caffe 源码学习笔记(9) reduce layer

背景

reduce layer其实就是做reduce操作,把一个任意shape的blob通过某种运算变成一个scalar.

caffe目前支持求和(SUM),绝对值的和(ASUM),平方和(SUMSQ),以及对得到的scalar的总数求平均的求和(MEAN).

proto

`````` 1
2message ReductionParameter {
3  enum ReductionOp {
4    SUM = 1;
5    ASUM = 2;
6    SUMSQ = 3;
7    MEAN = 4;
8  }
9
10  optional ReductionOp operation = 1 [default = SUM]; // reduction operation
11
12  // The first axis to reduce to a scalar -- may be negative to index from the
13  // end (e.g., -1 for the last axis).
14  // (Currently, only reduction along ALL "tail" axes is supported; reduction
15  // of axis M through N, where N < num_axes - 1, is unsupported.)
16  // Suppose we have an n-axis bottom Blob with shape:
17  //     (d0, d1, d2, ..., d(m-1), dm, d(m+1), ..., d(n-1)).
18  // If axis == m, the output Blob will have shape
19  //     (d0, d1, d2, ..., d(m-1)),
20  // and the ReductionOp operation is performed (d0 * d1 * d2 * ... * d(m-1))
21  // times, each including (dm * d(m+1) * ... * d(n-1)) individual data.
22  // If axis == 0 (the default), the output Blob always has the empty shape
23  // (count 1), performing reduction across the entire input --
24  // often useful for creating new loss functions.
25  optional int32 axis = 2 [default = 0];
26
27  optional float coeff = 3 [default = 1.0]; // coefficient for output
28}
29
``````

operation不用说,表示要进行哪种reduce操作. coeff也比较简单,就是最后结果中每个元素都可以乘一个额外的系数,默认是1.

axis的含义是,从axis这个维度开始做reduce.

c++ 代码

`````` 1
2template <typename Dtype>
3class ReductionLayer : public Layer<Dtype> {
4 public:
5//    ...省略无关部分
6
7  /// @brief the reduction operation performed by the layer
8  ReductionParameter_ReductionOp op_;
9  /// @brief a scalar coefficient applied to all outputs
10  Dtype coeff_;
11  /// @brief the index of the first input axis to reduce
12  int axis_;
13  /// @brief the number of reductions performed
14  int num_;
15  /// @brief the input size of each reduction
16  int dim_;
17  /// @brief a helper Blob used for summation (op_ == SUM)
18  Blob<Dtype> sum_multiplier_;
19};
20
21
22num_表示的是要做reduce操作的次数
23dim_表示的是每次reduce的维度部分的count
24
25Reshape部分,主要就是计算了几个成员变量 num_,dim_,coeff_
26```c++
27
28template <typename Dtype>
29void ReductionLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
30      const vector<Blob<Dtype>*>& top) {
31  axis_ = bottom[0]->CanonicalAxisIndex(
32      this->layer_param_.reduction_param().axis());
33  // In the output, we'll keep all axes up to the reduction axis, but
34  // throw away any after that.
35  // Note: currently reducing along non-tail axes is not supported; otherwise,
36  // we'd need to also copy any axes following an "end_axis".
37  vector<int> top_shape(bottom[0]->shape().begin(),
38                        bottom[0]->shape().begin() + axis_);
39  top[0]->Reshape(top_shape);
40  //  count都是左闭又开,一个参数就是到末尾
41  num_ = bottom[0]->count(0, axis_);
42  //  num_表示的是要做reduce操作的次数
43  //  range [0,axis_)
44  dim_ = bottom[0]->count(axis_);
45  //  dim_表示的是reduce操作部分的count
46  //  range [axis_,num_axes())
47  CHECK_EQ(num_, top[0]->count());
48  if (op_ == ReductionParameter_ReductionOp_SUM ||
49      op_ == ReductionParameter_ReductionOp_MEAN) {
50    vector<int> sum_mult_shape(1, dim_);
51    // sum_mult_shape[0] = dim_
52    sum_multiplier_.Reshape(sum_mult_shape);
53
54    caffe_set(dim_, Dtype(1), sum_multiplier_.mutable_cpu_data());
55    // sum_multiplier_是dim_长度的vector,值都为1
56  }
57  coeff_ = this->layer_param().reduction_param().coeff();
58  if (op_ == ReductionParameter_ReductionOp_MEAN) {
59    coeff_ /= dim_;
60  }
61}
62
``````

`````` 1
2template <typename Dtype>
3void ReductionLayer<Dtype>::Forward_cpu(
4    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
5  const Dtype* bottom_data = bottom[0]->cpu_data();
6  const Dtype* mult_data = NULL;
7  if (sum_multiplier_.count() > 0) {
8    mult_data = sum_multiplier_.cpu_data();
9  }
10  Dtype* top_data = top[0]->mutable_cpu_data();
11  // 得到num_个结果
12  for (int i = 0; i < num_; ++i) {
13    switch (op_) {
14    case ReductionParameter_ReductionOp_SUM:
15    case ReductionParameter_ReductionOp_MEAN:
16      *top_data = caffe_cpu_dot(dim_, mult_data, bottom_data);
17      break;
18    case ReductionParameter_ReductionOp_ASUM:
19      *top_data = caffe_cpu_asum(dim_, bottom_data);
20      break;
21    case ReductionParameter_ReductionOp_SUMSQ:
22      *top_data = caffe_cpu_dot(dim_, bottom_data, bottom_data);
23      break;
24    default:
25      LOG(FATAL) << "Unknown reduction op: "
26          << ReductionParameter_ReductionOp_Name(op_);
27    }
28    bottom_data += dim_;
29    ++top_data;
30  }
31  //  coeff_是每一个位置的系数,默认为1
32  //  此时才真的把系数放在每一个元素上
33  if (coeff_ != Dtype(1)) {
34    // Reset the top_data pointer.
35    top_data = top[0]->mutable_cpu_data();
36    caffe_scal(num_, coeff_, top_data);
37  }
38}
``````