#ifndef CAFFE_NEURON_LAYERS_HPP_
#define CAFFE_NEURON_LAYERS_HPP_

#include <string>
#include <utility>
#include <vector>

#include "caffe/blob.hpp"
#include "caffe/common.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"

#define HDF5_DATA_DATASET_NAME "data"
#define HDF5_DATA_LABEL_NAME "label"

namespace caffe {

/**
 * @brief An interface for layers that take one blob as input (@f$ x @f$)
 *        and produce one equally-sized blob as output (@f$ y @f$), where
 *        each element of the output depends only on the corresponding input
 *        element.
 */
template <typename Dtype>
class NeuronLayer : public Layer<Dtype> {
 public:
  explicit NeuronLayer(const LayerParameter& param)
     : Layer<Dtype>(param) {}
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline int ExactNumBottomBlobs() const { return 1; }
  virtual inline int ExactNumTopBlobs() const { return 1; }
};

/**
 * @brief Computes @f$ y = |x| @f$
 *
 * @param bottom input Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the inputs @f$ x @f$
 * @param top output Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the computed outputs @f$ y = |x| @f$
 */
template <typename Dtype>
class AbsValLayer : public NeuronLayer<Dtype> {
 public:
  explicit AbsValLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "AbsVal"; }
  virtual inline int ExactNumBottomBlobs() const { return 1; }
  virtual inline int ExactNumTopBlobs() const { return 1; }

 protected:
  /// @copydoc AbsValLayer
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the absolute value inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 2)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x} =
   *            \mathrm{sign}(x) \frac{\partial E}{\partial y}
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

/**
 * @brief Computes @f$ y = x + \log(1 + \exp(-x)) @f$ if @f$ x > 0 @f$;
 *        @f$ y = \log(1 + \exp(x)) @f$ otherwise.
 *
 * @param bottom input Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the inputs @f$ x @f$
 * @param top output Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the computed outputs @f$
 *      y = \left\{
 *         \begin{array}{ll}
 *            x + \log(1 + \exp(-x)) & \mbox{if } x > 0 \\
 *            \log(1 + \exp(x)) & \mbox{otherwise}
 *         \end{array} \right.
 *      @f$
 */
template <typename Dtype>
class BNLLLayer : public NeuronLayer<Dtype> {
 public:
  explicit BNLLLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual inline const char* type() const { return "BNLL"; }

 protected:
  /// @copydoc BNLLLayer
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the BNLL inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 2)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x}
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

/**
 * @brief During training only, sets a random portion of @f$x@f$ to 0, adjusting
 *        the rest of the vector magnitude accordingly.
 *
 * @param bottom input Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the inputs @f$ x @f$
 * @param top output Blob vector (length 1)
 *   -# @f$ (N \times C \times H \times W) @f$
 *      the computed outputs @f$ y = |x| @f$
 */
template <typename Dtype>
class DropoutLayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides DropoutParameter dropout_param,
   *     with DropoutLayer options:
   *   - dropout_ratio (\b optional, default 0.5).
   *     Sets the probability @f$ p @f$ that any given unit is dropped.
   */
  explicit DropoutLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "Dropout"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs. At training time, we have @f$
   *      y_{\mbox{train}} = \left\{
   *         \begin{array}{ll}
   *            \frac{x}{1 - p} & \mbox{if } u > p \\
   *            0 & \mbox{otherwise}
   *         \end{array} \right.
   *      @f$, where @f$ u \sim U(0, 1)@f$ is generated independently for each
   *      input at each iteration. At test time, we simply have
   *      @f$ y_{\mbox{test}} = \mathbb{E}[y_{\mbox{train}}] = x @f$.
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  /// when divided by UINT_MAX, the randomly generated values @f$u\sim U(0,1)@f$
  Blob<unsigned int> rand_vec_;
  /// the probability @f$ p @f$ of dropping any input
  Dtype threshold_;
  /// the scale for undropped inputs at train time @f$ 1 / (1 - p) @f$
  Dtype scale_;
  unsigned int uint_thres_;
};

/**
 * @brief Computes @f$ y = \gamma ^ {\alpha x + \beta} @f$,
 *        as specified by the scale @f$ \alpha @f$, shift @f$ \beta @f$,
 *        and base @f$ \gamma @f$.
 */
template <typename Dtype>
class ExpLayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides ExpParameter exp_param,
   *     with ExpLayer options:
   *   - scale (\b optional, default 1) the scale @f$ \alpha @f$
   *   - shift (\b optional, default 0) the shift @f$ \beta @f$
   *   - base (\b optional, default -1 for a value of @f$ e \approx 2.718 @f$)
   *         the base @f$ \gamma @f$
   */
  explicit ExpLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "Exp"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = \gamma ^ {\alpha x + \beta}
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the exp inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x} =
   *            \frac{\partial E}{\partial y} y \alpha \log_e(gamma)
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  Dtype inner_scale_, outer_scale_;
};

/**
 * @brief Computes @f$ y = log_{\gamma}(\alpha x + \beta) @f$,
 *        as specified by the scale @f$ \alpha @f$, shift @f$ \beta @f$,
 *        and base @f$ \gamma @f$.
 */
template <typename Dtype>
class LogLayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides LogParameter log_param,
   *     with LogLayer options:
   *   - scale (\b optional, default 1) the scale @f$ \alpha @f$
   *   - shift (\b optional, default 0) the shift @f$ \beta @f$
   *   - base (\b optional, default -1 for a value of @f$ e \approx 2.718 @f$)
   *         the base @f$ \gamma @f$
   */
  explicit LogLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "Log"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = log_{\gamma}(\alpha x + \beta)
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the exp inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x} =
   *            \frac{\partial E}{\partial y} y \alpha \log_e(gamma)
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  Dtype base_scale_;
  Dtype input_scale_, input_shift_;
  Dtype backward_num_scale_;
};

/**
 * @brief Computes @f$ y = (\alpha x + \beta) ^ \gamma @f$,
 *        as specified by the scale @f$ \alpha @f$, shift @f$ \beta @f$,
 *        and power @f$ \gamma @f$.
 */
template <typename Dtype>
class PowerLayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides PowerParameter power_param,
   *     with PowerLayer options:
   *   - scale (\b optional, default 1) the scale @f$ \alpha @f$
   *   - shift (\b optional, default 0) the shift @f$ \beta @f$
   *   - power (\b optional, default 1) the power @f$ \gamma @f$
   */
  explicit PowerLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "Power"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = (\alpha x + \beta) ^ \gamma
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the power inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x} =
   *            \frac{\partial E}{\partial y}
   *            \alpha \gamma (\alpha x + \beta) ^ {\gamma - 1} =
   *            \frac{\partial E}{\partial y}
   *            \frac{\alpha \gamma y}{\alpha x + \beta}
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  /// @brief @f$ \gamma @f$ from layer_param_.power_param()
  Dtype power_;
  /// @brief @f$ \alpha @f$ from layer_param_.power_param()
  Dtype scale_;
  /// @brief @f$ \beta @f$ from layer_param_.power_param()
  Dtype shift_;
  /// @brief Result of @f$ \alpha \gamma @f$
  Dtype diff_scale_;
};

/**
 * @brief Rectified Linear Unit non-linearity @f$ y = \max(0, x) @f$.
 *        The simple max is fast to compute, and the function does not saturate.
 */
template <typename Dtype>
class ReLULayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides ReLUParameter relu_param,
   *     with ReLULayer options:
   *   - negative_slope (\b optional, default 0).
   *     the value @f$ \nu @f$ by which negative values are multiplied.
   */
  explicit ReLULayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual inline const char* type() const { return "ReLU"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = \max(0, x)
   *      @f$ by default.  If a non-zero negative_slope @f$ \nu @f$ is provided,
   *      the computed outputs are @f$ y = \max(0, x) + \nu \min(0, x) @f$.
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the ReLU inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x} = \left\{
   *        \begin{array}{lr}
   *            0 & \mathrm{if} \; x \le 0 \\
   *            \frac{\partial E}{\partial y} & \mathrm{if} \; x > 0
   *        \end{array} \right.
   *      @f$ if propagate_down[0], by default.
   *      If a non-zero negative_slope @f$ \nu @f$ is provided,
   *      the computed gradients are @f$
   *        \frac{\partial E}{\partial x} = \left\{
   *        \begin{array}{lr}
   *            \nu \frac{\partial E}{\partial y} & \mathrm{if} \; x \le 0 \\
   *            \frac{\partial E}{\partial y} & \mathrm{if} \; x > 0
   *        \end{array} \right.
   *      @f$.
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

#ifdef USE_CUDNN
/**
 * @brief CuDNN acceleration of ReLULayer.
 */
template <typename Dtype>
class CuDNNReLULayer : public ReLULayer<Dtype> {
 public:
  explicit CuDNNReLULayer(const LayerParameter& param)
      : ReLULayer<Dtype>(param), handles_setup_(false) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual ~CuDNNReLULayer();

 protected:
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  bool handles_setup_;
  cudnnHandle_t             handle_;
  cudnnTensorDescriptor_t bottom_desc_;
  cudnnTensorDescriptor_t top_desc_;
};
#endif

/**
 * @brief Sigmoid function non-linearity @f$
 *         y = (1 + \exp(-x))^{-1}
 *     @f$, a classic choice in neural networks.
 *
 * Note that the gradient vanishes as the values move away from 0.
 * The ReLULayer is often a better choice for this reason.
 */
template <typename Dtype>
class SigmoidLayer : public NeuronLayer<Dtype> {
 public:
  explicit SigmoidLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual inline const char* type() const { return "Sigmoid"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = (1 + \exp(-x))^{-1}
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the sigmoid inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x}
   *            = \frac{\partial E}{\partial y} y (1 - y)
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

#ifdef USE_CUDNN
/**
 * @brief CuDNN acceleration of SigmoidLayer.
 */
template <typename Dtype>
class CuDNNSigmoidLayer : public SigmoidLayer<Dtype> {
 public:
  explicit CuDNNSigmoidLayer(const LayerParameter& param)
      : SigmoidLayer<Dtype>(param), handles_setup_(false) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual ~CuDNNSigmoidLayer();

 protected:
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  bool handles_setup_;
  cudnnHandle_t             handle_;
  cudnnTensorDescriptor_t bottom_desc_;
  cudnnTensorDescriptor_t top_desc_;
};
#endif

/**
 * @brief TanH hyperbolic tangent non-linearity @f$
 *         y = \frac{\exp(2x) - 1}{\exp(2x) + 1}
 *     @f$, popular in auto-encoders.
 *
 * Note that the gradient vanishes as the values move away from 0.
 * The ReLULayer is often a better choice for this reason.
 */
template <typename Dtype>
class TanHLayer : public NeuronLayer<Dtype> {
 public:
  explicit TanHLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual inline const char* type() const { return "TanH"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *        y = \frac{\exp(2x) - 1}{\exp(2x) + 1}
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the sigmoid inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times H \times W) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$; Backward fills their diff with
   *      gradients @f$
   *        \frac{\partial E}{\partial x}
   *            = \frac{\partial E}{\partial y}
   *              \left(1 - \left[\frac{\exp(2x) - 1}{exp(2x) + 1} \right]^2 \right)
   *            = \frac{\partial E}{\partial y} (1 - y^2)
   *      @f$ if propagate_down[0]
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
};

#ifdef USE_CUDNN
/**
 * @brief CuDNN acceleration of TanHLayer.
 */
template <typename Dtype>
class CuDNNTanHLayer : public TanHLayer<Dtype> {
 public:
  explicit CuDNNTanHLayer(const LayerParameter& param)
      : TanHLayer<Dtype>(param), handles_setup_(false) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual ~CuDNNTanHLayer();

 protected:
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  bool handles_setup_;
  cudnnHandle_t             handle_;
  cudnnTensorDescriptor_t bottom_desc_;
  cudnnTensorDescriptor_t top_desc_;
};
#endif

/**
 * @brief Tests whether the input exceeds a threshold: outputs 1 for inputs
 *        above threshold; 0 otherwise.
 */
template <typename Dtype>
class ThresholdLayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides ThresholdParameter threshold_param,
   *     with ThresholdLayer options:
   *   - threshold (\b optional, default 0).
   *     the threshold value @f$ t @f$ to which the input values are compared.
   */
  explicit ThresholdLayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}
  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "Threshold"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times H \times W) @f$
   *      the computed outputs @f$
   *       y = \left\{
   *       \begin{array}{lr}
   *         0 & \mathrm{if} \; x \le t \\
   *         1 & \mathrm{if} \; x > t
   *       \end{array} \right.
   *      @f$
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  /// @brief Not implemented (non-differentiable function)
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
    NOT_IMPLEMENTED;
  }

  Dtype threshold_;
};

/**
 * @brief Parameterized Rectified Linear Unit non-linearity @f$
 *        y_i = \max(0, x_i) + a_i \min(0, x_i)
 *        @f$. The differences from ReLULayer are 1) negative slopes are
 *        learnable though backprop and 2) negative slopes can vary across
 *        channels. The number of axes of input blob should be greater than or
 *        equal to 2. The 1st axis (0-based) is seen as channels.
 */
template <typename Dtype>
class PReLULayer : public NeuronLayer<Dtype> {
 public:
  /**
   * @param param provides PReLUParameter prelu_param,
   *     with PReLULayer options:
   *   - filler (\b optional, FillerParameter,
   *     default {'type': constant 'value':0.25}).
   *   - channel_shared (\b optional, default false).
   *     negative slopes are shared across channels.
   */
  explicit PReLULayer(const LayerParameter& param)
      : NeuronLayer<Dtype>(param) {}

  virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  virtual inline const char* type() const { return "PReLU"; }

 protected:
  /**
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times ...) @f$
   *      the inputs @f$ x @f$
   * @param top output Blob vector (length 1)
   *   -# @f$ (N \times C \times ...) @f$
   *      the computed outputs for each channel @f$i@f$ @f$
   *        y_i = \max(0, x_i) + a_i \min(0, x_i)
   *      @f$.
   */
  virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);
  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top);

  /**
   * @brief Computes the error gradient w.r.t. the PReLU inputs.
   *
   * @param top output Blob vector (length 1), providing the error gradient with
   *      respect to the outputs
   *   -# @f$ (N \times C \times ...) @f$
   *      containing error gradients @f$ \frac{\partial E}{\partial y} @f$
   *      with respect to computed outputs @f$ y @f$
   * @param propagate_down see Layer::Backward.
   * @param bottom input Blob vector (length 1)
   *   -# @f$ (N \times C \times ...) @f$
   *      the inputs @f$ x @f$; For each channel @f$i@f$, backward fills their
   *      diff with gradients @f$
   *        \frac{\partial E}{\partial x_i} = \left\{
   *        \begin{array}{lr}
   *            a_i \frac{\partial E}{\partial y_i} & \mathrm{if} \; x_i \le 0 \\
   *            \frac{\partial E}{\partial y_i} & \mathrm{if} \; x_i > 0
   *        \end{array} \right.
   *      @f$.
   *      If param_propagate_down_[0] is true, it fills the diff with gradients
   *      @f$
   *        \frac{\partial E}{\partial a_i} = \left\{
   *        \begin{array}{lr}
   *            \sum_{x_i} x_i \frac{\partial E}{\partial y_i} & \mathrm{if} \; x_i \le 0 \\
   *            0 & \mathrm{if} \; x_i > 0
   *        \end{array} \right.
   *      @f$.
   */
  virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

  bool channel_shared_;
  Blob<Dtype> multiplier_;  // dot multiplier for backward computation of params
  Blob<Dtype> backward_buff_;  // temporary buffer for backward computation
  Blob<Dtype> bottom_memory_;  // memory for in-place computation
};

}  // namespace caffe

#endif  // CAFFE_NEURON_LAYERS_HPP_