| 143 | |
| 144 | template <typename Dtype> |
| 145 | void ScaleLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top, |
| 146 | const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) { |
| 147 | if (bias_layer_ && |
| 148 | this->param_propagate_down_[this->param_propagate_down_.size() - 1]) { |
| 149 | bias_layer_->Backward(top, bias_propagate_down_, bias_bottom_vec_); |
| 150 | } |
| 151 | const bool scale_param = (bottom.size() == 1); |
| 152 | Blob<Dtype>* scale = scale_param ? this->blobs_[0].get() : bottom[1]; |
| 153 | if ((!scale_param && propagate_down[1]) || |
| 154 | (scale_param && this->param_propagate_down_[0])) { |
| 155 | const Dtype* top_diff = top[0]->cpu_diff(); |
| 156 | const bool in_place = (bottom[0] == top[0]); |
| 157 | const Dtype* bottom_data = (in_place ? &temp_ : bottom[0])->cpu_data(); |
| 158 | // Hack: store big eltwise product in bottom[0] diff, except in the special |
| 159 | // case where this layer itself does the eltwise product, in which case we |
| 160 | // can store it directly in the scale diff, and we're done. |
| 161 | // If we're computing in-place (and not doing eltwise computation), this |
| 162 | // hack doesn't work and we store the product in temp_. |
| 163 | const bool is_eltwise = (bottom[0]->count() == scale->count()); |
| 164 | Dtype* product = (is_eltwise ? scale->mutable_cpu_diff() : |
| 165 | (in_place ? temp_.mutable_cpu_data() : bottom[0]->mutable_cpu_diff())); |
| 166 | caffe_mul(top[0]->count(), top_diff, bottom_data, product); |
| 167 | if (!is_eltwise) { |
| 168 | Dtype* sum_result = NULL; |
| 169 | if (inner_dim_ == 1) { |
| 170 | sum_result = product; |
| 171 | } else if (sum_result_.count() == 1) { |
| 172 | const Dtype* sum_mult = sum_multiplier_.cpu_data(); |
| 173 | Dtype* scale_diff = scale->mutable_cpu_diff(); |
| 174 | if (scale_param) { |
| 175 | Dtype result = caffe_cpu_dot(inner_dim_, product, sum_mult); |
| 176 | *scale_diff += result; |
| 177 | } else { |
| 178 | *scale_diff = caffe_cpu_dot(inner_dim_, product, sum_mult); |
| 179 | } |
| 180 | } else { |
| 181 | const Dtype* sum_mult = sum_multiplier_.cpu_data(); |
| 182 | sum_result = (outer_dim_ == 1) ? |
| 183 | scale->mutable_cpu_diff() : sum_result_.mutable_cpu_data(); |
| 184 | caffe_cpu_gemv(CblasNoTrans, sum_result_.count(), inner_dim_, |
| 185 | Dtype(1), product, sum_mult, Dtype(0), sum_result); |
| 186 | } |
| 187 | if (outer_dim_ != 1) { |
| 188 | const Dtype* sum_mult = sum_multiplier_.cpu_data(); |
| 189 | Dtype* scale_diff = scale->mutable_cpu_diff(); |
| 190 | if (scale_dim_ == 1) { |
| 191 | if (scale_param) { |
| 192 | Dtype result = caffe_cpu_dot(outer_dim_, sum_mult, sum_result); |
| 193 | *scale_diff += result; |
| 194 | } else { |
| 195 | *scale_diff = caffe_cpu_dot(outer_dim_, sum_mult, sum_result); |
| 196 | } |
| 197 | } else { |
| 198 | caffe_cpu_gemv(CblasTrans, outer_dim_, scale_dim_, |
| 199 | Dtype(1), sum_result, sum_mult, Dtype(scale_param), |
| 200 | scale_diff); |
| 201 | } |
| 202 | } |
nothing calls this directly
no test coverage detected