diff options
| -rw-r--r-- | SpatialDivisiveNormalization.lua | 134 | ||||
| -rw-r--r-- | init.lua | 1 | ||||
| -rw-r--r-- | test/test.lua | 32 |
3 files changed, 167 insertions, 0 deletions
diff --git a/SpatialDivisiveNormalization.lua b/SpatialDivisiveNormalization.lua new file mode 100644 index 0000000..eaf7465 --- /dev/null +++ b/SpatialDivisiveNormalization.lua @@ -0,0 +1,134 @@ +local SpatialDivisiveNormalization, parent = torch.class('nn.SpatialDivisiveNormalization','nn.Module') + +function SpatialDivisiveNormalization:__init(nInputPlane, kernel, threshold, thresval) + parent.__init(self) + + -- get args + self.nInputPlane = nInputPlane or 1 + self.kernel = kernel or torch.Tensor(9,9):fill(1) + self.threshold = threshold or 1e-4 + self.thresval = thresval or 1e-4 + local kdim = self.kernel:nDimension() + + -- check args + if kdim ~= 2 and kdim ~= 1 then + error('<SpatialDivisiveNormalization> averaging kernel must be 2D or 1D') + end + if (self.kernel:size(1) % 2) == 0 or (kdim == 2 and (self.kernel:size(2) % 2) == 0) then + error('<SpatialDivisiveNormalization> averaging kernel must have ODD dimensions') + end + + -- normalize kernel + self.kernel:div(self.kernel:sum() * self.nInputPlane) + + -- padding values + local padH = math.floor(self.kernel:size(1)/2) + local padW = padH + if kdim == 2 then + padW = math.floor(self.kernel:size(2)/2) + end + + -- create convolutional mean estimator + self.meanestimator = nn.Sequential() + self.meanestimator:add(nn.SpatialZeroPadding(padW, padW, padH, padH)) + if kdim == 2 then + self.meanestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + self.kernel:size(2), self.kernel:size(1))) + else + self.meanestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + self.kernel:size(1), 1)) + self.meanestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + 1, self.kernel:size(1))) + end + self.meanestimator:add(nn.Sum(1)) + self.meanestimator:add(nn.Replicate(self.nInputPlane)) + + -- create convolutional std estimator + self.stdestimator = nn.Sequential() + self.stdestimator:add(nn.Square()) + self.stdestimator:add(nn.SpatialZeroPadding(padW, padW, padH, padH)) + if kdim == 2 then + self.stdestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + self.kernel:size(2), self.kernel:size(1))) + else + self.stdestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + self.kernel:size(1), 1)) + self.stdestimator:add(nn.SpatialConvolutionMap(nn.tables.oneToOne(self.nInputPlane), + 1, self.kernel:size(1))) + end + self.stdestimator:add(nn.Sum(1)) + self.stdestimator:add(nn.Replicate(self.nInputPlane)) + self.stdestimator:add(nn.Sqrt()) + + -- set kernel and bias + if kdim == 2 then + for i = 1,self.nInputPlane do + self.meanestimator.modules[2].weight[i] = self.kernel + self.stdestimator.modules[3].weight[i] = self.kernel + end + self.meanestimator.modules[2].bias:zero() + self.stdestimator.modules[3].bias:zero() + else + for i = 1,self.nInputPlane do + self.meanestimator.modules[2].weight[i]:copy(self.kernel) + self.meanestimator.modules[3].weight[i]:copy(self.kernel) + self.stdestimator.modules[3].weight[i]:copy(self.kernel) + self.stdestimator.modules[4].weight[i]:copy(self.kernel) + end + self.meanestimator.modules[2].bias:zero() + self.meanestimator.modules[3].bias:zero() + self.stdestimator.modules[3].bias:zero() + self.stdestimator.modules[4].bias:zero() + end + + -- other operation + self.normalizer = nn.CDivTable() + self.divider = nn.CDivTable() + self.thresholder = nn.Threshold(self.threshold, self.thresval) + + -- coefficient array, to adjust side effects + self.coef = torch.Tensor(1,1,1) +end + +function SpatialDivisiveNormalization:updateOutput(input) + -- compute side coefficients + if (input:size(3) ~= self.coef:size(3)) or (input:size(2) ~= self.coef:size(2)) then + local ones = input.new():resizeAs(input):fill(1) + self.coef = self.meanestimator:updateOutput(ones) + self.coef = self.coef:clone() + end + + -- normalize std dev + self.localstds = self.stdestimator:updateOutput(input) + self.adjustedstds = self.divider:updateOutput{self.localstds, self.coef} + self.thresholdedstds = self.thresholder:updateOutput(self.adjustedstds) + self.output = self.normalizer:updateOutput{input, self.thresholdedstds} + + -- done + return self.output +end + +function SpatialDivisiveNormalization:updateGradInput(input, gradOutput) + -- resize grad + self.gradInput:resizeAs(input):zero() + + -- backprop through all modules + local gradnorm = self.normalizer:updateGradInput({input, self.thresholdedstds}, gradOutput) + local gradadj = self.thresholder:updateGradInput(self.adjustedstds, gradnorm[2]) + local graddiv = self.divider:updateGradInput({self.localstds, self.coef}, gradadj) + self.gradInput:add(self.stdestimator:updateGradInput(input, graddiv[1])) + self.gradInput:add(gradnorm[1]) + + -- done + return self.gradInput +end + +function SpatialDivisiveNormalization:type(type) + parent.type(self,type) + self.meanestimator:type(type) + self.stdestimator:type(type) + self.divider:type(type) + self.normalizer:type(type) + self.thresholder:type(type) + return self +end @@ -62,6 +62,7 @@ torch.include('nn', 'SpatialLPPooling.lua') torch.include('nn', 'TemporalConvolution.lua') torch.include('nn', 'TemporalSubSampling.lua') torch.include('nn', 'SpatialSubtractiveNormalization.lua') +torch.include('nn', 'SpatialDivisiveNormalization.lua') torch.include('nn', 'SpatialZeroPadding.lua') torch.include('nn', 'VolumetricConvolution.lua') diff --git a/test/test.lua b/test/test.lua index d67165f..31a6049 100644 --- a/test/test.lua +++ b/test/test.lua @@ -556,6 +556,38 @@ function nntest.SpatialSubtractiveNormalization_1dkernel() mytester:asserteq(berr, 0, torch.typename(module) .. ' - i/o backward err ') end +function nntest.SpatialDivisiveNormalization_2dkernel() + local inputSize = math.random(11,20) + local kersize = 9 + local nbfeatures = math.random(5,10) + local kernel = torch.Tensor(kersize,kersize):fill(1) + local module = nn.SpatialDivisiveNormalization(nbfeatures,kernel) + local input = torch.rand(nbfeatures,inputSize,inputSize) + + local err = jac.testJacobian(module,input) + mytester:assertlt(err,precision, 'error on state ') + + local ferr,berr = jac.testIO(module,input) + mytester:asserteq(ferr, 0, torch.typename(module) .. ' - i/o forward err ') + mytester:asserteq(berr, 0, torch.typename(module) .. ' - i/o backward err ') +end + +function nntest.SpatialDivisiveNormalization_1dkernel() + local inputSize = math.random(11,20) + local kersize = 9 + local nbfeatures = math.random(5,10) + local kernel = torch.Tensor(kersize):fill(1) + local module = nn.SpatialDivisiveNormalization(nbfeatures,kernel) + local input = torch.rand(nbfeatures,inputSize,inputSize) + + local err = jac.testJacobian(module,input) + mytester:assertlt(err,precision, 'error on state ') + + local ferr,berr = jac.testIO(module,input) + mytester:asserteq(ferr, 0, torch.typename(module) .. ' - i/o forward err ') + mytester:asserteq(berr, 0, torch.typename(module) .. ' - i/o backward err ') +end + function nntest.SpatialConvolution() local from = math.random(1,10) local to = math.random(1,10) |