fully_connected_op_prune.h

#ifndef CAFFE2_OPERATORS_FULLY_CONNECTED_OP_PRUNE_H_
#define CAFFE2_OPERATORS_FULLY_CONNECTED_OP_PRUNE_H_

#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"

namespace caffe2 {

  namespace {

    template<int N>
      using Shape = std::array<int, N>;

    template<int N>
      const std::vector<TIndex>& shape(Shape<N> vs) {
        static thread_local std::vector<TIndex> cache;
        cache.resize(vs.size());
        for (auto i = 0; i < vs.size(); ++i) {
          cache[i] = vs[i];
        }
        return cache;
      }

    inline const std::vector<TIndex>& shape(int i) {
      return shape<1>(Shape<1>({i}));
    }

    inline const std::vector<TIndex>& shape(int i, int j) {
      return shape<2>(Shape<2>({i, j}));
    }

    template <typename T, class Context>
      void MaskMatrix(const T* mask, T* mat,
          int M, int N);

    template <typename T, class Context>
      void MaskMatrix_Inc(T* mask_seq, T* mat,
          int M, int N, int seq_len, T target);

    template <typename T, class Context>
      void AggrDW(T* ag_dw, const T* dw, int N, int K, Context* context);

    template <typename T>
      int MatrixCompare_LT(const T* mat, float thres,
                           T* mask_seq, int M, int N);

    // TODO(wyiming): write an incremental Mask
    // Incremental Mask: only give the new mask positions;
    // Assuming that weights masked will not be mask again;
    // The incremental mask can also be used to update mask matrix;
    // But this will include template for bool and float;
    template <>
      void MaskMatrix<float, CPUContext>(
          const float* mask, float* mat, int M, int N) {
        int offset = 0;
        for (int i = 0; i < M; ++i) {
          for (int j = 0; j < N; ++j) {
            mat[offset] = mask[offset]? mat[offset] : 0;
            offset++;
          }
        }
      }

      template <>
      void MaskMatrix_Inc<float, CPUContext>(
          float* mask_seq,
          float* mat,
          int /*M*/,
          int /*N*/,
          int seq_len,
          float target) {
        for (int i = 0; i < seq_len; ++i) {
          // assume that the mask_seq is smaller than size
          // Although it seems that random access gets bad performance,
          // we make sure that seq is in order;
          mat[static_cast<int>(mask_seq[i])] = target;
        }
      }

    template <>
      void AggrDW<float, CPUContext>(
          float* ag_dw, const float* dw,
          int N, int K, CPUContext* context) {
        math::Add<float, CPUContext>(N*K, dw, ag_dw, ag_dw, context);
      }

    template <>
      int MatrixCompare_LT<float>(
          const float* mat, float thres,
          float* mask_seq, int M, int N) {
        int seq_len = 0;
        int offset = 0;
        for (int i = 0 ; i < M; ++i) {
          for (int j = 0; j < N; ++j) {
            if (mat[offset] != 0 &&
                (mat[offset] < thres && mat[offset] > -thres)) {
              mask_seq[seq_len++] = static_cast<float>(offset);
            }
            offset++;
          }
        }
        return seq_len;
      }

  }

  // This is Caffe's InnerProductOp, with a name that fits its purpose better.
  template <typename T, class Context, class Engine=DefaultEngine>
    class FullyConnectedOpPrune final : public Operator<Context> {
      public:
        USE_OPERATOR_CONTEXT_FUNCTIONS;
        FullyConnectedOpPrune(const OperatorDef& operator_def, Workspace* ws)
          : Operator<Context>(operator_def, ws) {}
        ~FullyConnectedOpPrune() {}

        bool RunOnDevice() override {
          const auto& X = Input(0);
          const auto& W = Input(1);
          const auto& Mask = Input(2);
          const auto& b = Input(3);
          auto* Y = Output(0);
          CAFFE_ENFORCE_GE(X.ndim(), 1);
          CAFFE_ENFORCE_GE(W.ndim(), 2);
          if (X.ndim() > 2 || W.ndim() > 2) {
            VLOG(1) << "Using legacy support for arbitrary input and weight "
              "dimensions.";
          }
          CAFFE_ENFORCE_EQ(b.ndim(), 1);
          // batch size
          int M = X.ndim() > 1 ? X.dim32(0) : 1;
          // Feature dimension
          int K = X.size() / M;
          // number of outputs.
          int N = W.dim32(0);
          CAFFE_ENFORCE_EQ(K, W.size() / W.dim32(0));
          CAFFE_ENFORCE_EQ(N, b.dim32(0));
          if (X.ndim() > 1) {
            Y->Resize(M, N);
          } else {
            Y->Resize(N);
          }
          // W * x
          math::Gemm<T, Context, Engine>(
              CblasNoTrans, CblasTrans, M, N, K, 1, X.template data<T>(),
              W.template data<T>(), 0, Y->template mutable_data<T>(),
              &context_);
          // Add bias term
          if (bias_multiplier_.size() != M) {
            // If the helper bias multiplier is not M,
            // reshape and fill it with one.
            bias_multiplier_.Resize(M);
            math::Set<T, Context>(
                M, static_cast<T>(1),
                bias_multiplier_.template mutable_data<T>(),
                &context_);
          }
          math::Gemm<T, Context, Engine>(
              CblasNoTrans, CblasNoTrans, M, N, 1, 1,
              bias_multiplier_.template data<T>(), b.template data<T>(), 1,
              Y->template mutable_data<T>(), &context_);
          if (OutputSize() == 2){
            auto* Comp_rate = Output(1);
            Comp_rate->Resize(vector<TIndex>());
            T* comp_data = Comp_rate->template mutable_data<T>();
            math::Sum<T, Context>(
                Mask.size(), Mask.template data<T>(), comp_data, &context_);
            math::Scale<T, Context>(
                1, static_cast<T>(1.) / Mask.size(), comp_data, comp_data,
                &context_);
          }
          return true;
        }

      protected:
        Tensor<Context> bias_multiplier_;
    };

  template <typename T, class Context, class Engine=DefaultEngine>
    class FullyConnectedPruneGradientOp : public Operator<Context> {
      public:
        int iter_offset;
      public:
        USE_OPERATOR_CONTEXT_FUNCTIONS;
        FullyConnectedPruneGradientOp
          (const OperatorDef& operator_def, Workspace* ws)
          : Operator<Context>(operator_def, ws) { iter_offset = 0; }
        ~FullyConnectedPruneGradientOp() {}

        bool RunOnDevice() override {
          const auto& X = Input(0);
          //const auto& W = Input(1);
          auto* W_ptr = Output(2);
          auto& W = *W_ptr;
          //const auto& Mask = Input(2);
          auto* Mask_ptr = Output(3);
          auto& Mask = *Mask_ptr;
          const auto& dY = Input(3);
          //const auto& Ag_dW = Input(4);
          auto* Ag_dW_ptr = Output(4);
          auto& Ag_dW = *Ag_dW_ptr;
          // it is also the Input(5)
          auto* mask_seq_auto = Output(5);
          // how about get threshold
          auto& thres = Input(6);
          //TODO(wyiming): check comp_lb is a float
          auto& comp_lb = Input(7);
          DCHECK_GE(X.ndim(), 1);
          DCHECK_GE(W.ndim(), 2);
          DCHECK_LE(dY.ndim(), 2);
          // batch size
          int M = X.ndim() > 1 ? X.dim32(0) : 1;
          // Feature dimension
          int K = X.size() / M;
          // number of outputs.
          int N = W.dim32(0);
          // TODO(wyiming): add this window_size to workspace?
          int window_size = 100;
          // TODO(wyiming): this threshold should be
          // based on distribution of the layer weight
          float thr = 0.01;
          DCHECK_EQ(Mask.dim32(0), W.dim32(0));
          DCHECK_EQ(Mask.dim32(1), W.dim32(1));
          DCHECK_EQ(Ag_dW.dim32(0), W.dim32(0));
          DCHECK_EQ(Ag_dW.dim32(1), W.dim32(1));
          DCHECK_EQ(K, W.size() / W.dim32(0));
          if (dY.ndim() > 1) {
            DCHECK_EQ(M, dY.dim32(0));
            DCHECK_EQ(N, dY.dim32(1));
          } else {
            DCHECK_EQ(X.ndim(), 1);
            DCHECK_EQ(N, dY.size());
          }
          auto* dW = Output(0);
          auto* db = Output(1);
          dW->ResizeLike(W);
          db->Resize(N);

          // Compute dW
          math::Gemm<T, Context, Engine>(
              CblasTrans, CblasNoTrans, N, K, M, 1,
              dY.template data<T>(), X.template data<T>(),
              0, dW->template mutable_data<T>(),
              &context_);

          comp_r_buf_.Resize(vector<TIndex>());
          T* comp_data = comp_r_buf_.template mutable_data<T>();
          math::Sum<T, Context>(
              Mask.size(), Mask.template data<T>(), comp_data, &context_);
          math::Scale<T, Context>(
              1, static_cast<T>(1.) / Mask.size(), comp_data, comp_data,
              &context_);
          // update W size window
          // Notice here we need to maintain state in OP.
          // This is new in Caffe2.
          // And this is something we might need to discuss in the future.
          // at most mask half of the matrix at time
          // 1. mask dw with previous mask
          MaskMatrix<T, Context>(Mask.template mutable_data<T>(),
              dW->template mutable_data<T>(), N, K);
          if(*comp_data > *(comp_lb.template data<T>())){
            iter_offset++;
            if (iter_offset % window_size == 0) {
              // TODO(wyiming):do the prune here;
              sum_buffer_.ResizeLike(W);
              math::Add<T, Context>(W.size(),
                  W.template mutable_data<T>(),
                  Ag_dW.template mutable_data<T>(),
                  sum_buffer_.template mutable_data<T>(),
                  &context_);
              mask_seq_auto->ResizeLike(W);
              T* mask_seq = mask_seq_auto->template mutable_data<T>();
              math::Set<T, Context>(N*K, static_cast<T>(0),
                  mask_seq_auto->template mutable_data<T>(), &context_);
              // 2. find dw below thres but not eq 0
              int seq_len = MatrixCompare_LT<T>(
                  Ag_dW_ptr->template mutable_data<T>(),
                  *thres.template data<T>(), mask_seq, N, K);
              // 3. use the mask_seq to update W and dw
              MaskMatrix_Inc<T, Context>(mask_seq,
                                         dW->template mutable_data<T>(),
                                         N, K, seq_len, 0);
              MaskMatrix_Inc<T, Context>(mask_seq,
                                         W.template mutable_data<T>(),
                                         N, K, seq_len, 0);
              MaskMatrix_Inc<T, Context>(mask_seq,
                                         Mask.template mutable_data<T>(),
                                         N, K, seq_len, 0);
              math::Set<T, Context>(N*K, static_cast<T>(0),
                  Ag_dW.template mutable_data<T>(),
                  &context_);
            } else {
              // add dW to Aggregate dW.
              AggrDW<T, Context>(
                  Ag_dW.template mutable_data<T>(),
                  dW->template mutable_data<T>(),
                  N, K, &context_);
            }
          }
          if (bias_multiplier_.size() != M) {
            // If the helper bias multiplier is not M,
            // reshape and fill it with one.
            bias_multiplier_.Resize(M);
            math::Set<T, Context>(
                M, static_cast<T>(1),
                bias_multiplier_.template mutable_data<T>(),
                &context_);
          }
          // Compute dB
          math::Gemv<T, Context>(
              CblasTrans, M, N, 1, dY.template data<T>(),
              bias_multiplier_.template data<T>(), 0,
              db->template mutable_data<T>(),
              &context_);
          // Compute dX if necessary.
          if (OutputSize() == 7) {
            auto* dX = Output(6);
            dX->ResizeLike(X);
            math::Gemm<T, Context, Engine>(
                CblasNoTrans, CblasNoTrans, M, K, N, 1,
                dY.template data<T>(), W.template data<T>(),
                0, dX->template mutable_data<T>(),
                &context_);
          }

          return true;
        }

      protected:
        Tensor<Context> bias_multiplier_;
        Tensor<Context> sum_buffer_;
        Tensor<Context> comp_r_buf_;
    };

}  // namespace caffe2

#endif  // CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_