Caffe
Public Member Functions | Protected Member Functions | List of all members
caffe::HingeLossLayer< Dtype > Class Template Reference

Computes the hinge loss for a one-of-many classification task. More...

#include <hinge_loss_layer.hpp>

Inheritance diagram for caffe::HingeLossLayer< Dtype >:
caffe::LossLayer< Dtype > caffe::Layer< Dtype >

Public Member Functions

 HingeLossLayer (const LayerParameter &param)
 
virtual const char * type () const
 Returns the layer type.
 
- Public Member Functions inherited from caffe::LossLayer< Dtype >
 LossLayer (const LayerParameter &param)
 
virtual void LayerSetUp (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Does layer-specific setup: your layer should implement this function as well as Reshape. More...
 
virtual void Reshape (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Adjust the shapes of top blobs and internal buffers to accommodate the shapes of the bottom blobs. More...
 
virtual int ExactNumBottomBlobs () const
 Returns the exact number of bottom blobs required by the layer, or -1 if no exact number is required. More...
 
virtual bool AutoTopBlobs () const
 For convenience and backwards compatibility, instruct the Net to automatically allocate a single top Blob for LossLayers, into which they output their singleton loss, (even if the user didn't specify one in the prototxt, etc.).
 
virtual int ExactNumTopBlobs () const
 Returns the exact number of top blobs required by the layer, or -1 if no exact number is required. More...
 
virtual bool AllowForceBackward (const int bottom_index) const
 
- Public Member Functions inherited from caffe::Layer< Dtype >
 Layer (const LayerParameter &param)
 
void SetUp (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Implements common layer setup functionality. More...
 
Dtype Forward (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Given the bottom blobs, compute the top blobs and the loss. More...
 
void Backward (const vector< Blob< Dtype > *> &top, const vector< bool > &propagate_down, const vector< Blob< Dtype > *> &bottom)
 Given the top blob error gradients, compute the bottom blob error gradients. More...
 
vector< shared_ptr< Blob< Dtype > > > & blobs ()
 Returns the vector of learnable parameter blobs.
 
const LayerParameter & layer_param () const
 Returns the layer parameter.
 
virtual void ToProto (LayerParameter *param, bool write_diff=false)
 Writes the layer parameter to a protocol buffer.
 
Dtype loss (const int top_index) const
 Returns the scalar loss associated with a top blob at a given index.
 
void set_loss (const int top_index, const Dtype value)
 Sets the loss associated with a top blob at a given index.
 
virtual int MinBottomBlobs () const
 Returns the minimum number of bottom blobs required by the layer, or -1 if no minimum number is required. More...
 
virtual int MaxBottomBlobs () const
 Returns the maximum number of bottom blobs required by the layer, or -1 if no maximum number is required. More...
 
virtual int MinTopBlobs () const
 Returns the minimum number of top blobs required by the layer, or -1 if no minimum number is required. More...
 
virtual int MaxTopBlobs () const
 Returns the maximum number of top blobs required by the layer, or -1 if no maximum number is required. More...
 
virtual bool EqualNumBottomTopBlobs () const
 Returns true if the layer requires an equal number of bottom and top blobs. More...
 
bool param_propagate_down (const int param_id)
 Specifies whether the layer should compute gradients w.r.t. a parameter at a particular index given by param_id. More...
 
void set_param_propagate_down (const int param_id, const bool value)
 Sets whether the layer should compute gradients w.r.t. a parameter at a particular index given by param_id.
 

Protected Member Functions

virtual void Forward_cpu (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Computes the hinge loss for a one-of-many classification task. More...
 
virtual void Backward_cpu (const vector< Blob< Dtype > *> &top, const vector< bool > &propagate_down, const vector< Blob< Dtype > *> &bottom)
 Computes the hinge loss error gradient w.r.t. the predictions. More...
 
- Protected Member Functions inherited from caffe::Layer< Dtype >
virtual void Forward_gpu (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 Using the GPU device, compute the layer output. Fall back to Forward_cpu() if unavailable.
 
virtual void Backward_gpu (const vector< Blob< Dtype > *> &top, const vector< bool > &propagate_down, const vector< Blob< Dtype > *> &bottom)
 Using the GPU device, compute the gradients for any parameters and for the bottom blobs if propagate_down is true. Fall back to Backward_cpu() if unavailable.
 
virtual void CheckBlobCounts (const vector< Blob< Dtype > *> &bottom, const vector< Blob< Dtype > *> &top)
 
void SetLossWeights (const vector< Blob< Dtype > *> &top)
 

Additional Inherited Members

- Protected Attributes inherited from caffe::Layer< Dtype >
LayerParameter layer_param_
 
Phase phase_
 
vector< shared_ptr< Blob< Dtype > > > blobs_
 
vector< bool > param_propagate_down_
 
vector< Dtype > loss_
 

Detailed Description

template<typename Dtype>
class caffe::HingeLossLayer< Dtype >

Computes the hinge loss for a one-of-many classification task.

Parameters
bottominput Blob vector (length 2)
  1. $ (N \times C \times H \times W) $ the predictions $ t $, a Blob with values in $ [-\infty, +\infty] $ indicating the predicted score for each of the $ K = CHW $ classes. In an SVM, $ t $ is the result of taking the inner product $ X^T W $ of the D-dimensional features $ X \in \mathcal{R}^{D \times N} $ and the learned hyperplane parameters $ W \in \mathcal{R}^{D \times K} $, so a Net with just an InnerProductLayer (with num_output = D) providing predictions to a HingeLossLayer and no other learnable parameters or losses is equivalent to an SVM.
  2. $ (N \times 1 \times 1 \times 1) $ the labels $ l $, an integer-valued Blob with values $ l_n \in [0, 1, 2, ..., K - 1] $ indicating the correct class label among the $ K $ classes
topoutput Blob vector (length 1)
  1. $ (1 \times 1 \times 1 \times 1) $ the computed hinge loss: $ E = \frac{1}{N} \sum\limits_{n=1}^N \sum\limits_{k=1}^K [\max(0, 1 - \delta\{l_n = k\} t_{nk})] ^ p $, for the $ L^p $ norm (defaults to $ p = 1 $, the L1 norm; L2 norm, as in L2-SVM, is also available), and $ \delta\{\mathrm{condition}\} = \left\{ \begin{array}{lr} 1 & \mbox{if condition} \\ -1 & \mbox{otherwise} \end{array} \right. $

In an SVM, $ t \in \mathcal{R}^{N \times K} $ is the result of taking the inner product $ X^T W $ of the features $ X \in \mathcal{R}^{D \times N} $ and the learned hyperplane parameters $ W \in \mathcal{R}^{D \times K} $. So, a Net with just an InnerProductLayer (with num_output = $k$) providing predictions to a HingeLossLayer is equivalent to an SVM (assuming it has no other learned outside the InnerProductLayer and no other losses outside the HingeLossLayer).

Member Function Documentation

§ Backward_cpu()

template<typename Dtype >
void caffe::HingeLossLayer< Dtype >::Backward_cpu ( const vector< Blob< Dtype > *> &  top,
const vector< bool > &  propagate_down,
const vector< Blob< Dtype > *> &  bottom 
)
protectedvirtual

Computes the hinge loss error gradient w.r.t. the predictions.

Gradients cannot be computed with respect to the label inputs (bottom[1]), so this method ignores bottom[1] and requires !propagate_down[1], crashing if propagate_down[1] is set.

Parameters
topoutput Blob vector (length 1), providing the error gradient with respect to the outputs
  1. $ (1 \times 1 \times 1 \times 1) $ This Blob's diff will simply contain the loss_weight* $ \lambda $, as $ \lambda $ is the coefficient of this layer's output $\ell_i$ in the overall Net loss $ E = \lambda_i \ell_i + \mbox{other loss terms}$; hence $ \frac{\partial E}{\partial \ell_i} = \lambda_i $. (*Assuming that this top Blob is not used as a bottom (input) by any other layer of the Net.)
propagate_downsee Layer::Backward. propagate_down[1] must be false as we can't compute gradients with respect to the labels.
bottominput Blob vector (length 2)
  1. $ (N \times C \times H \times W) $ the predictions $t$; Backward computes diff $ \frac{\partial E}{\partial t} $
  2. $ (N \times 1 \times 1 \times 1) $ the labels – ignored as we can't compute their error gradients

Implements caffe::Layer< Dtype >.

§ Forward_cpu()

template<typename Dtype >
void caffe::HingeLossLayer< Dtype >::Forward_cpu ( const vector< Blob< Dtype > *> &  bottom,
const vector< Blob< Dtype > *> &  top 
)
protectedvirtual

Computes the hinge loss for a one-of-many classification task.

Parameters
bottominput Blob vector (length 2)
  1. $ (N \times C \times H \times W) $ the predictions $ t $, a Blob with values in $ [-\infty, +\infty] $ indicating the predicted score for each of the $ K = CHW $ classes. In an SVM, $ t $ is the result of taking the inner product $ X^T W $ of the D-dimensional features $ X \in \mathcal{R}^{D \times N} $ and the learned hyperplane parameters $ W \in \mathcal{R}^{D \times K} $, so a Net with just an InnerProductLayer (with num_output = D) providing predictions to a HingeLossLayer and no other learnable parameters or losses is equivalent to an SVM.
  2. $ (N \times 1 \times 1 \times 1) $ the labels $ l $, an integer-valued Blob with values $ l_n \in [0, 1, 2, ..., K - 1] $ indicating the correct class label among the $ K $ classes
topoutput Blob vector (length 1)
  1. $ (1 \times 1 \times 1 \times 1) $ the computed hinge loss: $ E = \frac{1}{N} \sum\limits_{n=1}^N \sum\limits_{k=1}^K [\max(0, 1 - \delta\{l_n = k\} t_{nk})] ^ p $, for the $ L^p $ norm (defaults to $ p = 1 $, the L1 norm; L2 norm, as in L2-SVM, is also available), and $ \delta\{\mathrm{condition}\} = \left\{ \begin{array}{lr} 1 & \mbox{if condition} \\ -1 & \mbox{otherwise} \end{array} \right. $

In an SVM, $ t \in \mathcal{R}^{N \times K} $ is the result of taking the inner product $ X^T W $ of the features $ X \in \mathcal{R}^{D \times N} $ and the learned hyperplane parameters $ W \in \mathcal{R}^{D \times K} $. So, a Net with just an InnerProductLayer (with num_output = $k$) providing predictions to a HingeLossLayer is equivalent to an SVM (assuming it has no other learned outside the InnerProductLayer and no other losses outside the HingeLossLayer).

Implements caffe::Layer< Dtype >.


The documentation for this class was generated from the following files: