void LogLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const int count = bottom[0]->count();
const Dtype* bottom_data = bottom[0]->cpu_data();
Dtype* top_data = top[0]->mutable_cpu_data();
if (input_scale_ == Dtype(1) && input_shift_ == Dtype(0)) {
caffe_log(count, bottom_data, top_data);
} else {
caffe_cpu_copy(count, bottom_data, top_data);
if (input_scale_ != Dtype(1)) {
caffe_scal(count, input_scale_, top_data);
}
if (input_shift_ != Dtype(0)) {
caffe_add_scalar(count, input_shift_, top_data);
}
caffe_log(count, top_data, top_data);
}
if (base_scale_ != Dtype(1)) {
caffe_scal(count, base_scale_, top_data);
}
}
void CRFWithLossLayer<Dtype>::Reshape(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
CHECK_EQ(bottom.size(), 2)
<< "CRFWithLossLayer need both feature table and labels as inputs";
CHECK_EQ(bottom[0]->count(1, 2), feature_num_)
<< "The state num of input feature table doesn't match the state-feature"
<< "number of the CRF";
CHECK_EQ(bottom[0]->count(2), max_seq_length_)
<< "The input feature table must have a length of max_seq_length_ as defined in crf layer"
<< "for a requirement of alignment in blobs";
CHECK_EQ(bottom[1]->count(2), max_seq_length_)
<< "The input label sequence should be of the same length as the input feature table";
CHECK_EQ(bottom[1]->count(1, 2), 1)
<< "The label sequence is a one dimensional vector with each element of it referring to a class number";
// Training top: (nbest label, nbest prob); Test top: (-logProb)
if (!for_training_) {
CHECK_EQ(top.size(), 2)
<< "The Top should have two blobs: pred labels, predictive probability";
vector<int> output_seq_shape = bottom[0]->shape();
output_seq_shape[1] = nbest_;
top[0]->Reshape(output_seq_shape);
vector<int> output_prob_shape = bottom[0]->shape();
output_prob_shape[1] = nbest_;
output_prob_shape[2] = 1;
output_prob_shape[3] = 1;
top[1]->Reshape(output_prob_shape);
} else {
vector<int> loss_shape(1);
loss_shape[0] = bottom[0]->count(0,1);
top[0]->Reshape(loss_shape);
}
// Do The Transpose operation input bottom , which is important for the RowMajor routine
Dtype* ptr_buf = buf_bottom_transposed_.mutable_cpu_data();
// Table size ( each matrix size)
int ts = bottom[0]->count(1);
// Transpose to a buffer and make a log energy transition of each element
// for following matrix multiplication of LogSumExp version
for (int i = 0; i < num_; ++i) {
caffe_cpu_transpose(feature_num_, max_seq_length_, bottom[0]->cpu_data() + i * ts, ptr_buf + i * ts);
caffe_log(ts, buf_bottom_transposed_.cpu_data() + i * ts, ptr_buf);
}
}
