Config class for the CollapseModel function. More...
#include <nnet-utils.h>
Public Member Functions | |
CollapseModelConfig () | |
Public Attributes | |
bool | collapse_dropout |
bool | collapse_batchnorm |
bool | collapse_affine |
bool | collapse_scale |
Config class for the CollapseModel function.
This function is reponsible for collapsing together sequential components where doing so could make the test-time operation more efficient. For example, dropout components and batch-norm components that are in test mode can be combined with the next layer; and if there are successive affine components it may also be possible to combine these under some circumstances.
It expects batch-norm components to be in test mode; you should probably call SetBatchnormTestMode() and SetDropoutTestMode() before CollapseModel().
Definition at line 240 of file nnet-utils.h.
|
inline |
Definition at line 245 of file nnet-utils.h.
References kaldi::nnet3::ApplyL2Regularization(), kaldi::nnet3::CollapseModel(), kaldi::nnet3::ConsolidateMemory(), kaldi::nnet3::ConstrainOrthonormal(), kaldi::nnet3::GetNumNvalues(), kaldi::nnet3::ReadEditConfig(), kaldi::nnet3::ScaleBatchnormStats(), and kaldi::nnet3::UpdateNnetWithMaxChange().
bool collapse_affine |
Definition at line 243 of file nnet-utils.h.
bool collapse_batchnorm |
Definition at line 242 of file nnet-utils.h.
bool collapse_dropout |
Definition at line 241 of file nnet-utils.h.
bool collapse_scale |
Definition at line 244 of file nnet-utils.h.