bool ConvolutionLayer::init() {
if(initialized) return true;
if((layer_param_.blobs_size() < 1) ||
(layer_param_.convolution_param().bias_term() &&
layer_param_.blobs_size() < 2)) return false;
LOG(INFO) << "Starting to initialize " << this->name() << endl;
const BlobProto& weights_blob = layer_param_.blobs(0);
LOG(INFO) << "CONV WEIGHTS: " << endl << weights_blob.width() << endl
<< weights_blob.height() << endl << weights_blob.channels() << endl
<< weights_blob.num() << endl;
weights_host_ = new float[weights_blob.width()*weights_blob.height()*
weights_blob.channels()*weights_blob.num()];
Buffer weights_buf(type_of<float>(), weights_blob.width(), weights_blob.height(),
weights_blob.channels(), weights_blob.num(),
(uint8_t*) (weights_host_));
for(int x = 0, x_end = weights_blob.width(); x < x_end; x++) {
for(int y = 0, y_end = weights_blob.height(); y < y_end; y++) {
for(int c = 0, c_end = weights_blob.channels(); c < c_end; c++) {
for(int n = 0, n_end = weights_blob.num(); n < n_end; n++) {
int index = x + y*weights_blob.width() + c*weights_blob.width()*
weights_blob.height() + n*weights_blob.width()*
weights_blob.height()*weights_blob.channels();
weights_host_[index] = weights_blob.data(index);
}
}
}
}
if(!weights_.defined())
weights_ = ImageParam(weights_buf.type(), 4);
weights_.set(weights_buf);
if(layer_param_.convolution_param().bias_term()) {
const BlobProto bias_blob = layer_param_.blobs(1);
bias_host_ = new float[bias_blob.width()*bias_blob.height()*
bias_blob.channels()*bias_blob.num()];
Buffer bias_buf(type_of<float>(), bias_blob.width(), bias_blob.height(),
bias_blob.channels(), bias_blob.num(),(uint8_t*) bias_host_);
LOG(INFO) << "CONV BIAS" << endl << bias_blob.width() << endl
<< bias_blob.height() << endl << bias_blob.channels() << endl
<< bias_blob.num() << endl;
int bias_size = bias_blob.width()*bias_blob.height()*bias_blob.channels()*
bias_blob.num();
for(int i = 0; i < bias_size; i++) {
bias_host_[i] = bias_blob.data(i);
}
if(!bias_.defined())
bias_ = ImageParam(bias_buf.type(), 4);
bias_.set(bias_buf);
}
LOG(INFO) << "Completed initializing " << this->name() << endl;
initialized = true;
return initialized;
}
void Blob<Dtype>::FromProto(const BlobProto& proto, bool reshape) {
if (reshape) {
vector<int> shape;
if (proto.has_num() || proto.has_channels() ||
proto.has_height() || proto.has_width()) {
// Using deprecated 4D Blob dimensions --
// shape is (num, channels, height, width).
shape.resize(4);
shape[0] = proto.num();
shape[1] = proto.channels();
shape[2] = proto.height();
shape[3] = proto.width();
} else {
shape.resize(proto.shape().dim_size());
for (int i = 0; i < proto.shape().dim_size(); ++i) {
shape[i] = proto.shape().dim(i);
}
}
Reshape(shape);
} else {
CHECK(ShapeEquals(proto)) << "shape mismatch (reshape not set)";
}
// copy data
Dtype* data_vec = mutable_cpu_data();
for (int i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i);
}
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (int i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i);
}
}
}
void Blob<Dtype>::FromProto(const BlobProto& proto, const bool init_ps_table) {
Reshape(proto.num(), proto.channels(), proto.height(), proto.width());
if (blob_mode_ == BlobProto_BlobMode_GLOBAL) {
if (init_ps_table) { // initialize ps table
// update values in ps table
Dtype* data_vec = ReadPSTable(0);
for (int i = 0; i < count_; ++i) {
data_vec[i] = data_vec[i] - proto.data(i);
}
diff_->set_cpu_data(data_vec);
UpdatePSTable();
//TODO: 2016-6-16
//// fetch the newest values
//data_vec = ReadPSTable(0);
//data_->set_cpu_ps_data(data_vec);
}
} else {
//copy data
Dtype* data_vec = mutable_cpu_data();
for (int i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i);
}
}
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (int i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i);
}
}
}
void Blob<Dtype>::FromProto(const BlobProto& proto) {
LOG(INFO)<<"FromProto size = "<<proto.num() <<" "<<proto.channels() << " "<<proto.height()<<" "<< proto.width()<<" "<< proto.depth();
if(proto.depth() ==0)
{
LOG(INFO)<< "proto depth is 0, converting to 1 for 2D models to 3D models...";
Reshape(proto.num(), proto.channels(), proto.height(), proto.width(), 1);
}else{
Reshape(proto.num(), proto.channels(), proto.height(), proto.width(), proto.depth());
}
// copy data
Dtype* data_vec = mutable_cpu_data();
/* for testing only
Dtype data_vec_sum=0;
for (int i = 0; i < count_; ++i) {
data_vec_sum=data_vec_sum+data_vec[i];
}
LOG(INFO)<<"bolb sum value = "<<data_vec_sum;
*/
for (size_t i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i);
//LOG(INFO)<<"proto.data(i) ="<<proto.data(i);
}
//LOG(INFO)<<"proto.data[11870779] ="<<proto.data(11870779);
//sleep(20);
// LOG(INFO)<<"proto.diff_size= "<<proto.diff_size();
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (size_t i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i);
}
}
//if (Caffe::mode()==Caffe::GPU) {
// gpu_data();
// }
}
void Blob<Dtype>::FromProto(const BlobProto& proto) {
Reshape(proto.num(), proto.channels(), proto.height(), proto.width());
// copy data
Dtype* data_vec = mutable_cpu_data();
for (int i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i);
}
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (int i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i);
}
}
}
void Blob<Dtype>::FromProtoWithExtraCopy(const BlobProto& proto){
// reshape other dimention but remain the same number of channels;
Reshape(proto.num(), this->channels(), proto.height(), proto.width(), proto.depth());
// copy data
Dtype* data_vec = mutable_cpu_data();
int sourceDataSize =proto.data_size();
for (int i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i%sourceDataSize);
}
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (int i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i%sourceDataSize);
}
}
}
void Blob<Dtype>::FromProtoFC2Conv(const BlobProto& proto){
// Note that we do not change the shape of taget layer blob (convolution)
// reshape other dimention but remain the same number of channels;
//Reshape(proto.num(), this->channels(), proto.height(), proto.width(), proto.depth());
// copy data
Dtype* data_vec = mutable_cpu_data();
size_t proto_count =proto.num()*proto.channels()*proto.height()*proto.width()*proto.depth();
CHECK_EQ(proto_count,count_);
for (size_t i = 0; i < count_; ++i) {
data_vec[i] = proto.data(i);
}
// LOG(INFO)<<"proto.diff_size= "<<proto.diff_size();
if (proto.diff_size() > 0) {
Dtype* diff_vec = mutable_cpu_diff();
for (size_t i = 0; i < count_; ++i) {
diff_vec[i] = proto.diff(i);
}
}
}
bool Blob<Dtype>::ShapeEquals(const BlobProto& other) {
if (other.has_num() || other.has_channels() || other.has_height()
|| other.has_width()) {
// Using deprecated 4D Blob dimensions --
// shape is (num, channels, height, width).
// Note: we do not use the normal Blob::num(), Blob::channels(), etc.
// methods as these index from the beginning of the blob shape, where legacy
// parameter blobs were indexed from the end of the blob shape (e.g., bias
// Blob shape (1 x 1 x 1 x N), IP layer weight Blob shape (1 x 1 x M x N)).
return shape_.size() <= 4 && LegacyShape(-4) == other.num()
&& LegacyShape(-3) == other.channels()
&& LegacyShape(-2) == other.height()
&& LegacyShape(-1) == other.width();
}
vector<int> other_shape(other.shape().dim_size());
for (int i = 0; i < other.shape().dim_size(); ++i) {
other_shape[i] = other.shape().dim(i);
}
return shape_ == other_shape;
}