diff options
| -rw-r--r-- | SpatialLogSoftMax.lua | 17 | ||||
| -rw-r--r-- | generic/SpatialLogSoftMax.c | 116 | ||||
| -rw-r--r-- | init.c | 5 | ||||
| -rw-r--r-- | init.lua | 1 | ||||
| -rw-r--r-- | nnx-1.0-1.rockspec | 1 |
5 files changed, 140 insertions, 0 deletions
diff --git a/SpatialLogSoftMax.lua b/SpatialLogSoftMax.lua new file mode 100644 index 0000000..d145caa --- /dev/null +++ b/SpatialLogSoftMax.lua @@ -0,0 +1,17 @@ +local SpatialLogSoftMax, parent = torch.class('nn.SpatialLogSoftMax', 'nn.Module') + +function SpatialLogSoftMax:__init() + parent.__init(self) +end + +function SpatialLogSoftMax:forward(input) + self.output:resizeAs(input) + input.nn.SpatialLogSoftMax_forward(self, input) + return self.output +end + +function SpatialLogSoftMax:backward(input, gradOutput) + self.gradInput:resizeAs(input) + input.nn.SpatialLogSoftMax_backward(self, input, gradOutput) + return self.gradInput +end diff --git a/generic/SpatialLogSoftMax.c b/generic/SpatialLogSoftMax.c new file mode 100644 index 0000000..0fc3d9d --- /dev/null +++ b/generic/SpatialLogSoftMax.c @@ -0,0 +1,116 @@ +#ifndef TH_GENERIC_FILE +#define TH_GENERIC_FILE "generic/SpatialLogSoftMax.c" +#else + +static int nn_(SpatialLogSoftMax_forward)(lua_State *L) +{ + // get all params + THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); + THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id)); + + // dims + int width = input->size[0]; + int height = input->size[1]; + + // select planes + THTensor *input_row = THTensor_(new)(); + THTensor *input_point = THTensor_(new)(); + THTensor *output_row = THTensor_(new)(); + THTensor *output_point = THTensor_(new)(); + + // process the whole plane + int x,y; + for (y=0; y<height; y++) { + THTensor_(select)(input_row, input, 1, y); + THTensor_(select)(output_row, output, 1, y); + for (x=0; x<width; x++) { + THTensor_(select)(input_point, input_row, 1, x); + THTensor_(select)(output_point, output_row, 1, x); + + real sum = THLogZero; + + TH_TENSOR_APPLY2(real, output_point, real, input_point, \ + real z = *input_point_data; \ + *output_point_data = z; \ + sum = THLogAdd(sum, z);) + + THTensor_(add)(output_point, -sum); + } + } + + // cleanup + THTensor_(free)(input_row); + THTensor_(free)(input_point); + THTensor_(free)(output_row); + THTensor_(free)(output_point); + return 1; +} + +static int nn_(SpatialLogSoftMax_backward)(lua_State *L) +{ + // get all params + THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); + THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id)); + THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_(Tensor_id)); + THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id)); + + // dims + int width = input->size[0]; + int height = input->size[1]; + + // zero gradInput + THTensor_(zero)(gradInput); + + // select planes + THTensor *gradOutput_row = THTensor_(new)(); + THTensor *gradOutput_point = THTensor_(new)(); + THTensor *gradInput_row = THTensor_(new)(); + THTensor *gradInput_point = THTensor_(new)(); + THTensor *output_row = THTensor_(new)(); + THTensor *output_point = THTensor_(new)(); + + // compute gradients for each point + int x,y; + for (y=0; y<height; y++) { + THTensor_(select)(gradInput_row, gradInput, 1, y); + THTensor_(select)(gradOutput_row, gradOutput, 1, y); + THTensor_(select)(output_row, output, 1, y); + for (x=0; x<width; x++) { + THTensor_(select)(gradInput_point, gradInput_row, 1, x); + THTensor_(select)(gradOutput_point, gradOutput_row, 1, x); + THTensor_(select)(output_point, output_row, 1, x); + + real sum = THTensor_(sum)(gradOutput_point); + + TH_TENSOR_APPLY3(real, gradInput_point, \ + real, gradOutput_point, \ + real, output_point, \ + *gradInput_point_data = *gradOutput_point_data - exp(*output_point_data)*sum;); + } + } + + // cleanup + THTensor_(free)(gradInput_row); + THTensor_(free)(gradInput_point); + THTensor_(free)(gradOutput_row); + THTensor_(free)(gradOutput_point); + THTensor_(free)(output_row); + THTensor_(free)(output_point); + + return 1; +} + +static const struct luaL_Reg nn_(SpatialLogSoftMax__) [] = { + {"SpatialLogSoftMax_forward", nn_(SpatialLogSoftMax_forward)}, + {"SpatialLogSoftMax_backward", nn_(SpatialLogSoftMax_backward)}, + {NULL, NULL} +}; + +static void nn_(SpatialLogSoftMax_init)(lua_State *L) +{ + luaT_pushmetaclass(L, torch_(Tensor_id)); + luaT_registeratname(L, nn_(SpatialLogSoftMax__), "nn"); + lua_pop(L,1); +} + +#endif @@ -20,6 +20,9 @@ static const void* torch_DoubleTensor_id = NULL; #include "generic/SpatialConvolutionTable.c" #include "THGenerateFloatTypes.h" +#include "generic/SpatialLogSoftMax.c" +#include "THGenerateFloatTypes.h" + #include "generic/Threshold.c" #include "THGenerateFloatTypes.h" @@ -33,12 +36,14 @@ DLL_EXPORT int luaopen_libnnx(lua_State *L) nn_FloatAbs_init(L); nn_FloatThreshold_init(L); nn_FloatSpatialConvolutionTable_init(L); + nn_FloatSpatialLogSoftMax_init(L); nn_DoubleSpatialLinear_init(L); nn_DoubleHardShrink_init(L); nn_DoubleAbs_init(L); nn_DoubleThreshold_init(L); nn_DoubleSpatialConvolutionTable_init(L); + nn_DoubleSpatialLogSoftMax_init(L); return 1; } @@ -60,6 +60,7 @@ torch.include('nnx', 'Narrow.lua') -- spatial (images) operators: torch.include('nnx', 'SpatialLinear.lua') +torch.include('nnx', 'SpatialLogSoftMax.lua') torch.include('nnx', 'SpatialConvolutionTable.lua') -- criterions: diff --git a/nnx-1.0-1.rockspec b/nnx-1.0-1.rockspec index 49284a8..44ae451 100644 --- a/nnx-1.0-1.rockspec +++ b/nnx-1.0-1.rockspec @@ -57,6 +57,7 @@ build = { install_files(/lua/nnx Sqrt.lua) install_files(/lua/nnx Threshold.lua) install_files(/lua/nnx SpatialConvolutionTable.lua) + install_files(/lua/nnx SpatialLogSoftMax.lua) install_files(/lua/nnx SpatialLinear.lua) install_files(/lua/nnx SuperCriterion.lua) add_subdirectory (test) |