void CNeuralLinearLayer::compute_gradients(
SGVector<float64_t> parameters,
SGMatrix<float64_t> targets,
CDynamicObjectArray* layers,
SGVector<float64_t> parameter_gradients)
{
compute_local_gradients(targets);
// compute bias gradients
float64_t* bias_gradients = parameter_gradients.vector;
#ifdef HAVE_EIGEN3
typedef Eigen::Map<Eigen::MatrixXd> EMappedMatrix;
typedef Eigen::Map<Eigen::VectorXd> EMappedVector;
EMappedVector BG(bias_gradients, m_num_neurons);
EMappedMatrix LG(m_local_gradients.matrix, m_num_neurons, m_batch_size);
BG = LG.rowwise().sum();
#else
for (int32_t i=0; i<m_num_neurons; i++)
{
float64_t sum = 0;
for (int32_t j=0; j<m_batch_size; j++)
{
sum += m_local_gradients[i+j*m_num_neurons];
}
bias_gradients[i] = sum;
}
#endif
// apply dropout to the local gradients
if (dropout_prop>0.0)
{
int32_t len = m_num_neurons*m_batch_size;
for (int32_t i=0; i<len; i++)
m_local_gradients[i] *= m_dropout_mask[i];
}
int32_t weights_index_offset = m_num_neurons;
for (int32_t l=0; l<m_input_indices.vlen; l++)
{
CNeuralLayer* layer =
(CNeuralLayer*)layers->element(m_input_indices[l]);
float64_t* weights = parameters.vector + weights_index_offset;
float64_t* weight_gradients = parameter_gradients.vector +
weights_index_offset;
weights_index_offset += m_num_neurons*layer->get_num_neurons();
#ifdef HAVE_EIGEN3
EMappedMatrix X(layer->get_activations().matrix,
layer->get_num_neurons(), m_batch_size);
EMappedMatrix W(weights, m_num_neurons, layer->get_num_neurons());
EMappedMatrix WG(weight_gradients,
m_num_neurons, layer->get_num_neurons());
EMappedMatrix IG(layer->get_activation_gradients().matrix,
layer->get_num_neurons(), m_batch_size);
// compute weight gradients
WG = LG*X.transpose();
// compute input gradients
if (!layer->is_input())
IG += W.transpose()*LG;
#else
// weight_gradients=local_gradients*previous_layer_activations.T
for (int32_t i=0; i<m_num_neurons; i++)
{
for (int32_t j=0; j<layer->get_num_neurons(); j++)
{
float64_t sum = 0;
for (int32_t k=0; k<m_batch_size; k++)
{
sum += m_local_gradients(i,k)*layer->get_activations()(j,k);
}
weight_gradients[i+j*m_num_neurons] = sum;
}
}
if (!layer->is_input())
{
// input_gradients = weights.T*local_gradients
for (int32_t i=0; i<layer->get_num_neurons(); i++)
{
for (int32_t j=0; j<m_batch_size; j++)
{
float64_t sum = 0;
for (int32_t k=0; k<m_num_neurons; k++)
{
sum += weights[k+i*m_num_neurons]*
m_local_gradients[k+j*m_num_neurons];
}
layer->get_activation_gradients()(i,j) += sum;
}
}
}
#endif
SG_UNREF(layer);
}
if (contraction_coefficient != 0)
{
compute_contraction_term_gradients(parameters, parameter_gradients);
}
}
void CConvolutionalFeatureMap::compute_gradients(
SGVector< float64_t > parameters,
SGMatrix<float64_t> activations,
SGMatrix< float64_t > activation_gradients,
CDynamicObjectArray* layers,
SGVector< int32_t > input_indices,
SGVector< float64_t > parameter_gradients)
{
int32_t batch_size = activation_gradients.num_cols;
if (m_activation_function==CMAF_LOGISTIC)
{
for (int32_t i=0; i<m_output_num_neurons; i++)
{
for (int32_t j=0; j<batch_size; j++)
{
activation_gradients(i+m_row_offset,j) *=
activation_gradients(i+m_row_offset,j) *
(1.0-activation_gradients(i+m_row_offset,j));
}
}
}
else if (m_activation_function==CMAF_RECTIFIED_LINEAR)
{
for (int32_t i=0; i<m_output_num_neurons; i++)
for (int32_t j=0; j<batch_size; j++)
if (activations(i+m_row_offset,j)==0)
activation_gradients(i+m_row_offset,j) = 0;
}
float64_t bias_gradient = 0;
for (int32_t i=0; i<m_output_num_neurons; i++)
for (int32_t j=0; j<batch_size; j++)
bias_gradient += activation_gradients(i+m_row_offset,j);
parameter_gradients[0] = bias_gradient;
int32_t weights_index_offset = 1;
for (int32_t l=0; l<input_indices.vlen; l++)
{
CNeuralLayer* layer =
(CNeuralLayer*)layers->element(input_indices[l]);
int32_t num_maps = layer->get_num_neurons()/m_input_num_neurons;
for (int32_t m=0; m<num_maps; m++)
{
SGMatrix<float64_t> W(parameters.vector+weights_index_offset,
m_filter_height, m_filter_width, false);
SGMatrix<float64_t> WG(parameter_gradients.vector+weights_index_offset,
m_filter_height, m_filter_width, false);
weights_index_offset += m_filter_height*m_filter_width;
compute_weight_gradients(layer->get_activations(),
activation_gradients, WG, m*m_input_num_neurons, m_row_offset);
if (!layer->is_input())
convolve(activation_gradients, W,
layer->get_activation_gradients(), true, false,
m_row_offset, m*m_input_num_neurons);
}
SG_UNREF(layer);
}
}