob<Dtype>(param.input_dim(i * 4),
param.input_dim(i * 4 + 1),
param.input_dim(i * 4 + 2),
param.input_dim(i * 4 + 3)));
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(param.force_backward());
net_input_blob_indices_.push_back(i);
net_input_blobs_.push_back(blob_pointer.get());
blob_name_to_idx[blob_name] = i;
available_blobs.insert(blob_name);
}
// For each layer, set up their input and output
bottom_vecs_.resize(param.layers_size());
top_vecs_.resize(param.layers_size());
bottom_id_vecs_.resize(param.layers_size());
top_id_vecs_.resize(param.layers_size());
for (int i = 0; i < param.layers_size(); ++i) {
const LayerConnection& layer_connection = param.layers(i);
const LayerParameter& layer_param = layer_connection.layer();
layers_.push_back(shared_ptr<Layer<Dtype> >(GetLayer<Dtype>(layer_param)));
layer_names_.push_back(layer_param.name());
LOG(INFO) << "Creating Layer " << layer_param.name();
bool need_backward = param.force_backward();
// Figure out this layer's input and output
for (int j = 0; j < layer_connection.bottom_size(); ++j) {
const string& blob_name = layer_connection.bottom(j);
const int blob_id = blob_name_to_idx[blob_name];
if (available_blobs.find(blob_name) == available_blobs.end()) {
LOG(FATAL) << "Unknown blob input " << blob_name <<
" to layer" << j;
}
LOG(INFO) << layer_param.name() << " <- " << blob_name;
bottom_vecs_[i].push_back(
blobs_[blob_id].get());
bottom_id_vecs_[i].push_back(blob_id);
// If a blob needs backward, this layer should provide it.
need_backward |= blob_need_backward_[blob_id];
available_blobs.erase(blob_name);
}
for (int j = 0; j < layer_connection.top_size(); ++j) {
const string& blob_name = layer_connection.top(j);
// Check if we are doing in-place computation
if (layer_connection.bottom_size() > j &&
blob_name == layer_connection.bottom(j)) {
// In-place computation
LOG(INFO) << layer_param.name() << " -> " << blob_name << " (in-place)";
available_blobs.insert(blob_name);
top_vecs_[i].push_back(
blobs_[blob_name_to_idx[blob_name]].get());
top_id_vecs_[i].push_back(blob_name_to_idx[blob_name]);
} else if (blob_name_to_idx.find(blob_name) != blob_name_to_idx.end()) {
// If we are not doing in-place computation but has duplicated blobs,
// raise an error.
LOG(FATAL) << "Duplicate blobs produced by multiple sources.";
} else {
// Normal output.
LOG(INFO) << layer_param.name() << " -> " << blob_name;
shared_ptr<Blob<Dtype> > blob_pointer(new Blob<Dtype>());
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(param.force_backward());
blob_name_to_idx[blob_name] = blob_names_.size() - 1;
available_blobs.insert(blob_name);
top_vecs_[i].push_back(blobs_[blob_names_.size() - 1].get());
top_id_vecs_[i].push_back(blob_names_.size() - 1);
}
}
// After this layer is connected, set it up.
// LOG(INFO) << "Setting up " << layer_names_[i];
layers_[i]->SetUp(bottom_vecs_[i], &top_vecs_[i]);
for (int topid = 0; topid < top_vecs_[i].size(); ++topid) {
LOG(INFO) << "Top shape: " << top_vecs_[i][topid]->channels() << " "
<< top_vecs_[i][topid]->height() << " "
<< top_vecs_[i][topid]->width();
}
// Check if this layer needs backward operation itself
for (int j = 0; j < layers_[i]->layer_param().blobs_lr_size(); ++j) {
need_backward |= (layers_[i]->layer_param().blobs_lr(j) > 0);
}
// Finally, set the backward flag
layer_need_backward_.push_back(need_backward);
if (need_backward) {
LOG(INFO) << layer_names_[i] << " needs backward computation.";
for (int j = 0; j < top_id_vecs_[i].size(); ++j) {
blob_need_backward_[top_id_vecs_[i][j]] = true;
}
} else {
LOG(INFO) << layer_names_[i] << " does not need backward computation.";
}
}
// In the end, all remaining blobs are considered output blobs.
for (set<string>::iterator it = available_blobs.begin();
it != available_blobs.end(); ++it) {
LOG(INFO) << "This network produces output " << *it;
net_output_blobs_.push_back(blobs_[blob_name_to_idx[*it]].get());
}
GetLearningRateAndWeightDecay();
LOG(INFO) << "Network initialization done.";
}
bool NetNeedsV1ToV2Upgrade(const NetParameter& net_param) {
return net_param.layers_size() > 0;
}
void insertSplits(const NetParameter& param, NetParameter* splitted_param){
splitted_param->CopyFrom(param);
splitted_param->clear_layer();
// pair<layer_idx,blob_idx>
map<string, pair<int, int> > blob_name_to_last_top_idx;
map<pair<int, int>, pair<int, int> > bottom_idx_to_source_top_idx;
map<pair<int, int>, int> top_idx_to_bottom_count;
map<pair<int, int>, float> top_idx_to_loss_weight;
map<pair<int, int>, int> top_idx_to_bottom_split_idx;
map<int, string> layer_idx_to_layer_name;
layer_idx_to_layer_name[-1] = "input";
// scan and stuff all input blobs into a virtual layer named as "input" at -1
// input blobs do not belong to any layers and we stuff them into a virtual layer
// usually use for viewing a Net(e.g: examples\cifar10\cifar10_full.prototxt
// input: "data" *** ¡û_¡û specify it as a temporary data blob ***
// input_shape{ *** ¡û_¡û specify it as shape***
// dim: 1
// dim : 3
// dim : 32
// dim : 32
// }
// pay attention: input blobs should not use in train/test prototxt
// because they are not specified vaild data sources
// you can regard them as viewing toys
for (int i = 0; i < param.input_size(); i++){
const string& blob_name = param.input(i);
blob_name_to_last_top_idx[blob_name] = make_pair(-1, i);
}
for (int i = 0; i < param.layer_size(); i++){
const LayerParameter& layer_param = param.layer(i);
// bind layer idx to layer name
layer_idx_to_layer_name[i] = layer_param.name();
// a layer has several bottom blobs(e.g DataLayer)
for (int j = 0; j < layer_param.bottom_size(); j++){
const string& blob_name = layer_param.bottom(j);
// ensure that all bottom blobs must have the same name as one top blob
if (!blob_name_to_last_top_idx.count(blob_name)){
LOG(FATAL) << "Unknown bottom blob: " << blob_name
<< " at layer: " << layer_param.name() << ".";
}
const pair<int, int>& bottom_idx = make_pair(i, j);
const pair<int, int>& top_idx = blob_name_to_last_top_idx[blob_name];
// a bottom's name must be as same as one top's name
// find a bottom's parent top (<- backward direction)
// note that top name must declare before bottom name
// or a bottom will bind to layer_{-1}
bottom_idx_to_source_top_idx[bottom_idx] = top_idx;
top_idx_to_bottom_count[top_idx]++;
}
// update top name's position for following bottom names
for (int j = 0; j < layer_param.top_size(); j++){
const string& blob_name = layer_param.top(j);
blob_name_to_last_top_idx[blob_name] = make_pair(i, j);
}
const int last_loss = min(layer_param.loss_weight_size(), layer_param.top_size());
// only work in LossLayer
for (int j = 0; j < last_loss; j++){
const string& blob_name = layer_param.top(j);
// updated before
const pair<int, int>& top_idx = blob_name_to_last_top_idx[blob_name];
top_idx_to_loss_weight[top_idx] = layer_param.loss_weight(j);
// from loss(top) backward to bottom
if (top_idx_to_loss_weight[top_idx]) top_idx_to_bottom_count[top_idx]++;
}
}
// special case: data blob shared by other blobs in the virtual layer
// split it also
for (int i = 0; i < param.input_size(); i++){
const int split_count = top_idx_to_bottom_count[make_pair(-1, i)];
if (split_count > 1){
// "input"
const string& layer_name = layer_idx_to_layer_name[-1];
const string& blob_name = param.input(i);
// push_back a new param
LayerParameter* split_layer_param = splitted_param->add_layer();
const float kZeroLossWeight = 0;
configureSplitLayer(layer_name, blob_name, i, split_count, kZeroLossWeight, split_layer_param);
}
}
for (int i = 0; i < param.layer_size(); i++){
// push_back a new param
LayerParameter* layer_param = splitted_param->add_layer();
layer_param->CopyFrom(param.layer(i));
for (int j = 0; j < layer_param->bottom_size(); j++){
// call the top before bottom
const pair<int, int>& top_idx = bottom_idx_to_source_top_idx[make_pair(i, j)];
// check top's count
const int split_count = top_idx_to_bottom_count[top_idx];
if (split_count > 1){
// previous layer_name
const string& layer_name = layer_idx_to_layer_name[top_idx.first];
const string& blob_name = layer_param->bottom(j);
// e.g: conv1 => conv1_conv1_0_split_0
// once used then ++ for next
layer_param->set_bottom(j, splitBlobName(layer_name, blob_name, top_idx.second,
top_idx_to_bottom_split_idx[top_idx]++));
}
}
for (int j = 0; j < layer_param->top_size(); j++){
const pair<int, int>& top_idx = make_pair(i, j);
const int split_count = top_idx_to_bottom_count[top_idx];
if (split_count > 1){
// now layer_name
const string& layer_name = layer_idx_to_layer_name[top_idx.first];
const string& blob_name = layer_param->top(j);
// add a split layer
LayerParameter *split_layer_param = splitted_param->add_layer();
const float loss_weight = top_idx_to_loss_weight[top_idx];
configureSplitLayer(layer_name, blob_name, j, split_count, loss_weight,split_layer_param);
if (loss_weight){
layer_param->clear_loss_weight();
// loss as bottom should split from 1 ???
top_idx_to_bottom_split_idx[top_idx]++;
}
}
}
}
}
void InsertSplits(const NetParameter& param, NetParameter* param_split) {
// Initialize by copying from the input NetParameter.
param_split->CopyFrom(param);
param_split->clear_layer();
map<string, pair<int, int> > blob_name_to_last_top_idx;
map<pair<int, int>, pair<int, int> > bottom_idx_to_source_top_idx;
map<pair<int, int>, int> top_idx_to_bottom_count;
map<pair<int, int>, float> top_idx_to_loss_weight;
map<pair<int, int>, int> top_idx_to_bottom_split_idx;
map<int, string> layer_idx_to_layer_name;
for (int i = 0; i < param.layer_size(); ++i) {
const LayerParameter& layer_param = param.layer(i);
layer_idx_to_layer_name[i] = layer_param.name();
for (int j = 0; j < layer_param.bottom_size(); ++j) {
const string& blob_name = layer_param.bottom(j);
if (blob_name_to_last_top_idx.find(blob_name) ==
blob_name_to_last_top_idx.end()) {
LOG(FATAL) << "Unknown bottom blob '" << blob_name << "' (layer '"
<< layer_param.name() << "', bottom index " << j << ")";
}
const pair<int, int>& bottom_idx = make_pair(i, j);
const pair<int, int>& top_idx = blob_name_to_last_top_idx[blob_name];
bottom_idx_to_source_top_idx[bottom_idx] = top_idx;
++top_idx_to_bottom_count[top_idx];
}
for (int j = 0; j < layer_param.top_size(); ++j) {
const string& blob_name = layer_param.top(j);
blob_name_to_last_top_idx[blob_name] = make_pair(i, j);
}
// A use of a top blob as a loss should be handled similarly to the use of
// a top blob as a bottom blob to another layer.
const int last_loss =
std::min(layer_param.loss_weight_size(), layer_param.top_size());
for (int j = 0; j < last_loss; ++j) {
const string& blob_name = layer_param.top(j);
const pair<int, int>& top_idx = blob_name_to_last_top_idx[blob_name];
top_idx_to_loss_weight[top_idx] = layer_param.loss_weight(j);
if (top_idx_to_loss_weight[top_idx]) {
++top_idx_to_bottom_count[top_idx];
}
}
}
for (int i = 0; i < param.layer_size(); ++i) {
LayerParameter* layer_param = param_split->add_layer();
layer_param->CopyFrom(param.layer(i));
// Replace any shared bottom blobs with split layer outputs.
for (int j = 0; j < layer_param->bottom_size(); ++j) {
const pair<int, int>& top_idx =
bottom_idx_to_source_top_idx[make_pair(i, j)];
const int split_count = top_idx_to_bottom_count[top_idx];
if (split_count > 1) {
const string& layer_name = layer_idx_to_layer_name[top_idx.first];
const string& blob_name = layer_param->bottom(j);
layer_param->set_bottom(j, SplitBlobName(layer_name,
blob_name, top_idx.second, top_idx_to_bottom_split_idx[top_idx]++));
}
}
// Create split layer for any top blobs used by other layer as bottom
// blobs more than once.
for (int j = 0; j < layer_param->top_size(); ++j) {
const pair<int, int>& top_idx = make_pair(i, j);
const int split_count = top_idx_to_bottom_count[top_idx];
if (split_count > 1) {
const string& layer_name = layer_idx_to_layer_name[i];
const string& blob_name = layer_param->top(j);
LayerParameter* split_layer_param = param_split->add_layer();
const float loss_weight = top_idx_to_loss_weight[top_idx];
ConfigureSplitLayer(layer_name, blob_name, j, split_count,
loss_weight, split_layer_param);
if (loss_weight) {
layer_param->clear_loss_weight();
top_idx_to_bottom_split_idx[top_idx]++;
}
}
}
}
}
void BaseProducer::to_proto (NetParameter& net) const {
*net.add_layers() = layer_param_;
}
void Net<Dtype>::Init(const NetParameter& in_param) {
// Create a copy of in_param with splits added where necessary.
NetParameter param;
InsertSplits(in_param, ¶m);
// Basically, build all the layers and set up its connections.
name_ = param.name();
map<string, int> blob_name_to_idx;
set<string> available_blobs;
int num_layers = param.layers_size();
CHECK_EQ(param.input_size() * 4, param.input_dim_size())
<< "Incorrect bottom blob dimension specifications.";
size_t memory_used = 0;
// set the input blobs
for (int i = 0; i < param.input_size(); ++i) {
const string& blob_name = param.input(i);
shared_ptr<Blob<Dtype> > blob_pointer(
new Blob<Dtype>(param.input_dim(i * 4),
param.input_dim(i * 4 + 1),
param.input_dim(i * 4 + 2),
param.input_dim(i * 4 + 3)));
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(param.force_backward());
net_input_blob_indices_.push_back(i);
net_input_blobs_.push_back(blob_pointer.get());
blob_name_to_idx[blob_name] = i;
available_blobs.insert(blob_name);
memory_used += blob_pointer->count();
}
DLOG(INFO) << "Memory required for Data" << memory_used*sizeof(Dtype);
// For each layer, set up their input and output
bottom_vecs_.resize(param.layers_size());
top_vecs_.resize(param.layers_size());
bottom_id_vecs_.resize(param.layers_size());
top_id_vecs_.resize(param.layers_size());
for (int i = 0; i < param.layers_size(); ++i) {
bool in_place = false;
const LayerParameter& layer_param = param.layers(i);
layers_.push_back(shared_ptr<Layer<Dtype> >(GetLayer<Dtype>(layer_param)));
layer_names_.push_back(layer_param.name());
LOG(INFO) << "Creating Layer " << layer_param.name();
bool need_backward = param.force_backward();
// Figure out this layer's input
for (int j = 0; j < layer_param.bottom_size(); ++j) {
const string& blob_name = layer_param.bottom(j);
const int blob_id = blob_name_to_idx[blob_name];
if (available_blobs.find(blob_name) == available_blobs.end()) {
LOG(FATAL) << "Unknown blob input " << blob_name <<
" to layer" << j;
}
LOG(INFO) << layer_param.name() << " <- " << blob_name;
bottom_vecs_[i].push_back(blobs_[blob_id].get());
bottom_id_vecs_[i].push_back(blob_id);
// If a blob needs backward, this layer should provide it.
need_backward |= blob_need_backward_[blob_id];
available_blobs.erase(blob_name);
}
// Figure out this layer's output
for (int j = 0; j < layer_param.top_size(); ++j) {
const string& blob_name = layer_param.top(j);
// Check if we are doing in-place computation
if (layer_param.bottom_size() > j &&
blob_name == layer_param.bottom(j)) {
// In-place computation
LOG(INFO) << layer_param.name() << " -> " << blob_name << " (in-place)";
in_place = true;
available_blobs.insert(blob_name);
top_vecs_[i].push_back(
blobs_[blob_name_to_idx[blob_name]].get());
top_id_vecs_[i].push_back(blob_name_to_idx[blob_name]);
} else if (blob_name_to_idx.find(blob_name) != blob_name_to_idx.end()) {
// If we are not doing in-place computation but has duplicated blobs,
// raise an error.
LOG(FATAL) << "Duplicate blobs produced by multiple sources.";
} else {
// Normal output.
LOG(INFO) << layer_param.name() << " -> " << blob_name;
shared_ptr<Blob<Dtype> > blob_pointer(new Blob<Dtype>());
blobs_.push_back(blob_pointer);
blob_names_.push_back(blob_name);
blob_need_backward_.push_back(param.force_backward());
blob_name_to_idx[blob_name] = blob_names_.size() - 1;
available_blobs.insert(blob_name);
top_vecs_[i].push_back(blobs_[blob_names_.size() - 1].get());
top_id_vecs_[i].push_back(blob_names_.size() - 1);
}
}
// After this layer is connected, set it up.
//LOG(INFO) << "Setting up " << layer_names_[i];
layers_[i]->SetUp(bottom_vecs_[i], &(top_vecs_[i]));
for (int topid = 0; topid < top_vecs_[i].size(); ++topid) {
LOG(INFO) << "Top shape: " << top_vecs_[i][topid]->num() << " "
<< top_vecs_[i][topid]->channels() << " "
<< top_vecs_[i][topid]->height() << " "
<< top_vecs_[i][topid]->width() << " ("
<< top_vecs_[i][topid]->count() << ")";
if (!in_place)
memory_used += top_vecs_[i][topid]->count();
}
DLOG(INFO) << "Memory required for Data " << memory_used*sizeof(Dtype);
// blobs: 0# weights, 1# bias term; blob_lr: 1# learning rate for weights, 2# learning rate for bias
int blobs_lr_size = layers_[i]->layer_param().blobs_lr_size();
CHECK(blobs_lr_size == layers_[i]->blobs().size() || blobs_lr_size == 0) // 0, 1, 2
<< "Incorrect blobs lr size: should be either 0 or the same as "
"the number of the layer's parameter blobs.";
if (blobs_lr_size) {
// Check if this layer needs backward operation itself
for (int j = 0; j < blobs_lr_size; ++j) {
need_backward |= (layers_[i]->layer_param().blobs_lr(j) > 0);
}
} else if (layers_[i]->blobs().size()) {
// catch: if a layer param does not specify blobs_lr, we should assume the
// learning rate to be 1. Thus we will need to perform backward.
need_backward = true;
}
// Finally, set the backward flag
layer_need_backward_.push_back(need_backward);
if (need_backward) {
LOG(INFO) << layer_names_[i] << " needs backward computation.";
for (int j = 0; j < top_id_vecs_[i].size(); ++j) {
blob_need_backward_[top_id_vecs_[i][j]] = true;
}
} else {
LOG(INFO) << layer_names_[i] << " does not need backward computation.";
}
}
// In the end, all remaining blobs are considered output blobs.
for (set<string>::iterator it = available_blobs.begin();
it != available_blobs.end(); ++it) {
LOG(INFO) << "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 i = 0; i < blob_names_.size(); ++i) {
blob_names_index_[blob_names_[i]] = i;
}
for (size_t i = 0; i < layer_names_.size(); ++i) {
layer_names_index_[layer_names_[i]] = i;
}
GetLearningRateAndWeightDecay();
LOG(INFO) << "Network initialization done.";
LOG(INFO) << "Memory required for Data " << memory_used*sizeof(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){
boost::shared_ptr<LayerParameter> layer_param(
layer_id >= 0 ? new LayerParameter(param.layer(layer_id)) : NULL);
// use (layer_id//top_id) or (-1//top_id) to get a blob name
const string& blob_name = layer_param ?
(top_id<layer_param->top_size() ? layer_param->top(top_id) : "(automatic)") : param.input(top_id);
// in-place case (e.g:
// I0721 10:38 : 16.722070 4692 net.cpp : 84] relu1 <-conv1
// I0721 10:38 : 16.722082 4692 net.cpp : 98] relu1->conv1(in-place)
// check a blob whether at the same postion in both bottom and top
if (blob_name_to_idx && layer_param && top_id < layer_param->bottom_size()
&& blob_name == layer_param->bottom(top_id)){
LOG_IF(INFO, Dragon::get_root_solver())
<< layer_param->name() << "[Layer-Produce]->" << blob_name << " [Blob-Name] (in-place)";
// add into this layer's top blob using blob_name
top_vecs[layer_id].push_back(blobs[(*blob_name_to_idx)[blob_name]].get());
// log the id
top_id_vecs[layer_id].push_back((*blob_name_to_idx)[blob_name]);
}
else if (blob_name_to_idx && (*blob_name_to_idx).count(blob_name) ){
LOG(FATAL) << "Top blob:" << blob_name << " propogate from multiple sources.";
}
// normal top blob stuffing
else{
// debug info
if (Dragon::get_root_solver()){
if (layer_param) LOG(INFO) << layer_param->name() << "[Layer-Produce] ->" << blob_name << " [Blob-Name]";
// special case and only used when viewing a Net's structure
// because they need not specify data source and can not train or test
// virtual data input blobs do not belong to any layers
// see more in insert_splits.cpp/void InsertSplits()
else LOG(INFO) << "Input " << top_id << "[Blob-Code] -> " << blob_name << "[Blob - Name]";
}
// allocate a null blob at first
boost::shared_ptr<Blob<Dtype>> ptr_blob(new Blob<Dtype>());
// store global blob infos
const int blob_id = blobs.size();
blobs.push_back(ptr_blob);
blobs_name.push_back(blob_name);
blobs_need_backward.push_back(false);
// encode index number for a name
// which also represent this top blob is binded from a bottom
// check it before can know whether a top blob has multiple sources(Forbidden)
if (blob_name_to_idx) (*blob_name_to_idx)[blob_name] = blob_id;
// reshape for virtual input blobs solely
// becaude they do not exist into a DataLayer(provide reshape/transfrom service)
if (layer_id == -1){
ptr_blob->reshape(param.input_shape(top_id));
// store solely for virtual input blobs
net_input_blobs.push_back(ptr_blob.get());
net_input_blob_indices.push_back(blob_id);
}
else{
top_vecs[layer_id].push_back(ptr_blob.get());
top_id_vecs[layer_id].push_back(blob_id);
}
}
// a set used for listing all exsiting top blobs
if (available_blobs) available_blobs->insert(blob_name);
}
void InsertSplits(const NetParameter& param, NetParameter* param_split) {
// Initialize by copying from the input NetParameter.
param_split->CopyFrom(param);
param_split->clear_layers();
map<string, pair<int, int> > blob_name_to_last_top_idx;
map<pair<int, int>, pair<int, int> > bottom_idx_to_source_top_idx;
map<pair<int, int>, int> top_idx_to_bottom_count;
map<pair<int, int>, int> top_idx_to_bottom_split_idx;
map<int, string> layer_idx_to_layer_name;
layer_idx_to_layer_name[-1] = "input";
// Determine the number of times each blob is used as an input (bottom) blob.
for (int i = 0; i < param.input_size(); ++i) {
const string& blob_name = param.input(i);
blob_name_to_last_top_idx[blob_name] = make_pair(-1, i);
}
for (int i = 0; i < param.layers_size(); ++i) {
const LayerParameter& layer_param = param.layers(i);
layer_idx_to_layer_name[i] = layer_param.name();
for (int j = 0; j < layer_param.bottom_size(); ++j) {
const string& blob_name = layer_param.bottom(j);
if (blob_name_to_last_top_idx.find(blob_name) ==
blob_name_to_last_top_idx.end()) {
LOG(FATAL) << "Unknown blob input " << blob_name << " to layer " << j;
}
const pair<int, int>& bottom_idx = make_pair(i, j);
const pair<int, int>& top_idx = blob_name_to_last_top_idx[blob_name];
bottom_idx_to_source_top_idx[bottom_idx] = top_idx;
++top_idx_to_bottom_count[top_idx];
}
for (int j = 0; j < layer_param.top_size(); ++j) {
const string& blob_name = layer_param.top(j);
blob_name_to_last_top_idx[blob_name] = make_pair(i, j);
}
}
// Create split layer for any input blobs used by other layers as bottom
// blobs more than once.
for (int i = 0; i < param.input_size(); ++i) {
const int split_count = top_idx_to_bottom_count[make_pair(-1, i)];
if (split_count > 1) {
const string& layer_name = layer_idx_to_layer_name[-1];
const string& blob_name = param.input(i);
LayerParameter* split_layer_param = param_split->add_layers();
ConfigureSplitLayer(layer_name, blob_name, i, split_count,
split_layer_param);
}
}
for (int i = 0; i < param.layers_size(); ++i) {
LayerParameter* layer_param = param_split->add_layers();
layer_param->CopyFrom(param.layers(i));
// Replace any shared bottom blobs with split layer outputs.
for (int j = 0; j < layer_param->bottom_size(); ++j) {
const pair<int, int>& top_idx =
bottom_idx_to_source_top_idx[make_pair(i, j)];
const int split_count = top_idx_to_bottom_count[top_idx];
if (split_count > 1) {
const string& layer_name = layer_idx_to_layer_name[top_idx.first];
const string& blob_name = layer_param->bottom(j);
layer_param->set_bottom(j, SplitBlobName(layer_name,
blob_name, top_idx.second, top_idx_to_bottom_split_idx[top_idx]++));
}
}
// Create split layer for any top blobs used by other layers as bottom
// blobs more than once.
for (int j = 0; j < layer_param->top_size(); ++j) {
const int split_count = top_idx_to_bottom_count[make_pair(i, j)];
if (split_count > 1) {
const string& layer_name = layer_idx_to_layer_name[i];
const string& blob_name = layer_param->top(j);
LayerParameter* split_layer_param = param_split->add_layers();
ConfigureSplitLayer(layer_name, blob_name, j, split_count,
split_layer_param);
}
}
}
}
// Generate inception network specifications.
int main(int argc, char** argv) {
string prefix(argv[1]);
// Deployment network spec.
string deploy_fname = prefix + "_deploy.prototxt";
NetParameter deploy;
deploy.set_name("Inception");
deploy.add_input("data");
deploy.add_input_dim(10);
deploy.add_input_dim(3);
deploy.add_input_dim(227);
deploy.add_input_dim(227);
AddBody(&deploy);
AddSoftmaxLayer("prob", "linear", deploy.add_layers());
WriteProtoToTextFile(deploy, deploy_fname);
// Training network spec.
string train_fname = prefix + "_train.prototxt";
NetParameter train;
train.set_name("Inception");
AddDataLayer("data", "label", traindb, meanfile, 64, 227, true,
train.add_layers());
AddBody(&train);
AddSoftmaxLoss("prob", "linear", "label", train.add_layers());
WriteProtoToTextFile(train, train_fname);
// Validation network spec.
string val_fname = prefix + "_val.prototxt";
NetParameter val;
val.set_name("Inception");
AddDataLayer("data", "label", valdb, meanfile, 50, 227, false,
val.add_layers());
AddBody(&val);
AddSoftmaxLayer("prob", "linear", val.add_layers());
AddAccuracyLayer("accuracy", "prob", "label", val.add_layers());
WriteProtoToTextFile(val, val_fname);
}
void Net<Dtype>::appendParam(const NetParameter& param, const int layer_id, const int param_id){
const LayerParameter& layer_param = param.layer(layer_id);
Layer<Dtype>* layer = layers[layer_id].get();
const int param_size = layer_param.param_size();
// default name="" (not set)
string param_name = param_id<param_size? layer_param.param(param_id).name() : "";
// has name
if (param_name.size()) param_display_names.push_back(param_name);
// set param_id as name
else{
ostringstream display_name;
display_name << param_id;
param_display_names.push_back(display_name.str());
}
// each param blob has a net id(both weight and bias)
const int net_param_id = param_blobs.size();
// add param blob which can be used by a net id
param_blobs.push_back(layer->getBlobs()[param_id]);
// store a net id
// param_id_vecs[layer_id][param_id] can get the net_param_id
param_id_vecs[layer_id].push_back(net_param_id);
// store orginal id ( x_th layer/ y_th param )
// param_layer_indices[net_param_id] can get layer_id/param_id
param_layer_indices.push_back(make_pair(layer_id, param_id));
ParamSpec default_hyperparameter;
const ParamSpec* hyperparameter = param_id < param_size ?
&layer_param.param(param_id) : &default_hyperparameter;
// do not have a name or
if (!param_size || !param_name.size() ||
(param_name.size() && !param_names_index.count(param_name))){
param_owners.push_back(-1);
// has a name(non-empty) but has not logged before
if (param_name.size()) param_names_index[param_name] = net_param_id;
const int learnable_param_id = learnable_params.size();
learnable_params.push_back(param_blobs[net_param_id].get());
learnable_param_ids.push_back(learnable_param_id);
has_params_lr.push_back(hyperparameter->has_lr_mult());
has_params_decay.push_back(hyperparameter->has_decay_mult());
params_lr.push_back(hyperparameter->lr_mult());
params_decay.push_back(hyperparameter->decay_mult());
}
else{
// has a name(non-empty) and has logged before
// it means to share this param and we need get the owner id
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, Dragon::get_root_solver())
<< "Share parameter: " << param_name << " ownd by layer: "
<< layer_names[owner_layer_id] << " param index: " << owner_layer_id;
Blob<Dtype>* this_blob = param_blobs[net_param_id].get();
Blob<Dtype>* owner_blob = param_blobs[owner_net_param_id].get();
CHECK(this_blob);CHECK(owner_blob);
// check before sharing
if (layer_param.param(param_id).share_mode() == ParamSpec_DimCheckMode_PERMISSIVE_MODE)
CHECK_EQ(this_blob->count(), owner_blob->count());
else CHECK(this_blob->shape() == owner_blob->shape());
// note that learnable_param_id = owner_net_param_id
const int learnable_param_id = learnable_param_ids[owner_net_param_id];
// store parent id
learnable_param_ids.push_back(learnable_param_id);
// check lr_mult
if (hyperparameter->has_lr_mult()){
if (has_params_lr[learnable_param_id])
CHECK_EQ(hyperparameter->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] = hyperparameter->lr_mult();
}
}
// check decay_mult
if (hyperparameter->has_decay_mult()){
if (has_params_decay[learnable_param_id])
CHECK_EQ(hyperparameter->decay_mult(), params_decay[learnable_param_id])
<< "Shared param: " << param_name << " has mismatched decay_mult.";
else{
has_params_decay[learnable_param_id] = true;
params_decay[learnable_param_id] = hyperparameter->decay_mult();
}
}
}
}
