void BNLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// Figure out the dimensions
N_ = bottom[0]->num();
C_ = bottom[0]->channels();
H_ = bottom[0]->height();
W_ = bottom[0]->width();
var_eps_ = 1e-9;
// Check if we need to set up the weights
if (this->blobs_.size() > 0) {
LOG(INFO) << "Skipping parameter initialization";
} else {
this->blobs_.resize(2);
// fill scale with scale_filler
this->blobs_[0].reset(new Blob<Dtype>(1, C_, 1, 1));
shared_ptr<Filler<Dtype> > scale_filler(GetFiller<Dtype>(
this->layer_param_.bn_param().scale_filler()));
scale_filler->Fill(this->blobs_[0].get());
// fill shift with shift_filler
this->blobs_[1].reset(new Blob<Dtype>(1, C_, 1, 1));
shared_ptr<Filler<Dtype> > shift_filler(GetFiller<Dtype>(
this->layer_param_.bn_param().shift_filler()));
shift_filler->Fill(this->blobs_[1].get());
} // parameter initialization
this->param_propagate_down_.resize(this->blobs_.size(), true);
}
void Scale2Layer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top){
num_output = this->layer_param_.convolution_param().num_output();
vector<int> scale_shape(1, bottom[0]->channels());
if (this->blobs_.size() > 0){
if (scale_shape != this->blobs_[0]->shape()){
Blob<Dtype> scale_shaped_blob(scale_shape);
LOG(FATAL) << "Incorrect weight shape: expected shape "
<< scale_shaped_blob.shape_string() << "; instead, shape was "
<< this->blobs_[0]->shape_string();
}
}
else{
this->blobs_.resize(2);
}
this->blobs_[0].reset(new Blob<Dtype>(scale_shape));
shared_ptr<Filler<Dtype> > scale_filler(GetFiller<Dtype>(
this->layer_param_.convolution_param().weight_filler()));
scale_filler->Fill(this->blobs_[0].get());
this->blobs_[1].reset(new Blob<Dtype>(scale_shape));
caffe_set(this->blobs_[1]->count(), (Dtype)-1, this->blobs_[1]->mutable_cpu_data());
}
void BNLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
// Figure out the dimensions
num_ = bottom[0]->num();
channels_ = bottom[0]->channels();
height_ = bottom[0]->height();
width_ = bottom[0]->width();
// extract param
var_eps_ = this->layer_param_.bn_param().var_eps();
decay_ = this->layer_param_.bn_param().decay();
moving_average_ = this->layer_param_.bn_param().moving_average();
// Check if we need to set up the weights
if (this->blobs_.size() > 0) {
LOG(INFO) << "Skipping parameter initialization";
} else {
this->blobs_.resize(4);
// fill scale with scale_filler
this->blobs_[0].reset(new Blob<Dtype>(1, channels_, 1, 1));
shared_ptr<Filler<Dtype> > scale_filler(GetFiller<Dtype>(
this->layer_param_.bn_param().scale_filler()));
scale_filler->Fill(this->blobs_[0].get());
// fill shift with shift_filler
this->blobs_[1].reset(new Blob<Dtype>(1, channels_, 1, 1));
shared_ptr<Filler<Dtype> > shift_filler(GetFiller<Dtype>(
this->layer_param_.bn_param().shift_filler()));
shift_filler->Fill(this->blobs_[1].get());
// history mean
this->blobs_[2].reset(new Blob<Dtype>(1, channels_, 1, 1));
caffe_set(channels_, Dtype(0), this->blobs_[2]->mutable_cpu_data());
// history variance
this->blobs_[3].reset(new Blob<Dtype>(1, channels_, 1, 1));
caffe_set(channels_, Dtype(0), this->blobs_[3]->mutable_cpu_data());
} // parameter initialization
this->param_propagate_down_.resize(this->blobs_.size(), true);
this->param_propagate_down_[2] = false;
this->param_propagate_down_[3] = false;
}
