void ProjectionLayer::BackPropagate( MatrixBase& outDiff,
const MatrixBase& in,
const MatrixBase& out,
float learningRate,
MatrixBase* inDiff ) {
int n_example = outDiff.Rows();
int n_project = m_projection.Columns();
int n_order = outDiff.Columns() / n_project;
ASSERT( n_order == in.Columns() );
bool useMomentum = m_momentum != 0.0f && m_weight_decay != 0.0f;
if ( m_norm_type == "Clip" ) {
outDiff.Clip( -m_norm_param, m_norm_param );
}
if ( m_norm_type == "L2Norm" ) {
float norm = outDiff.L2Norm();
if ( norm > m_norm_param ) {
outDiff.Scale( 1.0f / norm );
}
}
if ( useMomentum &&
(m_gradient.Rows() != m_projection.Rows() || m_gradient.Columns() != m_projection.Columns() ) ) {
m_gradient.Reshape( m_projection.Rows(), m_projection.Columns(), kSetZero );
}
for ( int i = 0; i < n_order; i++ ) {
SubMatrix gradient( outDiff,
MatrixRange(0, n_example, i * n_project, (i + 1) * n_project ) );
SubMatrix index( in,
MatrixRange(0, n_example, i, i + 1) );
if ( useMomentum ) {
m_gradient.Add( m_momentum, -learningRate * m_weight_decay, m_projection );
m_gradient.AddRows( gradient, index, -learningRate / n_example );
}
else {
m_projection.AddRows( gradient, index, -learningRate / n_example );
}
}
if ( useMomentum ) {
m_projection.Add( 1.0f, 1.0f, m_gradient );
}
} // end of BackPropagate
void vFSMNLayer::Compute( const MatrixBase& feature, MatrixBase* output ) {
m_linear.Compute( feature, output );
m_memory.Reshape( feature.Rows(), feature.Columns() );
m_memory.VfsmnMemory( feature, m_filter, m_position );
output->Sgemm( 1.0f, 1.0f, m_memory, CUBLAS_OP_N, m_weight, CUBLAS_OP_N );
output->ReLU( *output );
}
void sFSMNLayer::Compute( const MatrixBase& feature, MatrixBase* output ) {
m_linear.Compute( feature, output );
m_memory.Reshape( feature.Rows(), feature.Columns() );
m_memory.Strmm( 1.0f,
m_block_diagonal, CUBLAS_SIDE_LEFT, CUBLAS_FILL_MODE_LOWER,
feature, CUBLAS_OP_N ); // @xmb20160226
// m_memory.Sgemm( 0.0f, 1.0f, m_block_diagonal, CUBLAS_OP_N, feature, CUBLAS_OP_N );
output->Sgemm( 1.0f, 1.0f, m_memory, CUBLAS_OP_N, m_weight, CUBLAS_OP_N );
output->ReLU( *output );
}
void ProjectionLayer::Compute( const MatrixBase& feature, MatrixBase* output ) {
int n_order = feature.Columns();
int n_example = feature.Rows();
int n_project = m_projection.Columns();
assert( feature.Rows() == output->Rows() );
assert( output->Columns() == n_order * m_projection.Columns() );
for ( int i = 0; i < n_order; i++ ) {
SubMatrix gram( *output,
MatrixRange(0, n_example, i * n_project, (i + 1) * n_project ) );
SubMatrix index( feature,
MatrixRange(0, n_example, i, i + 1) );
gram.GetRows( m_projection, index );
}
} // end of ForwardPropagate
void sFSMNLayer::BackPropagate( MatrixBase& outDiff,
const MatrixBase& in,
const MatrixBase& out,
float learningRate,
MatrixBase* inDiff ) {
outDiff.DiffReLU( out, outDiff );
m_linear.BackPropagate( outDiff, in, Matrix(), learningRate, inDiff );
if ( m_w_momentum.Rows() != m_weight.Rows() ||
m_w_momentum.Columns() != m_weight.Columns() ) {
m_w_momentum.Reshape( m_weight.Rows(), m_weight.Columns(), kSetZero );
}
float avg_lr = -learningRate / (float)outDiff.Rows();
m_w_momentum.Sgemm( m_momentum,
avg_lr,
m_memory,
CUBLAS_OP_T,
outDiff,
CUBLAS_OP_N );
if ( m_weight_decay != 0.0f ) {
m_w_momentum.Add( 1.0f, -learningRate * m_weight_decay, m_weight );
}
m_memory_diff.Reshape( m_memory.Rows(), m_memory.Columns() );
m_memory_diff.Sgemm( 0.0f, 1.0f, outDiff, CUBLAS_OP_N, m_weight, CUBLAS_OP_T );
if ( NULL != inDiff ) {
inDiff->Sgemm( 1.0f,
1.0f,
m_block_diagonal,
CUBLAS_OP_T,
m_memory_diff,
CUBLAS_OP_N );
}
m_diagonal_diff.Reshape( m_block_diagonal.Rows(), m_block_diagonal.Columns() );
m_diagonal_diff.Sgemm( 0.0f,
avg_lr / in.Columns(),
m_memory_diff,
CUBLAS_OP_N,
in,
CUBLAS_OP_T );
if ( m_norm_type == "Clip" ) {
float range = m_norm_param * (-avg_lr);
m_w_momentum.Clip( -range, range );
m_diagonal_diff.Clip( -range, range );
}
if ( m_norm_type == "L2Norm" ) {
float w_norm = m_w_momentum.L2Norm() * (-avg_lr);
if ( w_norm > m_norm_param ) {
m_w_momentum.Scale( 1.0f / w_norm );
}
float d_norm = m_diagonal_diff.L2Norm() * (-avg_lr);
if ( d_norm > m_norm_param ) {
m_diagonal_diff.Scale( 1.0f / d_norm );
}
}
m_weight.Add( 1.0f, 1.0f, m_w_momentum );
m_filter.UpdateSfsmnFilter( m_length, m_diagonal_diff );
}
