static int nn_(SpatialConvolution_updateGradInput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
THTensor *tweight;
THArgCheck( nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1, "Number of output features is not equal to nOutputPlane" );
/* gradient to input */
tweight = THTensor_(newTranspose)(weight,0,1);
if (input->nDimension == 3)
{
THTensor_(conv2Dmv)(gradInput, 0.0, 1.0, gradOutput, tweight, dH, dW, "F","C");
}
else
{
THTensor_(conv2Dmm)(gradInput, 0.0, 1.0, gradOutput, tweight, dH, dW, "F","C");
}
THTensor_(free)(tweight);
return 1;
}
static int nn_(SpatialMaxPooling_updateGradInput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_Tensor);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
// get contiguous gradOutput
gradOutput = THTensor_(newContiguous)(gradOutput);
// resize
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
// sizes
int ichannels = input->size[0];
int iheight = input->size[1];
int iwidth = input->size[2];
int ochannels = ichannels;
int oheight = gradOutput->size[1];
int owidth = gradOutput->size[2];
// get raw pointers
real *gradInput_data = THTensor_(data)(gradInput);
real *gradOutput_data = THTensor_(data)(gradOutput);
real *indices_data = THTensor_(data)(indices);
// backprop
long k;
for (k = 0; k < input->size[0]; k++) {
// pointers to slices
real *gradOutput_p = gradOutput_data + k*owidth*oheight;
real *gradInput_p = gradInput_data + k*iwidth*iheight;
real *indy_p = indices_data + k*owidth*oheight;
real *indx_p = indices_data + (k+ochannels)*owidth*oheight;
// calculate max points
int i,j;
for(i = 0; i < oheight; i++) {
for(j = 0; j < owidth; j++) {
// retrieve position of max
long maxi = *(indy_p + i*owidth + j) - 1 + i*dH;
long maxj = *(indx_p + i*owidth + j) - 1 + j*dW;
// update gradient
*(gradInput_p + maxi*iwidth + maxj) += *(gradOutput_p + i*owidth + j);
}
}
}
// cleanup
THTensor_(free)(gradOutput);
return 1;
}
static int nnOmp_(SpatialMaxPooling_updateGradInputOmp)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id));
THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id));
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
setompnthread(L,1,"nThread");
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_(Tensor_id));
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_(Tensor_id));
THTensor *gradOutputPlane, *gradInputPlane, *unfoldedGradInputPlane, *gradLocalInput;
int k,i,j;
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
gradInputPlane = THTensor_(new)();
gradOutputPlane = THTensor_(new)();
gradLocalInput = THTensor_(new)();
unfoldedGradInputPlane = THTensor_(new)();
for (k = 0; k < input->size[0]; k++)
{
/* get input and output plane */
THTensor_(select)(gradOutputPlane, gradOutput, 0, k);
THTensor_(select)(gradInputPlane, gradInput, 0, k);
/* Unfold input to get each local window */
THTensor_(unfold)(unfoldedGradInputPlane, gradInputPlane, 0, kH, dH);
THTensor_(unfold)(unfoldedGradInputPlane, NULL, 1, kW, dW);
/* Calculate max points */
for(i = 0; i < gradOutputPlane->size[0]; i++) {
for(j = 0; j < gradOutputPlane->size[1]; j++) {
THTensor_(select)(gradLocalInput, unfoldedGradInputPlane,0,i);
THTensor_(select)(gradLocalInput, NULL, 0,j);
long maxi = THTensor_(get4d)(indices,0,k,i,j)-1;
long maxj = THTensor_(get4d)(indices,1,k,i,j)-1;
double gi = THTensor_(get2d)(gradLocalInput,maxi,maxj)+THTensor_(get2d)(gradOutputPlane,i,j);
THTensor_(set2d)(gradLocalInput,maxi,maxj,gi);
}
}
}
/* Cleanup */
THTensor_(free)(gradInputPlane);
THTensor_(free)(gradOutputPlane);
THTensor_(free)(unfoldedGradInputPlane);
THTensor_(free)(gradLocalInput);
return 1;
}
static int nn_(SpatialConvolutionMM_updateGradInput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *finput = luaT_getfieldcheckudata(L, 1, "finput", torch_Tensor);
THTensor *fgradInput = luaT_getfieldcheckudata(L, 1, "fgradInput", torch_Tensor);
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
THArgCheck( nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1, "Number of output features is not equal to nOutputPlane" );
THTensor_(resizeAs)(gradInput, input);
THTensor_(resizeAs)(fgradInput, finput);
THTensor_(transpose)(weight, weight, 0, 1);
if(input->nDimension == 3)
{
nn_(SpatialConvolutionMM_updateGradInput_frame)(gradInput, gradOutput, weight, fgradInput, kW, kH);
}
else
{
long T = input->size[0];
long t;
THStorage_(clearFlag)(gradInput->storage, TH_STORAGE_REFCOUNTED);
THStorage_(clearFlag)(gradOutput->storage, TH_STORAGE_REFCOUNTED);
THStorage_(clearFlag)(fgradInput->storage, TH_STORAGE_REFCOUNTED);
#pragma omp parallel for private(t)
for(t = 0; t < T; t++)
{
THTensor *gradInput_t = THTensor_(newSelect)(gradInput, 0, t);
THTensor *gradOutput_t = THTensor_(newSelect)(gradOutput, 0, t);
THTensor *fgradInput_t = THTensor_(newSelect)(fgradInput, 0, t);
nn_(SpatialConvolutionMM_updateGradInput_frame)(gradInput_t, gradOutput_t, weight, fgradInput_t, kW, kH);
THTensor_(free)(gradInput_t);
THTensor_(free)(gradOutput_t);
THTensor_(free)(fgradInput_t);
}
THStorage_(setFlag)(gradInput->storage, TH_STORAGE_REFCOUNTED);
THStorage_(setFlag)(gradOutput->storage, TH_STORAGE_REFCOUNTED);
THStorage_(setFlag)(fgradInput->storage, TH_STORAGE_REFCOUNTED);
}
THTensor_(transpose)(weight, weight, 0, 1);
return 1;
}
static int nn_TemporalSubSampling_forward(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor_id);
int kW = luaT_getfieldcheckint(L, 1, "kW");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int inputFrameSize = luaT_getfieldcheckint(L, 1, "inputFrameSize");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor_id);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor_id);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor_id);
THTensor *unfoldedInput, *unfoldedInputFrame, *unfoldedInputFrames;
THTensor *outputFrame;
int nInputFrame, nOutputFrame;
int i, k;
luaL_argcheck(L, input->nDimension == 2, 2, "2D tensor expected");
luaL_argcheck(L, input->size[0] == inputFrameSize, 2, "invalid input frame size");
luaL_argcheck(L, input->size[1] >= kW, 2, "input sequence smaller than kernel size");
nInputFrame = input->size[1];
nOutputFrame = (nInputFrame - kW) / dW + 1;
THTensor_resize2d(output,
inputFrameSize,
nOutputFrame);
outputFrame = THTensor_new();
unfoldedInput = THTensor_new();
unfoldedInputFrame = THTensor_new();
unfoldedInputFrames = THTensor_new();
THTensor_unfold(unfoldedInput, input, 1, kW, dW);
for(k = 0; k < nOutputFrame; k++)
{
THTensor_select(unfoldedInputFrames, unfoldedInput, 1, k);
THTensor_select(outputFrame, output, 1, k);
THTensor_zero(outputFrame);
for(i = 0; i < kW; i++)
{
THTensor_select(unfoldedInputFrame, unfoldedInputFrames, 1, i);
THTensor_addTensor(outputFrame, 1, unfoldedInputFrame);
}
THTensor_cmul(outputFrame, weight);
THTensor_addTensor(outputFrame, 1, bias);
}
THTensor_free(outputFrame);
THTensor_free(unfoldedInput);
THTensor_free(unfoldedInputFrame);
THTensor_free(unfoldedInputFrames);
return 1;
}
static int gpunn_SpatialConvolutionGPU_accGradParameters(lua_State *L)
{
THGPUTensor *input = (THGPUTensor *)luaT_checkudata(L, 2, "torch.GPUTensor");
THGPUTensor *gradOutput = (THGPUTensor *)luaT_checkudata(L, 3, "torch.GPUTensor");
THGPUTensor *gradWeight = (THGPUTensor *)luaT_getfieldcheckudata(L, 1, "gradWeight", "torch.GPUTensor");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int padding = luaT_getfieldcheckint(L, 1, "padding");
int partialSum = luaT_getfieldcheckint(L, 1, "partialSum");
float scale = luaL_optnumber(L, 4, 1);
long nOutputPlane = gradWeight->size[3];
long nInputPlane = gradWeight->size[0];
long kH = gradWeight->size[1];
long kW = gradWeight->size[2];
long inputHeight = input->size[1];
long inputWidth = input->size[2];
long batchSize = input->size[3];
long outputHeight = (padding + inputHeight - kH) / dH + 1;
long outputWidth = (padding + inputWidth - kW) / dW + 1;
// asserts
luaL_argcheck(L, inputWidth == inputHeight, 1, "input must be square");
luaL_argcheck(L, kH == kW, 1, "kH must be equal to kW");
luaL_argcheck(L, dH == dW, 1, "dH must be equal to dW");
if (partialSum)
{
// compute partial gradients for outputHeight*outputWidth/partialSum groups of filters separately
gradWeight = (THGPUTensor *)luaT_getfieldcheckudata(L, 1, "gradWeightPartial", "torch.GPUTensor");
THGPUTensor_resize4d(gradWeight, outputHeight * outputWidth / partialSum, nInputPlane, kH * kW,
nOutputPlane);
// numModuleY*numModulesX/partialSum, numFilterColors, filterPixels, numFilters
}
// all the data must be contiguous:
luaL_argcheck(L, THGPUTensor_isContiguous(input), 2, "input must be contiguous");
luaL_argcheck(L, THGPUTensor_isContiguous(gradWeight), 1, "weight must be contiguous");
luaL_argcheck(L, THGPUTensor_isContiguous(gradOutput), 1, "output must be contiguous");
auto avInput = input->get_array_view();
auto avGradOutput = gradOutput->get_array_view();
auto avGradWeight = gradWeight->get_array_view();
// convolutions
spatialConv_accGradParameters(avInput, avGradOutput, avGradWeight, nInputPlane, inputHeight,
inputWidth, batchSize, nOutputPlane, outputHeight, outputWidth, kH, kW,
-floor((double)padding/2), dW, 0, scale, partialSum);
return 0;
}
static int nn_(Min_updateOutput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
int dimension = luaT_getfieldcheckint(L, 1, "dimension")-1;
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
THLongStorage *dim;
long i;
luaL_argcheck(L, dimension >= 0 && dimension < input->nDimension, 2, "dimension out of range");
dim = THLongStorage_newWithSize(input->nDimension);
for(i = 0; i < input->nDimension; i++)
dim->data[i] = input->size[i];
dim->data[dimension] = 1;
THTensor_(resize)(output, dim, NULL);
THTensor_(resize)(indices, dim, NULL);
THLongStorage_free(dim);
TH_TENSOR_DIM_APPLY3(real, output, real, input, real, indices, dimension,
long theIndex = 0;
real theMin = input_data[0];
for(i = 1; i < input_size; i++)
{
if(input_data[i*input_stride] < theMin)
{
theIndex = i;
theMin = input_data[i*input_stride];
}
}
*indices_data = theIndex+1;
*output_data = theMin;)
if(output->nDimension > 1)
static int nn_(SpatialUpSampling_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));
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
// dims
int iwidth = input->size[2];
int iheight = input->size[1];
int ichannels = input->size[0];
int owidth = gradOutput->size[2];
int oheight = gradOutput->size[1];
int ochannels = gradOutput->size[0];
// resize gradInput
THTensor_(zero)(gradInput);
// get raw pointers
real *gradInput_data = THTensor_(data)(gradInput);
real *gradOutput_data = THTensor_(data)(gradOutput);
// compute gradients for each plane
int k;
for (k=0; k<ochannels; k++) {
// get planes
real *gradInput_p = gradInput_data + k*iwidth*iheight;
real *gradOutput_p = gradOutput_data + k*owidth*oheight;
// for each plane, resample
int x,y;
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
// input positions (floored)
int ix = x/dW;
int iy = y/dH;
// accumulate gradient
gradInput_p[iy*iwidth + ix] += gradOutput_p[y*owidth + x];
}
}
}
return 1;
}
static int gpunn_SpatialConvolutionGPU_updateOutput(lua_State *L)
{
THGPUTensor *input = (THGPUTensor*)luaT_checkudata(L, 2, "torch.GPUTensor");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int padding = luaT_getfieldcheckint(L, 1, "padding");
THGPUTensor *weight = (THGPUTensor*)luaT_getfieldcheckudata(L, 1, "weight", "torch.GPUTensor");
THGPUTensor *output = (THGPUTensor*)luaT_getfieldcheckudata(L, 1, "output", "torch.GPUTensor");
luaL_argcheck(L, input->nDimension == 4, 2, "4D (batch mode) tensor is expected");
long nOutputPlane = weight->size[3];
long nInputPlane = weight->size[0];
long kH = weight->size[1];
long kW = weight->size[2];
long inputHeight = input->size[1];
long inputWidth = input->size[2];
long batchSize = input->size[3];
long outputHeight = (padding + inputHeight - kH) / dH + 1;
long outputWidth = (padding + inputWidth - kW) / dW + 1;
// resize output
THGPUTensor_resize4d(output, nOutputPlane, outputHeight, outputWidth, batchSize);
// asserts
luaL_argcheck(L, inputWidth == inputHeight, 1, "input must be square");
luaL_argcheck(L, kH == kW, 1, "kH must be equal to kW");
luaL_argcheck(L, dH == dW, 1, "dH must be equal to dW");
// all the data must be contiguous:
luaL_argcheck(L, THGPUTensor_isContiguous(input), 2, "input must be contiguous");
luaL_argcheck(L, THGPUTensor_isContiguous(weight), 1, "weight must be contiguous");
luaL_argcheck(L, THGPUTensor_isContiguous(output), 1, "output must be contiguous");
auto avInput = input->get_array_view();
auto avOutput = output->get_array_view();
auto avWeight = weight->get_array_view();
// convolutions
spatialConv_updateOutput(avInput, avWeight, avOutput, nInputPlane, inputHeight, inputWidth,
batchSize, nOutputPlane, outputHeight, outputWidth, kH, kW,
-floor((double)padding/2), dW, 0, 1, true);
return 1;
}
static int nn_(SpatialUpSampling_forward)(lua_State *L)
{
// get all params
THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id));
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id));
// dims
int iwidth = input->size[2];
int iheight = input->size[1];
int ochannels = input->size[0];
int owidth = iwidth * dW;
int oheight = iheight * dH;
// get raw pointers
real *input_data = THTensor_(data)(input);
real *output_data = THTensor_(data)(output);
// resample each plane
int k;
for (k=0; k<ochannels; k++) {
// get planes
real *input_p = input_data + k*iwidth*iheight;
real *output_p = output_data + k*owidth*oheight;
// for each plane, resample
int x,y;
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
// input positions (floored)
int ix = x/dW;
int iy = y/dH;
// set output
output_p[y*owidth + x] = input_p[iy*iwidth + ix];
}
}
}
return 1;
}
static int nn_(TemporalMaxPooling_updateGradInput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id));
THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id));
int dW = luaT_getfieldcheckint(L, 1, "dW");
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_(Tensor_id));
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_(Tensor_id));
// get contiguous gradOutput
gradOutput = THTensor_(newContiguous)(gradOutput);
// resize and zero
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
// sizes
int noframe = gradOutput->size[0];
long framesize = gradOutput->size[1];
// get raw pointers
real *gradInput_data = THTensor_(data)(gradInput);
real *gradOutput_data = THTensor_(data)(gradOutput);
real *indices_data = THTensor_(data)(indices);
long t, y;
for(t = 0; t < noframe; t++)
{
real *gip = gradInput_data + t*framesize*dW;
real *gop = gradOutput_data + t*framesize;
real *xp = indices_data + t*framesize;
#pragma omp parallel for private(y)
for(y = 0; y < framesize; y++)
{
// compute local max:
long maxindex = (long)xp[y];
gip[maxindex*framesize+y] += gop[y];
}
}
// cleanup
THTensor_(free)(gradOutput);
return 1;
}
static int nn_(SpatialConvolutionMM_accGradParameters)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
real scale = luaL_optnumber(L, 4, 1);
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *finput = luaT_getfieldcheckudata(L, 1, "finput", torch_Tensor);
THTensor *gradWeight = luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor);
THTensor *gradBias = luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor);
THArgCheck( nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1, "Number of output features is not equal to nOutputPlane" );
if(input->nDimension == 3)
{
nn_(SpatialConvolutionMM_accGradParameters_frame)(gradOutput, gradWeight, gradBias, finput, scale);
}
else
{
long T = input->size[0];
long t;
for(t = 0; t < T; t++)
{
THTensor *gradOutput_t = THTensor_(newSelect)(gradOutput, 0, t);
THTensor *finput_t = THTensor_(newSelect)(finput, 0, t);
nn_(SpatialConvolutionMM_accGradParameters_frame)(gradOutput_t, gradWeight, gradBias, finput_t, scale);
THTensor_(free)(gradOutput_t);
THTensor_(free)(finput_t);
}
}
return 0;
}
static int nn_(SpatialGraph_updateOutput)(lua_State *L)
{
// get all params
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
int connex = luaT_getfieldcheckint(L, 1, "connex");
int dist = luaT_getfieldcheckint(L, 1, "dist");
int norm = luaT_getfieldcheckint(L, 1, "normalize");
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
// dims
int iwidth = input->size[2];
int iheight = input->size[1];
int ichannels = input->size[0];
int owidth = iwidth;
int oheight = iheight;
int ochannels = connex / 2;
// norm ?
double normer = (norm == 1) ? 1/sqrt(ichannels) : 1;
// zero output
THTensor_(zero)(output);
// Euclidean distance
if (dist == 0) {
// Sum[ (Xi - Xi+1)^2 ]
int x,y,k;
for (k=0; k<ichannels; k++) {
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
if (x < owidth-1) {
double temp = square(THTensor_(get3d)(input, k, y, x) - THTensor_(get3d)(input, k, y, x+1));
THTensor_(set3d)(output, 0, y, x, temp + THTensor_(get3d)(output, 0, y, x));
}
if (y < oheight-1) {
double temp = square(THTensor_(get3d)(input, k, y, x) - THTensor_(get3d)(input, k, y+1, x));
THTensor_(set3d)(output, 1, y, x, temp + THTensor_(get3d)(output, 1, y, x));
}
}
}
}
// Sqrt[ Sum[ (Xi - Xi+1)^2 ] ]
for (k=0; k<ochannels; k++) {
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
THTensor_(set3d)(output, k, y, x, sqrt(THTensor_(get3d)(output, k, y, x)) * normer);
}
}
}
// Cosine dissimilarity
} else {
// add epsilon to input (to get rid of 0s)
THTensor *inputb = THTensor_(newWithSize3d)(input->size[0], input->size[1], input->size[2]);
THTensor_(copy)(inputb, input);
THTensor_(add)(inputb, inputb, 1e-12);
// Sum[ (Xi * Xi+1) ]
int x,y,k;
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
double norm_A = 0;
double norm_B = 0;
double norm_C = 0;
for (k=0; k<ichannels; k++) {
norm_A += square(THTensor_(get3d)(inputb, k, y, x));
if (x < owidth-1) {
double temp = THTensor_(get3d)(inputb, k, y, x) * THTensor_(get3d)(inputb, k, y, x+1);
THTensor_(set3d)(output, 0, y, x, temp + THTensor_(get3d)(output, 0, y, x));
norm_B += square(THTensor_(get3d)(inputb, k, y, x+1));
}
if (y < oheight-1) {
double temp = THTensor_(get3d)(inputb, k, y, x) * THTensor_(get3d)(inputb, k, y+1, x);
THTensor_(set3d)(output, 1, y, x, temp + THTensor_(get3d)(output, 1, y, x));
norm_C += square(THTensor_(get3d)(inputb, k, y+1, x));
}
}
if (x < owidth-1) {
if (norm) {
THTensor_(set3d)(output, 0, y, x, 1 - THTensor_(get3d)(output, 0, y, x) / (sqrt(norm_A) * sqrt(norm_B)));
} else {
THTensor_(set3d)(output, 0, y, x, ichannels - THTensor_(get3d)(output, 0, y, x));
}
}
if (y < oheight-1) {
if (norm) {
THTensor_(set3d)(output, 1, y, x, 1 - THTensor_(get3d)(output, 1, y, x) / (sqrt(norm_A) * sqrt(norm_C)));
} else {
THTensor_(set3d)(output, 1, y, x, ichannels - THTensor_(get3d)(output, 1, y, x));
}
}
}
}
// Cleanup
THTensor_(free)(inputb);
}
return 1;
}
static int nn_(SpatialGraph_updateGradInput)(lua_State *L)
{
// get all params
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
//int connex = luaT_getfieldcheckint(L, 1, "connex");
int dist = luaT_getfieldcheckint(L, 1, "dist");
int norm = luaT_getfieldcheckint(L, 1, "normalize");
// dims
//int iwidth = input->size[2];
//int iheight = input->size[1];
int ichannels = input->size[0];
int owidth = gradOutput->size[2];
int oheight = gradOutput->size[1];
//int ochannels = gradOutput->size[0];
// norm ?
double normer = (norm == 1) ? 1/sqrt(ichannels)/sqrt(ichannels) : 1;
// resize gradInput
THTensor_(zero)(gradInput);
// compute derivatives, and backpropagate output error to input
if (dist == 0) {
int x,y,k;
for (k=0; k<ichannels; k++) {
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
if (x < owidth-1) {
double partial_d = THTensor_(get3d)(input, k, y, x) - THTensor_(get3d)(input, k, y, x+1);
if (partial_d != 0) partial_d /= THTensor_(get3d)(output, 0, y, x);
partial_d *= THTensor_(get3d)(gradOutput, 0, y, x) * normer;
THTensor_(set3d)(gradInput, k, y, x, partial_d + THTensor_(get3d)(gradInput, k, y, x));
THTensor_(set3d)(gradInput, k, y, x+1, -partial_d + THTensor_(get3d)(gradInput, k, y, x+1));
}
if (y < oheight-1) {
double partial_d = THTensor_(get3d)(input, k, y, x) - THTensor_(get3d)(input, k, y+1, x);
if (partial_d != 0) partial_d /= THTensor_(get3d)(output, 1, y, x);
partial_d *= THTensor_(get3d)(gradOutput, 1, y, x) * normer;
THTensor_(set3d)(gradInput, k, y, x, partial_d + THTensor_(get3d)(gradInput, k, y, x));
THTensor_(set3d)(gradInput, k, y+1, x, -partial_d + THTensor_(get3d)(gradInput, k, y+1, x));
}
}
}
}
// Cosine
} else {
int x,y,k;
for (y=0; y<oheight; y++) {
for (x=0; x<owidth; x++) {
double sum_A = 0;
double sum_B = 0;
double sum_C = 0;
double sum_AB = 0;
double sum_AC = 0;
if (norm) {
for (k=0; k<ichannels; k++) {
sum_A += square(THTensor_(get3d)(input, k, y, x));
if (x < owidth-1) {
sum_B += square(THTensor_(get3d)(input, k, y, x+1));
sum_AB += THTensor_(get3d)(input, k, y, x) * THTensor_(get3d)(input, k, y, x+1);
}
if (y < oheight-1) {
sum_C += square(THTensor_(get3d)(input, k, y+1, x));
sum_AC += THTensor_(get3d)(input, k, y, x) * THTensor_(get3d)(input, k, y+1, x);
}
}
}
double term1, term2, term3, partial_d;
double epsi = 1e-12;
if (x < owidth-1) {
if (norm) {
term1 = 1 / ( pow(sum_A, 1/2) * pow(sum_B, 1/2) + epsi );
term2 = sum_AB / ( pow(sum_A, 3/2) * pow(sum_B, 1/2) + epsi );
term3 = sum_AB / ( pow(sum_B, 3/2) * pow(sum_A, 1/2) + epsi );
}
for (k=0; k<ichannels; k++) {
if (norm) {
partial_d = term2 * THTensor_(get3d)(input, k, y, x)
- term1 * THTensor_(get3d)(input, k, y, x+1);
} else {
partial_d = -THTensor_(get3d)(input, k, y, x+1);
}
partial_d *= THTensor_(get3d)(gradOutput, 0, y, x);
THTensor_(set3d)(gradInput, k, y, x, partial_d + THTensor_(get3d)(gradInput, k, y, x));
if (norm) {
partial_d = term3 * THTensor_(get3d)(input, k, y, x+1)
- term1 * THTensor_(get3d)(input, k, y, x);
} else {
partial_d = -THTensor_(get3d)(input, k, y, x);
}
partial_d *= THTensor_(get3d)(gradOutput, 0, y, x);
THTensor_(set3d)(gradInput, k, y, x+1, partial_d + THTensor_(get3d)(gradInput, k, y, x+1));
}
}
if (y < oheight-1) {
if (norm) {
term1 = 1 / ( pow(sum_A, 1/2) * pow(sum_C, 1/2) + epsi );
term2 = sum_AC / ( pow(sum_A, 3/2) * pow(sum_C, 1/2) + epsi );
term3 = sum_AC / ( pow(sum_C, 3/2) * pow(sum_A, 1/2) + epsi );
}
for (k=0; k<ichannels; k++) {
if (norm) {
partial_d = term2 * THTensor_(get3d)(input, k, y, x)
- term1 * THTensor_(get3d)(input, k, y+1, x);
} else {
partial_d = -THTensor_(get3d)(input, k, y+1, x);
}
partial_d *= THTensor_(get3d)(gradOutput, 1, y, x);
THTensor_(set3d)(gradInput, k, y, x, partial_d + THTensor_(get3d)(gradInput, k, y, x));
if (norm) {
partial_d = term3 * THTensor_(get3d)(input, k, y+1, x)
- term1 * THTensor_(get3d)(input, k, y, x);
} else {
partial_d = -THTensor_(get3d)(input, k, y, x);
}
partial_d *= THTensor_(get3d)(gradOutput, 1, y, x);
THTensor_(set3d)(gradInput, k, y+1, x, partial_d + THTensor_(get3d)(gradInput, k, y+1, x));
}
}
}
}
}
return 1;
}
static int nn_SpatialConvolution_backward(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor_id);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor_id);
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor_id);
THTensor *gradWeight = luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor_id);
THTensor *gradBias = luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor_id);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor_id);
THTensor *gradInputPlane, *unfoldedInputPlane, *unfoldedGradInputPlane, *inputPlane;
THTensor *gradOutputPlane;
THTensor *weightPlane, *gradWeightPlane;
int i, k;
gradInputPlane = THTensor_new();
unfoldedInputPlane = THTensor_new();
unfoldedGradInputPlane = THTensor_new();
inputPlane = THTensor_new();
gradOutputPlane = THTensor_new();
weightPlane = THTensor_new();
gradWeightPlane = THTensor_new();
/* Not necessary with partial backprop: */
THTensor_resizeAs(gradInput, input);
THTensor_zero(gradInput);
for(k = 0; k < nOutputPlane; k++)
{
THTensor_select(gradOutputPlane, gradOutput, 2, k);
THTensor_set1d(gradBias, k, THTensor_get1d(gradBias, k) + THTensor_sum(gradOutputPlane));
for(i = 0; i < nInputPlane; i++)
{
/* ------------------------- gradWeight ------------------------------------- */
/* Get the input image */
THTensor_select(inputPlane, input, 2, i);
THTensor_unfold(unfoldedInputPlane, inputPlane, 0, kW, dW);
THTensor_unfold(unfoldedInputPlane, NULL, 1, kH, dH);
THTensor_transpose(unfoldedInputPlane,NULL,0,2);
THTensor_transpose(unfoldedInputPlane,NULL,1,3);
/* Get the good gradWeight for (k,i) (k out, i in) */
THTensor_select(gradWeightPlane, gradWeight, 3, k);
THTensor_select(gradWeightPlane, NULL, 2, i);
THTensor_addT4dotT2(gradWeightPlane, 1, unfoldedInputPlane, gradOutputPlane);
/* -------------------------- gradInput ------------------------------------- */
/* Not necessary with partial backprop: */
/* Get the gradInput image */
THTensor_select(gradInputPlane, gradInput, 2, i);
THTensor_unfold(unfoldedGradInputPlane, gradInputPlane, 0, kW, dW);
THTensor_unfold(unfoldedGradInputPlane, NULL , 1, kH, dH);
/* Get the good weight for (k,i) (k out, i in) */
THTensor_select(weightPlane, weight, 3, k);
THTensor_select(weightPlane, NULL, 2, i);
THTensor_addT2outT2(unfoldedGradInputPlane, 1, gradOutputPlane, weightPlane);
}
}
THTensor_free(gradInputPlane);
THTensor_free(unfoldedInputPlane);
THTensor_free(unfoldedGradInputPlane);
THTensor_free(inputPlane);
THTensor_free(gradOutputPlane);
THTensor_free(weightPlane);
THTensor_free(gradWeightPlane);
return 1;
}
static int nn_SpatialConvolution_forward(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor_id);
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor_id);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor_id);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor_id);
THTensor *outputPlane, *inputPlane, *weightPlane, *unfoldedInputPlane;
int i, k;
luaL_argcheck(L, input->nDimension == 3, 2, "3D tensor expected");
luaL_argcheck(L, input->size[2] == nInputPlane, 2, "invalid number of input planes");
luaL_argcheck(L, input->size[0] >= kW && input->size[1] >= kH, 2, "input image smaller than kernel size");
THTensor_resize3d(output,
(input->size[0] - kW) / dW + 1,
(input->size[1] - kH) / dH + 1,
nOutputPlane);
inputPlane = THTensor_new();
weightPlane = THTensor_new();
outputPlane = THTensor_new();
unfoldedInputPlane = THTensor_new();
for(k = 0; k < nOutputPlane; k++)
{
THTensor_select(outputPlane, output, 2, k);
/* Initialize to the bias */
THTensor_fill(outputPlane, THTensor_get1d(bias, k));
/* Go! */
for(i = 0; i < nInputPlane; i++)
{
THTensor_select(inputPlane, input, 2, i);
/* Get the good mask for (k,i) (k out, i in) */
THTensor_select(weightPlane, weight, 3, k);
THTensor_select(weightPlane, NULL, 2, i);
/* Get the input image */
THTensor_unfold(unfoldedInputPlane, inputPlane, 0, kW, dW);
THTensor_unfold(unfoldedInputPlane, NULL, 1, kH, dH);
THTensor_addT4dotT2(outputPlane, 1, unfoldedInputPlane, weightPlane);
}
}
THTensor_free(inputPlane);
THTensor_free(weightPlane);
THTensor_free(outputPlane);
THTensor_free(unfoldedInputPlane);
return 1;
}
static int nn_(SpatialFullConvolution_updateOutput)(lua_State *L) {
// Input
THTensor *input = (THTensor*)luaT_checkudata(L, 2, torch_Tensor);
// Params:
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int padW = luaT_getfieldcheckint(L, 1, "padW");
int padH = luaT_getfieldcheckint(L, 1, "padH");
int adjW = luaT_getfieldcheckint(L, 1, "adjW");
int adjH = luaT_getfieldcheckint(L, 1, "adjH");
THTensor *weight = (THTensor*)luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *bias = (THTensor*)luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor);
THTensor *columns = (THTensor*)luaT_getfieldcheckudata(L, 1, "finput", torch_Tensor);
THTensor *ones = (THTensor*)luaT_getfieldcheckudata(L, 1, "fgradInput", torch_Tensor);
THTensor *output = (THTensor*)luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch mode) tensor is expected");
int batch = 1;
if (input->nDimension == 3) {
luaL_argcheck(L, input->size[0] == nInputPlane, 2, "input channels and nInputPlane dont match");
// Force batch
batch = 0;
THTensor_(resize4d)(input, 1, input->size[0], input->size[1], input->size[2]);
} else {
luaL_argcheck(L, input->size[1] == nInputPlane, 2, "input channels and nInputPlane dont match");
}
long inputWidth = input->size[3];
long inputHeight = input->size[2];
long outputWidth = (inputWidth - 1) * dW - 2*padW + kW + adjW;
long outputHeight = (inputHeight - 1) * dH - 2*padH + kH + adjH;
// Batch size + input planes
long batchSize = input->size[0];
// Resize output
THTensor_(resize4d)(output, batchSize, nOutputPlane, outputHeight, outputWidth);
// Resize temporary columns
THTensor_(resize2d)(columns, nOutputPlane*kW*kH, inputHeight*inputWidth);
// Define a buffer of ones, for bias accumulation
// Note: this buffer can be shared with other modules, it only ever gets increased,
// and always contains ones.
if (ones->nDimension != 2 || ones->size[0]*ones->size[1] < outputHeight*outputWidth) {
// Resize plane and fill with ones...
THTensor_(resize2d)(ones, outputHeight, outputWidth);
THTensor_(fill)(ones, 1);
}
// Helpers
THTensor *input_n = THTensor_(new)();
THTensor *output_n = THTensor_(new)();
int elt;
// For each elt in batch, do:
for (elt = 0; elt < batchSize; elt ++) {
// Matrix mulitply per output:
THTensor_(select)(input_n, input, 0, elt);
THTensor_(select)(output_n, output, 0, elt);
// M,N,K are dims of matrix A and B
// (see http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-gemm)
long m = weight->size[1] * weight->size[2] * weight->size[3];
long n = columns->size[1];
long k = weight->size[0];
// Do GEMM (note: this is a bit confusing because gemm assumes column-major matrices)
THBlas_(gemm)(
'n', 't',
n, m, k,
1,
THTensor_(data)(input_n), n,
THTensor_(data)(weight), m,
0,
THTensor_(data)(columns), n
);
// Unpack columns back into input:
nn_(col2im)(
THTensor_(data)(columns),
nOutputPlane, outputHeight, outputWidth, kH, kW, padH, padW, dH, dW,
THTensor_(data)(output_n)
);
// Do Bias after:
// M,N,K are dims of matrix A and B
// (see http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-gemm)
long m_ = nOutputPlane;
long n_ = outputHeight * outputWidth;
long k_ = 1;
// Do GEMM (note: this is a bit confusing because gemm assumes column-major matrices)
THBlas_(gemm)(
't', 'n',
n_, m_, k_,
1,
THTensor_(data)(ones), k_,
THTensor_(data)(bias), k_,
1,
THTensor_(data)(output_n), n_
);
}
// Free
THTensor_(free)(input_n);
THTensor_(free)(output_n);
// Resize output
if (batch == 0) {
THTensor_(resize3d)(output, nOutputPlane, outputHeight, outputWidth);
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
}
// return output
return 1;
}
static int nxn_(Jitter_updateOutput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
int xstart = luaT_getfieldcheckint(L, 1, "xstart");
int ystart = luaT_getfieldcheckint(L, 1, "ystart");
int xcrop = luaT_getfieldcheckint(L, 1, "xcrop");
int ycrop = luaT_getfieldcheckint(L, 1, "ycrop");
int hflip = luaT_getfieldcheckint(L, 1, "randflip");
int bs = input->size[0];
int outy = input->size[1] - ycrop;
int outx = input->size[2] - xcrop;
int channels = input->size[3];
THTensor_(resize4d)(output, bs, outy, outx, channels);
real* idata = THTensor_(data)(input);
real* odata = THTensor_(data)(output);
int istr0 = input->stride[0];
int istr1 = input->stride[1];
int istr2 = input->stride[2];
int istr3 = input->stride[3];
int ostr0 = output->stride[0];
int ostr1 = output->stride[1];
int ostr2 = output->stride[2];
int ostr3 = output->stride[3];
/* This is jittering + hflip */
int batchidx, y, x, ch;
if(hflip==1)
{
#pragma omp parallel for private(batchidx)
for(batchidx=0; batchidx<bs; batchidx++)
{
for (y = 0; y<outy; y++)
{
for(x = 0; x<outx; x++)
{
for (ch = 0; ch < channels; ch++)
{
odata[batchidx*ostr0 + y*ostr1 + x*ostr2 + ch*ostr3] = idata[batchidx*istr0 + (y+ystart-1)*istr1 + (xstart-1+outx-1-x)*istr2 + ch*istr3];
}
}
}
}
}
else
/* This is only jittering */
{
#pragma omp parallel for private(batchidx)
for(batchidx=0; batchidx<bs; batchidx++)
{
for (y = 0; y<outy; y++)
{
for(x = 0; x<outx; x++)
{
for (ch = 0; ch < channels; ch++)
{
odata[batchidx*ostr0 + y*ostr1 + x*ostr2 + ch*ostr3] = idata[batchidx*istr0 + (y+ystart-1)*istr1 + (x+xstart-1)*istr2 + ch*istr3];
}
}
}
}
}
return 1;
}
static int nn_(SpatialConvolution_updateOutput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
int dimw = 2;
int dimh = 1;
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D(batch mode) tensor expected");
if (input->nDimension == 4) {
dimw++;
dimh++;
}
{
long nOutputPlane = weight->size[0];
long kW = weight->size[3];
long kH = weight->size[2];
long inputWidth = input->size[dimw];
long inputHeight = input->size[dimh];
long outputWidth = (inputWidth - kW) / dW + 1;
long outputHeight = (inputHeight - kH) / dH + 1;
if (input->nDimension == 3)
{
long i;
real* bias_data;
real* output_data;
THTensor_(resize3d)(output, nOutputPlane, outputHeight, outputWidth);
/* add bias */
bias_data = THTensor_(data)(bias);
output_data = THTensor_(data)(output);
#pragma omp parallel for private(i)
for (i=0; i<bias->size[0]; i++)
{
/*THTensor_(select)(outn,output,0,i);*/
/*TH_TENSOR_APPLY(real,outn, *outn_data = bias_data[i];);*/
real *ptr_output = output_data + i*outputWidth*outputHeight;
long j;
for(j = 0; j < outputWidth*outputHeight; j++)
ptr_output[j] = bias_data[i];
}
/*THTensor_(free)(outn);*/
/* do convolutions */
THTensor_(conv2Dmv)(output, 1.0, 1.0, input, weight, dH, dW, "V","X");
}
else
{
real* bias_data;
real* output_data;
long p;
THTensor_(resize4d)(output, input->size[0], nOutputPlane, outputHeight, outputWidth);
bias_data = THTensor_(data)(bias);
output_data = THTensor_(data)(output);
#pragma omp parallel for private(p)
for (p=0; p<input->size[0]; p++)
{
/* BIAS */
long i;
for (i=0; i<bias->size[0]; i++)
{
real *ptr_output = output_data + p*nOutputPlane*outputWidth*outputHeight + i*outputWidth*outputHeight;
long j;
for(j = 0; j < outputWidth*outputHeight; j++)
ptr_output[j] = bias_data[i];
}
}
/* do convolutions */
THTensor_(conv2Dmm)(output, 1.0, 1.0, input, weight, dH, dW, "V","X");
}
}
return 1;
}
static int nn_LcEncoder_forward(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor_id);
int winX = luaT_getfieldcheckint(L, 1, "winX");
int winY = luaT_getfieldcheckint(L, 1, "winY");
int woutX = luaT_getfieldcheckint(L, 1, "woutX");
int woutY = luaT_getfieldcheckint(L, 1, "woutY");
double xStep = luaT_getfieldchecknumber(L, 1, "xStep");
double yStep = luaT_getfieldchecknumber(L, 1, "yStep");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor_id);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor_id);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor_id);
luaL_argcheck(L, input->nDimension == 3, 2, "3D tensor expected");
luaL_argcheck(L, input->size[2] == 1, 2, "invalid input 3rd dim size has to be 1");
THTensor *inputPlane, *inputNarrowedX, *inputNarrowedYX;
THTensor *weightSelectedX, *weightSelectedYX;
inputPlane = THTensor_new();
inputNarrowedX = THTensor_new();
inputNarrowedYX = THTensor_new();
weightSelectedX = THTensor_new();
weightSelectedYX = THTensor_new();
// get output size from input
THTensor_resize3d(output,
(input->size[0] - winX+1) / xStep,
(input->size[1] - winY+1) / yStep,
1);
THTensor_select(inputPlane, input, 2, 0);
int y,x,iy,ix,wy,wx;
for (y = 0; y<output->size[1]; y++)
{
iy = (int)floor(y*yStep);
wy = y%woutY;
for (x = 0; x<output->size[0]; x++)
{
ix = (int)floor(x*xStep);
wx = x%woutX;
THTensor_narrow(inputNarrowedX, inputPlane, 0, ix, winX);
THTensor_narrow(inputNarrowedYX, inputNarrowedX, 1, iy, winY);
THTensor_select(weightSelectedX, weight, 3, wy);
THTensor_select(weightSelectedYX, weightSelectedX, 2, wx);
double dot = THTensor_dot(inputNarrowedYX, weightSelectedYX);
double biasSelect = THTensor_get2d(bias,wx,wy);
THTensor_set3d(output,x,y,0,dot+biasSelect);
}
}
THTensor_free(inputPlane);
THTensor_free(inputNarrowedX);
THTensor_free(inputNarrowedYX);
THTensor_free(weightSelectedX);
THTensor_free(weightSelectedYX);
return 1;
}
static int nn_LcEncoder_backward(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor_id);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor_id);
int winX = luaT_getfieldcheckint(L, 1, "winX");
int winY = luaT_getfieldcheckint(L, 1, "winY");
int woutX = luaT_getfieldcheckint(L, 1, "woutX");
int woutY = luaT_getfieldcheckint(L, 1, "woutY");
double xStep = luaT_getfieldchecknumber(L, 1, "xStep");
double yStep = luaT_getfieldchecknumber(L, 1, "yStep");
luaL_argcheck(L, input->nDimension == 3, 2, "input 3D tensor expected");
luaL_argcheck(L, input->size[2] == 1, 2, "invalid input 3rd dim size has to be 1");
luaL_argcheck(L, gradOutput->nDimension == 3, 3, "gradOutput 3D tensor expected");
luaL_argcheck(L, gradOutput->size[2] == 1, 3, "invalid gradOutput 3rd dim size has to be 1");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor_id);
THTensor *gradWeight = luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor_id);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor_id);
THTensor *gradBias = luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor_id);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor_id);
/* ----------------------- gradWeight ----------------------- */
THTensor_fill(gradWeight, 0);
THTensor *inputPlane, *inputNarrowedX, *inputNarrowedYX;
inputPlane = THTensor_new();
inputNarrowedX = THTensor_new();
inputNarrowedYX = THTensor_new();
THTensor_select(inputPlane, input, 2, 0);
THTensor *gradWeightSelectedX, *gradWeightSelectedYX;
gradWeightSelectedX = THTensor_new();
gradWeightSelectedYX = THTensor_new();
/* ----------------------- gradInput ------------------------ */
THTensor_resizeAs(gradInput, input);
THTensor_fill(gradInput, 0);
THTensor *gradInputPlane, *gradInputNarrowedX, *gradInputNarrowedYX;
gradInputPlane = THTensor_new();
gradInputNarrowedX = THTensor_new();
gradInputNarrowedYX = THTensor_new();
THTensor_select(gradInputPlane, gradInput, 2, 0);
THTensor *weightSelectedX, *weightSelectedYX;
weightSelectedX = THTensor_new();
weightSelectedYX = THTensor_new();
int y,x,iy,ix,wy,wx;
for (y = 0; y<gradOutput->size[1]; y++)
{
iy = (int)floor(y*yStep);
wy = y%woutY;
for (x = 0; x<gradOutput->size[0]; x++)
{
ix = (int)floor(x*xStep);
wx = x%woutX;
double gradOutVal = THTensor_get3d(gradOutput,x,y,0);
/* ----------------------- gradWeight ----------------------- */
THTensor_narrow(inputNarrowedX, inputPlane, 0, ix, winX);
THTensor_narrow(inputNarrowedYX, inputNarrowedX, 1, iy, winY);
THTensor_select(gradWeightSelectedX, gradWeight, 3, wy);
THTensor_select(gradWeightSelectedYX, gradWeightSelectedX, 2, wx);
THTensor_addTensor(gradWeightSelectedYX, gradOutVal, inputNarrowedYX);
/* ----------------------- gradBias ----------------------- */
THTensor_set2d(gradBias,wx,wy, THTensor_get2d(gradBias,wx,wy) + gradOutVal);
/* ----------------------- gradInput ------------------------ */
THTensor_narrow(gradInputNarrowedX, gradInputPlane, 0, ix, winX);
THTensor_narrow(gradInputNarrowedYX, gradInputNarrowedX, 1, iy, winY);
THTensor_select(weightSelectedX, weight, 3, wy);
THTensor_select(weightSelectedYX, weightSelectedX, 2, wx);
THTensor_addTensor(gradInputNarrowedYX, gradOutVal, weightSelectedYX);
}
}
/* free gradWeight */
THTensor_free(inputPlane);
THTensor_free(inputNarrowedX);
THTensor_free(inputNarrowedYX);
THTensor_free(gradWeightSelectedX);
THTensor_free(gradWeightSelectedYX);
/* free gradInput */
THTensor_free(gradInputPlane);
THTensor_free(gradInputNarrowedX);
THTensor_free(gradInputNarrowedYX);
THTensor_free(weightSelectedX);
THTensor_free(weightSelectedYX);
return 1;
}
static int nn_(SpatialFullConvolution_updateGradInput)(lua_State *L) {
// Inputs
THTensor *input = (THTensor *)luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = (THTensor *)luaT_checkudata(L, 3, torch_Tensor);
// Params
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int padW = luaT_getfieldcheckint(L, 1, "padW");
int padH = luaT_getfieldcheckint(L, 1, "padH");
int adjW = luaT_getfieldcheckint(L, 1, "adjW");
int adjH = luaT_getfieldcheckint(L, 1, "adjH");
THTensor *weight = (THTensor *)luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *gradColumns = (THTensor*)luaT_getfieldcheckudata(L, 1, "finput", torch_Tensor);
THTensor *gradInput = (THTensor *)luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch mode) tensor is expected");
int batch = 1;
if (input->nDimension == 3) {
// Force batch
batch = 0;
THTensor_(resize4d)(input, 1, input->size[0], input->size[1], input->size[2]);
THTensor_(resize4d)(gradOutput, 1, gradOutput->size[0], gradOutput->size[1], gradOutput->size[2]);
}
long inputWidth = input->size[3];
long inputHeight = input->size[2];
long outputWidth = (inputWidth - 1) * dW - 2*padW + kW + adjW;
long outputHeight = (inputHeight - 1) * dH - 2*padH + kH + adjH;
// Batch size + input planes
long batchSize = input->size[0];
// Resize output
THTensor_(resize4d)(gradInput, batchSize, nInputPlane, inputHeight, inputWidth);
// Resize temporary columns
THTensor_(resize2d)(gradColumns, nOutputPlane*kW*kH, inputHeight*inputWidth);
// Helpers
THTensor *gradInput_n = THTensor_(new)();
THTensor *gradOutput_n = THTensor_(new)();
int elt;
// For each elt in batch, do:
for (elt = 0; elt < batchSize; elt ++) {
// Matrix mulitply per sample:
THTensor_(select)(gradInput_n, gradInput, 0, elt);
THTensor_(select)(gradOutput_n, gradOutput, 0, elt);
// Extract columns:
nn_(im2col)(
THTensor_(data)(gradOutput_n),
nOutputPlane, outputHeight, outputWidth, kH, kW, padH, padW, dH, dW,
THTensor_(data)(gradColumns)
);
// M,N,K are dims of matrix A and B
// (see http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-gemm)
long m = weight->size[0];
long n = gradColumns->size[1];
long k = weight->size[1] * weight->size[2] * weight->size[3];
// Do GEMM (note: this is a bit confusing because gemm assumes column-major matrices)
THBlas_(gemm)(
'n', 'n',
n, m, k,
1,
THTensor_(data)(gradColumns), n,
THTensor_(data)(weight), k,
0,
THTensor_(data)(gradInput_n), n
);
}
// Free
THTensor_(free)(gradInput_n);
THTensor_(free)(gradOutput_n);
// Resize output
if (batch == 0) {
THTensor_(resize3d)(gradOutput, nOutputPlane, outputHeight, outputWidth);
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
THTensor_(resize3d)(gradInput, nInputPlane, inputHeight, inputWidth);
}
// Return gradInput
return 1;
}
static int nnconv1d_(HorizontalConvolution_updateOutput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int kL = luaT_getfieldcheckint(L, 1, "kL");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *bias = luaT_getfieldcheckudata(L, 1, "bias", torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
luaL_argcheck(L, input->nDimension == 3 ||
input->nDimension == 4, 2, "3D or 4D (batch mode) tensor expected");
// change to batch mode
int batch = 1;
if (input->nDimension == 3) {
batch = 0;
THTensor_(resize4d)(input, 1, nInputPlane, input->size[1], input->size[2]);
}
long batchSize = input->size[0];
long inputHeight = input->size[2];
long inputWidth = input->size[3];
long outputHeight = inputHeight;
long outputWidth = inputWidth - kL + 1;
THTensor_(resize4d)(output, batchSize, nOutputPlane, outputHeight, outputWidth);
int elt;
#pragma omp parallel for private(elt)
for (elt = 0; elt < batchSize; elt++) {
// select each batch
THTensor *input_t = THTensor_(newSelect)(input, 0, elt);
THTensor *output_t = THTensor_(newSelect)(output, 0, elt);
// fill biases
int i, j, k;
for (i = 0; i < nOutputPlane; i++) {
THVector_(fill)(output_t->storage->data+output_t->storageOffset+output_t->stride[0]*i,
THTensor_(get1d)(bias, i), outputHeight*outputWidth);
}
// convolve horizontally
for (i = 0; i < nInputPlane; i++) {
for (j = 0; j < inputHeight; j++) {
for (k = 0; k < kL; k++) {
THVector_(add)(output_t->storage->data + output_t->storageOffset +
output_t->stride[0]*i + output_t->stride[1]*j,
input_t->storage->data + input_t->storageOffset +
input_t->stride[0]*i + input_t->stride[1]*j + k,
*(THTensor_(data)(weight)+i*kL+k), outputWidth);
}
}
}
// release temp tensors
THTensor_(free)(input_t);
THTensor_(free)(output_t);
}
// revert to single batch
if (batch == 0) {
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
THTensor_(resize3d)(output, nOutputPlane, outputHeight, outputWidth);
}
return 1;
}
static int nnconv1d_(HorizontalConvolution_accGradParameters)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
real scale = luaL_optnumber(L, 4, 1);
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int kL = luaT_getfieldcheckint(L, 1, "kL");
THTensor *ones = luaT_getfieldcheckudata(L, 1, "ones", torch_Tensor);
THTensor *gradWeight = luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor);
THTensor *gradBias = luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor);
THArgCheck(nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1,
"Number of output features is not equal to nOutputPlane" );
// change to batch mode
int batch = 1;
if (input->nDimension == 3) {
batch = 0;
THTensor_(resize4d)(input, 1, input->size[0], input->size[1], input->size[2]);
THTensor_(resize4d)(gradOutput, 1, gradOutput->size[0], gradOutput->size[1], gradOutput->size[2]);
}
long batchSize = input->size[0];
long inputHeight = input->size[2];
long inputWidth = input->size[3];
long outputHeight = inputHeight;
long outputWidth = inputWidth - kL + 1;
if (ones->nDimension != 1 || ones->size[0] < outputHeight*outputWidth) {
THTensor_(resize1d)(ones, outputHeight*outputWidth);
THTensor_(fill)(ones, 1);
}
int elt;
for (elt = 0; elt < batchSize; elt++) {
// select each batch in 2D
THTensor *input_t = THTensor_(newSelect)(input, 0, elt);
THTensor *gradOutput_t = THTensor_(newSelect)(gradOutput, 0, elt);
THTensor *gradOutput2d = THTensor_(newWithStorage2d)(gradOutput->storage, gradOutput->storageOffset,
nOutputPlane, -1, outputWidth*outputHeight, -1);
// dot products
int i, j, k;
for (i = 0; i < nInputPlane; i++) {
for (k = 0; k < kL; k++) {
for (j = 0; j < outputHeight; j++) {
*(gradWeight->storage->data + gradWeight->storageOffset + i*gradWeight->stride[0] + k) +=
scale*THBlas_(dot)
(outputWidth,
gradOutput_t->storage->data + gradOutput_t->storageOffset +
i*gradOutput_t->stride[0] + j*gradOutput_t->stride[1],
gradOutput_t->stride[2],
input_t->storage->data + input_t->storageOffset +
i*input_t->stride[0] + j*input_t->stride[1] + k,
input_t->stride[2]);
}
}
}
// fill biases
THTensor_(addmv)(gradBias, 1, gradBias, scale, gradOutput2d, ones);
THTensor_(free)(gradOutput2d);
THTensor_(free)(input_t);
THTensor_(free)(gradOutput_t);
}
// revert to single batch
if (batch == 0) {
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
THTensor_(resize3d)(gradOutput, nOutputPlane, outputHeight, outputWidth);
}
return 0;
}
static int nn_(SpatialAdaptiveMaxPooling_updateOutput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
long oheight = luaT_getfieldcheckint(L, 1, "H");
long owidth = luaT_getfieldcheckint(L, 1, "W");
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_Tensor);
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
int dimw = 2;
int dimh = 1;
long nbatch = 1;
long nslices;
long iheight;
long iwidth;
long istride_d;
long istride_h;
long istride_w;
long istride_b;
real *input_data;
real *output_data;
real *indices_data;
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4 , 2, "3D or 4D (batch mode) tensor expected");
if (input->nDimension == 4)
{
istride_b = input->stride[0];
nbatch = input->size[0];
dimw++;
dimh++;
}
/* sizes */
nslices = input->size[dimh-1];
iheight = input->size[dimh];
iwidth = input->size[dimw];
/* strides */
istride_d = input->stride[dimh-1];
istride_h = input->stride[dimh];
istride_w = input->stride[dimw];
/* resize output */
if (input->nDimension == 3)
{
THTensor_(resize3d)(output, nslices, oheight, owidth);
/* indices will contain i,j locations for each output point */
THTensor_(resize4d)(indices, 2, nslices, oheight, owidth);
input_data = THTensor_(data)(input);
output_data = THTensor_(data)(output);
indices_data = THTensor_(data)(indices);
nn_(SpatialAdaptiveMaxPooling_updateOutput_frame)(input_data, output_data,
indices_data+nslices*owidth*oheight, indices_data,
nslices,
iwidth, iheight,
owidth, oheight,
istride_w,istride_h,
istride_d);
}
else
{
long p;
THTensor_(resize4d)(output, nbatch, nslices, oheight, owidth);
/* indices will contain i,j locations for each output point */
THTensor_(resize5d)(indices, 2, nbatch, nslices, oheight, owidth);
input_data = THTensor_(data)(input);
output_data = THTensor_(data)(output);
indices_data = THTensor_(data)(indices);
#pragma omp parallel for private(p)
for (p = 0; p < nbatch; p++)
{
nn_(SpatialAdaptiveMaxPooling_updateOutput_frame)(input_data+p*istride_b, output_data+p*nslices*owidth*oheight,
indices_data+(p+nbatch)*nslices*owidth*oheight, indices_data+p*nslices*owidth*oheight,
nslices,
iwidth, iheight,
owidth, oheight,
istride_w,istride_h,
istride_d);
}
}
return 1;
}
static int nn_(VolumetricAveragePooling_updateGradInput)(lua_State *L) {
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
int dT = luaT_getfieldcheckint(L, 1, "dT");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int kT = luaT_getfieldcheckint(L, 1, "kT");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput",
torch_Tensor);
int nslices;
int itime;
int iheight;
int iwidth;
int otime;
int oheight;
int owidth;
real *gradInput_data;
real *gradOutput_data;
int dimN = 0;
int dimt = 1;
int dimh = 2;
int dimw = 3;
/* get contiguous gradOutput */
gradOutput = THTensor_(newContiguous)(gradOutput);
/* resize */
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
if (input->nDimension == 5) {
dimN++;
dimt++;
dimh++;
dimw++;
}
/* sizes */
nslices = input->size[dimN];
itime = input->size[dimt];
iheight = input->size[dimh];
iwidth = input->size[dimw];
otime = gradOutput->size[dimt];
oheight = gradOutput->size[dimh];
owidth = gradOutput->size[dimw];
/* get raw pointers */
gradInput_data = THTensor_(data)(gradInput);
gradOutput_data = THTensor_(data)(gradOutput);
/* backprop */
if (input->nDimension == 4) { /* non-batch mode*/
nn_(VolumetricAveragePooling_updateGradInput_frame)(
gradInput_data, gradOutput_data, nslices,
itime, iwidth, iheight, otime, owidth, oheight,
kT, kW, kH, dT, dW, dH);
} else { /* batch mode */
long p;
long nBatch = input->size[0];
long istride = nslices * itime * iwidth * iheight;
long ostride = nslices * otime * owidth * oheight;
#pragma omp parallel for private(p)
for (p = 0; p < nBatch; p++) {
nn_(VolumetricAveragePooling_updateGradInput_frame)(
gradInput_data + p * istride, gradOutput_data + p * ostride, nslices,
itime, iwidth, iheight, otime, owidth, oheight,
kT, kW, kH, dT, dW, dH);
}
}
/* cleanup */
THTensor_(free)(gradOutput);
return 1;
}
static int nn_(SpatialConvolution_accGradParameters)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
real scale = luaL_optnumber(L, 4, 1);
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
THTensor *gradWeight = luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor);
THTensor *gradBias = luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor);
int dimw = 2;
int dimh = 1;
real *gradBias_data;
real *gradOutput_data;
long noutSlice;
THArgCheck( nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1, "Number of output features is not equal to nOutputPlane" );
if (input->nDimension == 4)
{
dimw++;
dimh++;
}
/* gradient to bias */
gradBias_data = THTensor_(data)(gradBias);
gradOutput_data = THTensor_(data)(gradOutput);
noutSlice = gradOutput->size[dimh]*gradOutput->size[dimw];
/*THTensor* gradOutSlice = THTensor_(new)();*/
if (input->nDimension == 3)
{
long k;
#pragma omp parallel for private(k)
for(k = 0; k < nOutputPlane; k++)
{
/*THTensor_(select)(gradOutSlice, gradOutput, 0, k);*/
real *ptr_gradOutput = gradOutput_data + k*noutSlice;
long l;
for(l = 0; l < noutSlice; l++)
gradBias_data[k] += scale*ptr_gradOutput[l];
}
/* gradient to kernels */
THTensor_(conv2DRevger)(gradWeight, 1.0, scale, input, gradOutput, dH, dW);
}
else
{
long k;
#pragma omp parallel for private(k)
for(k = 0; k < nOutputPlane; k++)
{
long p;
for(p = 0; p < input->size[0]; p++)
{
/* BIAS */
real *ptr_gradOutput = gradOutput_data + p*nOutputPlane*noutSlice + k*noutSlice;
long l;
for(l = 0; l < noutSlice; l++)
gradBias_data[k] += scale*ptr_gradOutput[l];
}
}
/* gradient to kernels */
THTensor_(conv2DRevgerm)(gradWeight, 1.0, scale, input, gradOutput, dH, dW);
}
return 0;
}
static int nn_(VolumetricAveragePooling_updateOutput)(lua_State *L) {
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
int kT = luaT_getfieldcheckint(L, 1, "kT");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int dT = luaT_getfieldcheckint(L, 1, "dT");
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_Tensor);
long nslices;
long itime;
long iheight;
long iwidth;
long otime;
long oheight;
long owidth;
real *input_data;
real *output_data;
luaL_argcheck(L, input->nDimension == 4 || input->nDimension == 5, 2,
"4D or 5D (batch-mode) tensor expected");
int dimN = 0;
int dimt = 1;
int dimh = 2;
int dimw = 3;
if (input->nDimension == 5) {
dimN++;
dimt++;
dimh++;
dimw++;
}
luaL_argcheck(L, input->size[dimw] >= kW && input->size[dimh] >= kH &&
input->size[dimt] >= kT, 2,
"input image smaller than kernel size");
/* sizes */
nslices = input->size[dimN];
itime = input->size[dimt];
iheight = input->size[dimh];
iwidth = input->size[dimw];
otime = (itime - kT) / dT + 1;
oheight = (iheight - kH) / dH + 1;
owidth = (iwidth - kW) / dW + 1;
/* get contiguous input */
input = THTensor_(newContiguous)(input);
if (input->nDimension == 4) { /* non-batch mode */
/* resize output */
THTensor_(resize4d)(output, nslices, otime, oheight, owidth);
input_data = THTensor_(data)(input);
output_data = THTensor_(data)(output);
nn_(VolumetricAveragePooling_updateOutput_frame)(input_data, output_data,
nslices,
itime, iwidth, iheight,
otime, owidth, oheight,
kT, kW, kH, dT, dW, dH);
} else { /* batch mode */
long p;
long nBatch = input->size[0];
long istride = nslices * itime * iwidth * iheight;
long ostride = nslices * otime * owidth * oheight;
/* resize output */
THTensor_(resize5d)(output, nBatch, nslices, otime, oheight, owidth);
input_data = THTensor_(data)(input);
output_data = THTensor_(data)(output);
#pragma omp parallel for private(p)
for (p=0; p < nBatch; p++) {
nn_(VolumetricAveragePooling_updateOutput_frame)(
input_data + p * istride, output_data + p * ostride,
nslices, itime, iwidth, iheight, otime, owidth, oheight,
kT, kW, kH, dT, dW, dH);
}
}
/* cleanup */
THTensor_(free)(input);
return 1;
}
static int nnconv1d_(HorizontalConvolution_updateGradInput)(lua_State *L)
{
THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = luaT_checkudata(L, 3, torch_Tensor);
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int kL = luaT_getfieldcheckint(L, 1, "kL");
THTensor *weight = luaT_getfieldcheckudata(L, 1, "weight", torch_Tensor);
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
THArgCheck(nOutputPlane == gradOutput->size[input->nDimension == 4 ? 1 : 0], 1,
"Number of output features is not equal to nOutputPlane" );
// change to batch mode
int batch = 1;
if (input->nDimension == 3) {
batch = 0;
THTensor_(resize4d)(input, 1, input->size[0], input->size[1], input->size[2]);
THTensor_(resize4d)(gradOutput, 1, nOutputPlane, gradOutput->size[1], gradOutput->size[2]);
}
long batchSize = input->size[0];
long inputHeight = input->size[2];
long inputWidth = input->size[3];
long outputHeight = inputHeight;
long outputWidth = inputWidth - kL + 1;
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
int elt;
#pragma omp parallel for private(elt)
for (elt = 0; elt < batchSize; elt++) {
// select each batch
THTensor *gradInput_t = THTensor_(newSelect)(gradInput, 0, elt);
THTensor *gradOutput_t = THTensor_(newSelect)(gradOutput, 0, elt);
// convolve horizontally
int i, j, k;
for (i = 0; i < nOutputPlane; i++) {
for (j = 0; j < outputHeight; j++) {
for (k = 0; k < kL; k++) {
THVector_(add)(gradInput_t->storage->data + gradInput_t->storageOffset +
gradInput_t->stride[0]*i + gradInput_t->stride[1]*j + k,
gradOutput_t->storage->data + gradOutput_t->storageOffset +
gradOutput_t->stride[0]*i + gradOutput_t->stride[1]*j,
*(THTensor_(data)(weight)+i*kL+k), outputWidth); // needs to change
}
}
}
// release temp tensors
THTensor_(free)(gradInput_t);
THTensor_(free)(gradOutput_t);
}
// revert to single batch
if (batch == 0) {
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
THTensor_(resize3d)(gradInput, nInputPlane, inputHeight, inputWidth);
THTensor_(resize3d)(gradOutput, nOutputPlane, outputHeight, outputWidth);
}
return 1;
}
static int nn_(SpatialFullConvolution_accGradParameters)(lua_State *L) {
// Inputs
THTensor *input = (THTensor *)luaT_checkudata(L, 2, torch_Tensor);
THTensor *gradOutput = (THTensor *)luaT_checkudata(L, 3, torch_Tensor);
// Params
int dW = luaT_getfieldcheckint(L, 1, "dW");
int dH = luaT_getfieldcheckint(L, 1, "dH");
int kW = luaT_getfieldcheckint(L, 1, "kW");
int kH = luaT_getfieldcheckint(L, 1, "kH");
int nInputPlane = luaT_getfieldcheckint(L, 1, "nInputPlane");
int nOutputPlane = luaT_getfieldcheckint(L, 1, "nOutputPlane");
int padW = luaT_getfieldcheckint(L, 1, "padW");
int padH = luaT_getfieldcheckint(L, 1, "padH");
int adjW = luaT_getfieldcheckint(L, 1, "adjW");
int adjH = luaT_getfieldcheckint(L, 1, "adjH");
float scale = luaL_optnumber(L, 4, 1);
THTensor *gradWeight = (THTensor *)luaT_getfieldcheckudata(L, 1, "gradWeight", torch_Tensor);
THTensor *gradBias = (THTensor *)luaT_getfieldcheckudata(L, 1, "gradBias", torch_Tensor);
THTensor *columns = (THTensor*)luaT_getfieldcheckudata(L, 1, "finput", torch_Tensor);
THTensor *ones = (THTensor*)luaT_getfieldcheckudata(L, 1, "fgradInput", torch_Tensor);
luaL_argcheck(L, input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch mode) tensor is expected");
int batch = 1;
if (input->nDimension == 3) {
// Force batch
batch = 0;
THTensor_(resize4d)(input, 1, input->size[0], input->size[1], input->size[2]);
THTensor_(resize4d)(gradOutput, 1, gradOutput->size[0], gradOutput->size[1], gradOutput->size[2]);
}
long inputWidth = input->size[3];
long inputHeight = input->size[2];
long outputWidth = (inputWidth - 1) * dW - 2*padW + kW + adjW;
long outputHeight = (inputHeight - 1) * dH - 2*padH + kH + adjH;
// Batch size + input planes
long batchSize = input->size[0];
// Define a buffer of ones, for bias accumulation
if (ones->nDimension != 2 || ones->size[0]*ones->size[1] < outputHeight*outputWidth) {
// Resize plane and fill with ones...
THTensor_(resize2d)(ones, outputHeight, outputWidth);
THTensor_(fill)(ones, 1);
}
// Resize temporary columns
THTensor_(resize2d)(columns, nOutputPlane*kW*kH, inputHeight*inputWidth);
// Helpers
THTensor *input_n = THTensor_(new)();
THTensor *gradOutput_n = THTensor_(new)();
int elt;
// For each elt in batch, do:
for (elt = 0; elt < batchSize; elt ++) {
// Matrix mulitply per output:
THTensor_(select)(input_n, input, 0, elt);
THTensor_(select)(gradOutput_n, gradOutput, 0, elt);
// Extract columns:
nn_(im2col)(
THTensor_(data)(gradOutput_n),
nOutputPlane, outputHeight, outputWidth, kH, kW, padH, padW, dH, dW,
THTensor_(data)(columns)
);
// M,N,K are dims of matrix A and B
// (see http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-gemm)
long n = columns->size[0]; // nOutputPlane * kh * kw
long m = input_n->size[0]; // nInputPlane
long k = columns->size[1]; // inputHeight * inputWidth
// Do GEMM (note: this is a bit confusing because gemm assumes column-major matrices)
THBlas_(gemm)(
't', 'n',
n, m, k,
scale,
THTensor_(data)(columns), k,
THTensor_(data)(input_n), k,
1,
THTensor_(data)(gradWeight), n
);
// Do Bias:
// M,N,K are dims of matrix A and B
// (see http://docs.nvidia.com/cuda/cublas/#cublas-lt-t-gt-gemm)
long m_ = nOutputPlane;
long k_ = outputHeight * outputWidth;
// Do GEMV (note: this is a bit confusing because gemv assumes column-major matrices)
THBlas_(gemv)(
't',
k_, m_,
scale,
THTensor_(data)(gradOutput_n), k_,
THTensor_(data)(ones), 1,
1,
THTensor_(data)(gradBias), 1
);
}
// Free
THTensor_(free)(input_n);
THTensor_(free)(gradOutput_n);
// Resize
if (batch == 0) {
THTensor_(resize3d)(gradOutput, nOutputPlane, outputHeight, outputWidth);
THTensor_(resize3d)(input, nInputPlane, inputHeight, inputWidth);
}
// Return nothing
return 0;
}
