def _createWeightsTensorAndConvolve(self, rng, filterShape, convWInitMethod,
inputToConvTrain, inputToConvVal, inputToConvTest,
inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest) :
# Behaviour: Create W, set self._W, set self.params, convolve, return ouput and outputShape.
# The created filters are either 1-dimensional (rank=1) or 2-dim (rank=2), depending on the self._rank
# If 1-dim: rSubconv is the input convolved with the row-1dimensional filter.
# If 2-dim: rSubconv is the input convolved with the RC-2D filter, cSubconv with CZ-2D filter, zSubconv with ZR-2D filter.
#----- Initialise the weights and Convolve for 3 separate, low rank filters, R,C,Z. -----
# W shape: [#FMs of this layer, #FMs of Input, rKernDim, cKernDim, zKernDim]
rSubconvFilterShape = [ filterShape[0]//3, filterShape[1], filterShape[2], 1 if self._rank == 1 else filterShape[3], 1 ]
rSubconvW = createAndInitializeWeightsTensor(rSubconvFilterShape, convWInitMethod, rng)
rSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(rSubconvW, rSubconvFilterShape, inputToConvTrain, inputToConvVal, inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest)
cSubconvFilterShape = [ filterShape[0]//3, filterShape[1], 1, filterShape[3], 1 if self._rank == 1 else filterShape[4] ]
cSubconvW = createAndInitializeWeightsTensor(cSubconvFilterShape, convWInitMethod, rng)
cSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(cSubconvW, cSubconvFilterShape, inputToConvTrain, inputToConvVal, inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest)
numberOfFmsForTotalToBeExact = filterShape[0] - 2*(filterShape[0]//3) # Cause of possibly inexact integer division.
zSubconvFilterShape = [ numberOfFmsForTotalToBeExact, filterShape[1], 1 if self._rank == 1 else filterShape[2], 1, filterShape[4] ]
zSubconvW = createAndInitializeWeightsTensor(zSubconvFilterShape, convWInitMethod, rng)
zSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(zSubconvW, zSubconvFilterShape, inputToConvTrain, inputToConvVal, inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest)
# Set the W attribute and trainable parameters.
self._WperSubconv = [rSubconvW, cSubconvW, zSubconvW] # Bear in mind that these sub tensors have different shapes! Treat carefully.
self.params = self._WperSubconv + self.params
# concatenate together.
(concatSubconvOutputsTrain, concatOutputShapeTrain) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(rSubconvTupleWithOuputAndShapeTrValTest[0], rSubconvTupleWithOuputAndShapeTrValTest[3],
cSubconvTupleWithOuputAndShapeTrValTest[0], cSubconvTupleWithOuputAndShapeTrValTest[3],
zSubconvTupleWithOuputAndShapeTrValTest[0], zSubconvTupleWithOuputAndShapeTrValTest[3],
filterShape)
(concatSubconvOutputsVal, concatOutputShapeVal) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(rSubconvTupleWithOuputAndShapeTrValTest[1], rSubconvTupleWithOuputAndShapeTrValTest[4],
cSubconvTupleWithOuputAndShapeTrValTest[1], cSubconvTupleWithOuputAndShapeTrValTest[4],
zSubconvTupleWithOuputAndShapeTrValTest[1], zSubconvTupleWithOuputAndShapeTrValTest[4],
filterShape)
(concatSubconvOutputsTest, concatOutputShapeTest) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(rSubconvTupleWithOuputAndShapeTrValTest[2], rSubconvTupleWithOuputAndShapeTrValTest[5],
cSubconvTupleWithOuputAndShapeTrValTest[2], cSubconvTupleWithOuputAndShapeTrValTest[5],
zSubconvTupleWithOuputAndShapeTrValTest[2], zSubconvTupleWithOuputAndShapeTrValTest[5],
filterShape)
return (concatSubconvOutputsTrain, concatSubconvOutputsVal, concatSubconvOutputsTest, concatOutputShapeTrain, concatOutputShapeVal, concatOutputShapeTest)
def _createWeightsTensorAndConvolve(self, rng, filterShape, convWInitMethod,
inputToConvTrain, inputToConvVal, inputToConvTest,
inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest) :
#-----------------------------------------------
#------------------ Convolution ----------------
#-----------------------------------------------
#----- Initialise the weights -----
# W shape: [#FMs of this layer, #FMs of Input, rKernDim, cKernDim, zKernDim]
self._W = createAndInitializeWeightsTensor(filterShape, convWInitMethod, rng)
self.params = [self._W] + self.params
#---------- Convolve --------------
tupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(self._W, filterShape, inputToConvTrain, inputToConvVal, inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal, inputToConvShapeTest)
return tupleWithOuputAndShapeTrValTest
def _createWeightsTensorAndConvolve(self, rng, filterShape,
convWInitMethod, inputToConvTrain,
inputToConvVal, inputToConvTest,
inputToConvShapeTrain,
inputToConvShapeVal,
inputToConvShapeTest):
#-----------------------------------------------
#------------------ Convolution ----------------
#-----------------------------------------------
#----- Initialise the weights -----
# W shape: [#FMs of this layer, #FMs of Input, rKernDim, cKernDim, zKernDim]
self._W = createAndInitializeWeightsTensor(filterShape,
convWInitMethod, rng)
self.params = [self._W] + self.params
#---------- Convolve --------------
tupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(
self._W, filterShape, inputToConvTrain, inputToConvVal,
inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal,
inputToConvShapeTest)
return tupleWithOuputAndShapeTrValTest
def _createWeightsTensorAndConvolve(self, rng, filterShape,
convWInitMethod, inputToConvTrain,
inputToConvVal, inputToConvTest,
inputToConvShapeTrain,
inputToConvShapeVal,
inputToConvShapeTest):
# Behaviour: Create W, set self._W, set self.params, convolve, return ouput and outputShape.
# The created filters are either 1-dimensional (rank=1) or 2-dim (rank=2), depending on the self._rank
# If 1-dim: rSubconv is the input convolved with the row-1dimensional filter.
# If 2-dim: rSubconv is the input convolved with the RC-2D filter, cSubconv with CZ-2D filter, zSubconv with ZR-2D filter.
#----- Initialise the weights and Convolve for 3 separate, low rank filters, R,C,Z. -----
# W shape: [#FMs of this layer, #FMs of Input, rKernDim, cKernDim, zKernDim]
rSubconvFilterShape = [
filterShape[0] // 3, filterShape[1], filterShape[2],
1 if self._rank == 1 else filterShape[3], 1
]
rSubconvW = createAndInitializeWeightsTensor(rSubconvFilterShape,
convWInitMethod, rng)
rSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(
rSubconvW, rSubconvFilterShape, inputToConvTrain, inputToConvVal,
inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal,
inputToConvShapeTest)
cSubconvFilterShape = [
filterShape[0] // 3, filterShape[1], 1, filterShape[3],
1 if self._rank == 1 else filterShape[4]
]
cSubconvW = createAndInitializeWeightsTensor(cSubconvFilterShape,
convWInitMethod, rng)
cSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(
cSubconvW, cSubconvFilterShape, inputToConvTrain, inputToConvVal,
inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal,
inputToConvShapeTest)
numberOfFmsForTotalToBeExact = filterShape[0] - 2 * (
filterShape[0] // 3) # Cause of possibly inexact integer division.
zSubconvFilterShape = [
numberOfFmsForTotalToBeExact, filterShape[1],
1 if self._rank == 1 else filterShape[2], 1, filterShape[4]
]
zSubconvW = createAndInitializeWeightsTensor(zSubconvFilterShape,
convWInitMethod, rng)
zSubconvTupleWithOuputAndShapeTrValTest = convolveWithGivenWeightMatrix(
zSubconvW, zSubconvFilterShape, inputToConvTrain, inputToConvVal,
inputToConvTest, inputToConvShapeTrain, inputToConvShapeVal,
inputToConvShapeTest)
# Set the W attribute and trainable parameters.
self._WperSubconv = [
rSubconvW, cSubconvW, zSubconvW
] # Bear in mind that these sub tensors have different shapes! Treat carefully.
self.params = self._WperSubconv + self.params
# concatenate together.
(concatSubconvOutputsTrain, concatOutputShapeTrain
) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(
rSubconvTupleWithOuputAndShapeTrValTest[0],
rSubconvTupleWithOuputAndShapeTrValTest[3],
cSubconvTupleWithOuputAndShapeTrValTest[0],
cSubconvTupleWithOuputAndShapeTrValTest[3],
zSubconvTupleWithOuputAndShapeTrValTest[0],
zSubconvTupleWithOuputAndShapeTrValTest[3], filterShape)
(concatSubconvOutputsVal, concatOutputShapeVal
) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(
rSubconvTupleWithOuputAndShapeTrValTest[1],
rSubconvTupleWithOuputAndShapeTrValTest[4],
cSubconvTupleWithOuputAndShapeTrValTest[1],
cSubconvTupleWithOuputAndShapeTrValTest[4],
zSubconvTupleWithOuputAndShapeTrValTest[1],
zSubconvTupleWithOuputAndShapeTrValTest[4], filterShape)
(concatSubconvOutputsTest, concatOutputShapeTest
) = self._cropSubconvOutputsToSameDimsAndConcatenateFms(
rSubconvTupleWithOuputAndShapeTrValTest[2],
rSubconvTupleWithOuputAndShapeTrValTest[5],
cSubconvTupleWithOuputAndShapeTrValTest[2],
cSubconvTupleWithOuputAndShapeTrValTest[5],
zSubconvTupleWithOuputAndShapeTrValTest[2],
zSubconvTupleWithOuputAndShapeTrValTest[5], filterShape)
return (concatSubconvOutputsTrain, concatSubconvOutputsVal,
concatSubconvOutputsTest, concatOutputShapeTrain,
concatOutputShapeVal, concatOutputShapeTest)