Forward propagation

template <typename Dtype>
void SoftmaxLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->cpu_data();
  Dtype* top_data = top[0]->mutable_cpu_data();
  Dtype* scale_data = scale_.mutable_cpu_data();
  int channels = bottom[0]->shape(softmax_axis_);
  int dim = bottom[0]->count() / outer_num_; //dim represents the number of categories to be classified, count() gets the total number of input blobs, and outer_num_ gets the number of blobs for each category
  caffe_copy(bottom[0]->count(), bottom_data, top_data); //First copy the input to the output buffer
  // We need to subtract the max to avoid numerical issues, compute the exp,
  // and then normalize, subtract the maximum value, avoid numerical problems, calculate the exponent, and normalize
  for (int i = 0; i < outer_num_; ++i) {
    // Initialize the data field of scale_ as the first plane, where scale is used to store temporary calculation results
    caffe_copy(inner_num_, bottom_data + i * dim, scale_data);
    for (int j = 0; j < channels; j++) {
      for (int k = 0; k < inner_num_; k++) {
        scale_data[k] = std::max(scale_data[k],
            bottom_data[i * dim + j * inner_num_ + k]);
      }
    }
    // output buffer minus maximum value
    //a_k = a_k - max(a_i)
    // C = alpha . A . B + beta * C
    // A is the unit 1 matrix
    caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels, inner_num_,
        1, -1., sum_multiplier_.cpu_data(), scale_data, 1., top_data);
    // exponentiation
    caffe_exp<Dtype>(dim, top_data, top_data);
    // sum after exp
    caffe_cpu_gemv<Dtype>(CblasTrans, channels, inner_num_, 1.,
        top_data, sum_multiplier_.cpu_data(), 0., scale_data);
    // division
    for (int j = 0; j < channels; j++) {
      caffe_div(inner_num_, top_data, scale_data, top_data);
      top_data += inner_num_;
    }
  }
}

Backward propagation

template <typename Dtype>
void SoftmaxLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down,
    const vector<Blob<Dtype>*>& bottom) {
    //Get data, diff pointer
  const Dtype* top_diff = top[0]->cpu_diff();
  const Dtype* top_data = top[0]->cpu_data();
  Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
  Dtype* scale_data = scale_.mutable_cpu_data();
  int channels = top[0]->shape(softmax_axis_);
  int dim = top[0]->count() / outer_num_;
  caffe_copy(top[0]->count(), top_diff, bottom_diff); //First initialize bottom_diff with top_diff
  for (int i = 0; i < outer_num_; ++i) {
    // Calculate the dot product of top_diff and top_data, and then subtract the value from bottom_diff
    for (int k = 0; k < inner_num_; ++k) {
      scale_data[k] = caffe_cpu_strided_dot<Dtype>(channels,
          bottom_diff + i * dim + k, inner_num_,
          top_data + i * dim + k, inner_num_);
    }
    // Impairment
    caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels, inner_num_, 1,
        -1., sum_multiplier_.cpu_data(), scale_data, 1., bottom_diff + i * dim);
  }
  // multiply point by point
  caffe_mul(top[0]->count(), bottom_diff, top_data, bottom_diff);
}