Apache SINGA
A distributed deep learning platform .
|
some extension of expressions, used to support something beyond elementwise op More...
Go to the source code of this file.
Namespaces | |
mshadow | |
namespace for mshadow | |
mshadow::expr | |
namespace for abstract expressions and expressions template, have no dependecy on tensor.h, These data structure takes no charge in computations, they are only used to define operations and represent expression in a symbolic way | |
Functions | |
template<typename E , typename R , int d> | |
ReduceTo1DExp< E, R, d > | mshadow::expr::operator* (const ReduceTo1DExp< E, R, d > &e, real_t scale) |
operator overload | |
template<typename E , typename R , int d> | |
ReduceTo1DExp< E, R, d > | mshadow::expr::operator* (real_t scale, const ReduceTo1DExp< E, R, d > &e) |
operator overload | |
template<int dimcast, typename Device , int dimdst> | |
Broadcast1DExp< Device, dimdst, dimcast > | mshadow::expr::broadcast (const Tensor< Device, 1 > &src, Shape< dimdst > shape) |
a expression that replicate a 1 dimension tensor in dimension dimcast More... | |
template<typename SrcExp , int etype> | |
UnpackPatchToColXExp< SrcExp, ExpInfo< SrcExp >::kDim > | mshadow::expr::unpack_patch2col (const Exp< SrcExp, etype > &img, index_t psize, index_t pstride) |
unpack local (overlap) patches of image to column of mat, can be used to implement convolution after getting unpacked mat, we can use: output = dot( weight, mat ) to get covolved results, the relations: More... | |
template<typename Device , int dstdim> | |
PackColToPatchXExp< Device, dstdim > | mshadow::expr::pack_col2patch (const Tensor< Device, 2 > &mat, Shape< dstdim > imshape, index_t psize, index_t pstride) |
reverse operation of pack_col2patch, can be used to implement deconvolution More... | |
template<typename SrcExp , int etype, int dimdst> | |
ReshapeExp< SrcExp, dimdst, ExpInfo< SrcExp >::kDim > | mshadow::expr::reshape (const Exp< SrcExp, etype > &src, Shape< dimdst > oshape) |
a expression that reshapes a tensor to another shape More... | |
template<int a1, int a2, typename SrcExp , int etype> | |
SwapAxisExp< SrcExp, ExpInfo < SrcExp >::kDim, a1, a2 > | mshadow::expr::swapaxis (const Exp< SrcExp, etype > &src) |
a expression that reshapes a tensor to another shape More... | |
template<int dimkeep, typename SrcExp , int etype> | |
ReduceTo1DExp< SrcExp, red::sum, dimkeep > | mshadow::expr::sumall_except_dim (const Exp< SrcExp, etype > &exp) |
a sum over all dimensions, except dimkeep More... | |
template<typename Reducer , typename SrcExp , int etype> | |
PoolingExp< Reducer, SrcExp, ExpInfo< SrcExp >::kDim > | mshadow::expr::pool (const Exp< SrcExp, etype > &src, index_t ksize, index_t kstride) |
pooling subregion results together More... | |
template<typename Reducer , typename SrcExp , int etype> | |
PoolingExp< Reducer, SrcExp, ExpInfo< SrcExp >::kDim > | mshadow::expr::pool (const Exp< SrcExp, etype > &src, Shape< 2 > pshape, index_t ksize, index_t kstride) |
same as pool, except the output shape is specified by pshape More... | |
template<typename Reducer , typename Device > | |
UnPoolingExp< Reducer, Device > | mshadow::expr::unpool (const Tensor< Device, 4 > &data_src, const Tensor< Device, 4 > &data_pooled, const Tensor< Device, 4 > &grad_pooled, index_t ksize, index_t kstride) |
unpooling gradient for 4D, backprop gradient value back, revserse operation of pooling More... | |
template<typename SrcExp , int etype> | |
PaddingExp< SrcExp, ExpInfo < SrcExp >::kDim > | mshadow::expr::pad (const Exp< SrcExp, etype > &src, index_t pad) |
padding expression, pad a image with zeros on boundaries, padding affects shape[0], and shape[1] More... | |
template<typename SrcExp , int etype> | |
CroppingExp< SrcExp, ExpInfo < SrcExp >::kDim > | mshadow::expr::crop (const Exp< SrcExp, etype > &src, Shape< 2 > oshape) |
revserse operationg of padding, cut off boundaries, crop output from center of input More... | |
template<typename SrcExp , int etype> | |
CroppingExp< SrcExp, ExpInfo < SrcExp >::kDim > | mshadow::expr::crop (const Exp< SrcExp, etype > &src, Shape< 2 > oshape, index_t start_height, index_t start_width) |
same as crop, but can specify starting position to do cropping More... | |
template<typename SrcExp , int etype> | |
MirroringExp< SrcExp, ExpInfo < SrcExp >::kDim > | mshadow::expr::mirror (const Exp< SrcExp, etype > &src) |
mirroring expression, mirror images in width More... | |
template<typename Reducer , typename SrcExp , int etype> | |
ChannelPoolingExp< Reducer, SrcExp, ExpInfo< SrcExp > ::kDim > | mshadow::expr::chpool (const Exp< SrcExp, etype > &src, index_t nsize) |
channel pooling, do reduction over (local nearby) channels, used to implement local response normalization More... | |
template<typename Device > | |
Broadcast1DExp< Device, 2, 0 > | mshadow::expr::repmat (const Tensor< Device, 1 > &src, index_t nrow) |
a expression that replicate a 1 dimension tensor for nrow times More... | |
template<typename SrcExp , int etype> | |
ReduceTo1DExp< SrcExp, red::sum, 0 > | mshadow::expr::sum_rows (const Exp< SrcExp, etype > &exp) |
a expression that sum over rows of a matrix More... | |
some extension of expressions, used to support something beyond elementwise op