def testNxNred(self, datatype, ncols, npoints, mode):
rtol, atol = 1e-5, 1e-8
# Configuration
nAngles = int(npoints * (npoints - 1) / 2)
# Expected values
expctdLeftTop = torch.tensor(1., dtype=datatype)
# Instantiation of target class
target = OrthonormalTransform(n=npoints, mode=mode)
target.angles = nn.init.uniform_(target.angles, a=0.0, b=2 * math.pi)
target.angles.data[:npoints - 1] = torch.zeros(npoints - 1)
# Actual values
with torch.no_grad():
matrix = target.forward(torch.eye(npoints, dtype=datatype))
actualLeftTop = matrix[0, 0] #.numpy()
# Evaluation
message = "actualLeftTop=%s differs from %s" % (str(actualLeftTop),
str(expctdLeftTop))
#self.assertTrue(np.isclose(actualLeftTop,expctdLeftTop,rtol=rtol,atol=atol),message)
self.assertTrue(
torch.isclose(actualLeftTop, expctdLeftTop, rtol=rtol, atol=atol),
message)
def testCallWithAngles(self, datatype, ncols, mode):
rtol, atol = 1e-4, 1e-7
# Expected values
X = torch.randn(2, ncols, dtype=datatype)
R = torch.tensor([[math.cos(math.pi / 4), -math.sin(math.pi / 4)],
[math.sin(math.pi / 4),
math.cos(math.pi / 4)]],
dtype=datatype)
if mode != 'Synthesis':
expctdZ = R @ X
else:
expctdZ = R.T @ X
# Instantiation of target class
target = OrthonormalTransform(mode=mode)
#target.angles.data = torch.tensor([math.pi/4])
target.angles = nn.init.constant_(target.angles, val=math.pi / 4)
# Actual values
with torch.no_grad():
actualZ = target.forward(X)
# Evaluation
self.assertTrue(torch.allclose(actualZ, expctdZ, rtol=rtol, atol=atol))
def testBackwardMultiColumns(self, datatype, ncols, mode):
rtol, atol = 1e-4, 1e-7
# Configuration
nPoints = 2
# Expected values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=True)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
R = torch.eye(nPoints, dtype=datatype)
dRdW = torch.tensor([[0., -1.], [1., 0.]], dtype=datatype)
if mode != 'Synthesis':
expctddLdX = R.T @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW @ X))
else:
expctddLdX = R @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW.T @ X))
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
# Actual values
torch.autograd.set_detect_anomaly(True)
Z = target.forward(X)
target.zero_grad()
Z.backward(dLdZ)
actualdLdX = X.grad
actualdLdW = target.angles.grad
# Evaluation
self.assertTrue(
torch.allclose(actualdLdX, expctddLdX, rtol=rtol, atol=atol))
self.assertTrue(
torch.allclose(actualdLdW, expctddLdW, rtol=rtol, atol=atol))
def __init__(self,
number_of_channels=[],
decimation_factor=[],
no_dc_leakage=False,
name=''):
super(NsoltFinalRotation2dLayer, self).__init__()
self.name = name
self.number_of_channels = number_of_channels
self.decimation_factor = decimation_factor
self.description = "NSOLT final rotation " \
+ "(ps,pa) = (" \
+ str(self.number_of_channels[0]) + "," \
+ str(self.number_of_channels[1]) + "), " \
+ "(mv,mh) = (" \
+ str(self.decimation_factor[Direction.VERTICAL]) + "," \
+ str(self.decimation_factor[Direction.HORIZONTAL]) + ")"
# Instantiation of orthormal transforms
ps, pa = self.number_of_channels
self.orthTransW0T = OrthonormalTransform(n=ps, mode='Synthesis')
self.orthTransW0T.angles = nn.init.zeros_(self.orthTransW0T.angles)
self.orthTransU0T = OrthonormalTransform(n=pa, mode='Synthesis')
self.orthTransU0T.angles = nn.init.zeros_(self.orthTransU0T.angles)
# No DC leakage
self.no_dc_leakage = no_dc_leakage
def testBackward8x8RandAngPdAng4(self, mode, ncols):
datatype = torch.double
rtol, atol = 1e-4, 1e-7
# Configuration
#mode = 'Synthesis'
nPoints = 8
#ncols = 2
angs0 = 2 * math.pi * torch.rand(28, dtype=datatype)
angs1 = angs0.clone()
angs2 = angs0.clone()
pdAng = 4
delta = 1e-4
angs1[pdAng] = angs0[pdAng] - delta / 2.
angs2[pdAng] = angs0[pdAng] + delta / 2.
# Expcted values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=True)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
omgs = OrthonormalMatrixGenerationSystem(dtype=datatype,
partial_difference=False)
R = omgs(angles=angs0, mus=1)
dRdW = (omgs(angles=angs2, mus=1) - omgs(angles=angs1, mus=1)) / delta
if mode != 'Synthesis':
expctddLdX = R.T @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW @ X))
else:
expctddLdX = R @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW.T @ X))
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target.angles.data = angs0
# Actual values
torch.autograd.set_detect_anomaly(True)
Z = target.forward(X)
target.zero_grad()
Z.backward(dLdZ)
actualdLdX = X.grad
actualdLdW = target.angles.grad[pdAng]
# Evaluation
self.assertTrue(
torch.allclose(actualdLdX, expctddLdX, rtol=rtol, atol=atol))
self.assertTrue(
torch.allclose(actualdLdW, expctddLdW, rtol=rtol, atol=atol))
def __init__(self,
number_of_channels=[],
mode='Synthesis',
name=''):
super(NsoltIntermediateRotation2dLayer, self).__init__()
self.name = name
self.number_of_channels = number_of_channels
self.description = mode \
+ " NSOLT intermediate rotation " \
+ "(ps,pa) = (" \
+ str(self.number_of_channels[0]) + "," \
+ str(self.number_of_channels[1]) + ")"
# Instantiation of orthormal transforms
ps,pa = self.number_of_channels
self.orthTransUn = OrthonormalTransform(n=pa,mode=mode)
self.orthTransUn.angles = nn.init.zeros_(self.orthTransUn.angles)
class NsoltIntermediateRotation2dLayer(nn.Module):
"""
NSOLTINTERMEDIATEROTATION2DLAYER
コンポーネント別に入力(nComponents):
nSamples x nRows x nCols x nChs
コンポーネント別に出力(nComponents):
nSamples x nRows x nCols x nChs
Requirements: Python 3.7.x, PyTorch 1.7.x
Copyright (c) 2020-2021, Shogo MURAMATSU
All rights reserved.
Contact address: Shogo MURAMATSU,
Faculty of Engineering, Niigata University,
8050 2-no-cho Ikarashi, Nishi-ku,
Niigata, 950-2181, JAPAN
http://msiplab.eng.niigata-u.ac.jp/
"""
def __init__(self,
number_of_channels=[],
mode='Synthesis',
name=''):
super(NsoltIntermediateRotation2dLayer, self).__init__()
self.name = name
self.number_of_channels = number_of_channels
self.description = mode \
+ " NSOLT intermediate rotation " \
+ "(ps,pa) = (" \
+ str(self.number_of_channels[0]) + "," \
+ str(self.number_of_channels[1]) + ")"
# Instantiation of orthormal transforms
ps,pa = self.number_of_channels
self.orthTransUn = OrthonormalTransform(n=pa,mode=mode)
self.orthTransUn.angles = nn.init.zeros_(self.orthTransUn.angles)
def forward(self,X):
nSamples = X.size(dim=0)
nrows = X.size(dim=1)
ncols = X.size(dim=2)
ps,pa = self.number_of_channels
# Process
Z = X.clone()
Ya = X[:,:,:,ps:].view(-1,pa).T
Za = self.orthTransUn.forward(Ya)
Z[:,:,:,ps:] = Za.T.view(nSamples,nrows,ncols,pa)
return Z
@property
def mode(self):
return self.orthTransUn.mode
def test8x8(self, datatype, ncols, mode):
rtol, atol = 1e-5, 1e-8
# Expected values
expctdNorm = torch.tensor(1., dtype=datatype)
# Instantiation of target class
target = OrthonormalTransform(n=8, mode=mode)
target.angles.data = torch.randn(28, dtype=datatype)
# Actual values
unitvec = torch.randn(8, ncols, dtype=datatype)
unitvec /= unitvec.norm()
with torch.no_grad():
actualNorm = target.forward(unitvec).norm() #.numpy()
# Evaluation
message = "actualNorm=%s differs from %s" % (str(actualNorm),
str(expctdNorm))
#self.assertTrue(np.isclose(actualNorm,expctdNorm,rtol=rtol,atol=atol),message)
self.assertTrue(
torch.isclose(actualNorm, expctdNorm, rtol=rtol, atol=atol),
message)
def testConstructor(self, datatype, ncols):
rtol, atol = 1e-5, 1e-8
# Expected values
X = torch.randn(2, ncols, dtype=datatype)
expctdZ = X
expctdNParams = 1
expctdMode = 'Analysis'
# Instantiation of target class
target = OrthonormalTransform()
# Actual values
with torch.no_grad():
actualZ = target.forward(X)
actualNParams = len(target.parameters().__next__())
actualMode = target.mode
# Evaluation
self.assertTrue(isinstance(target, nn.Module))
self.assertTrue(torch.allclose(actualZ, expctdZ, rtol=rtol, atol=atol))
self.assertEqual(actualNParams, expctdNParams)
self.assertEqual(actualMode, expctdMode)
def testBackwardAngsAndMus(self, datatype, mode, ncols):
rtol, atol = 1e-4, 1e-7
# Configuration
#mode = 'Analysis'
nPoints = 2
#ncols = 1
mus = [1, -1]
# Expected values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=True)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
# angle = 2.*math.pi*randn(1)
angle = 2. * math.pi * gauss(mu=0., sigma=1.) #randn(1)
R = torch.tensor([[math.cos(angle), -math.sin(angle)],
[-math.sin(angle), -math.cos(angle)]],
dtype=datatype) #.squeeze()
dRdW = torch.tensor(
[[-math.sin(angle), -math.cos(angle)],
[-math.cos(angle), math.sin(angle)]],
dtype=datatype) #.squeeze()
if mode != 'Synthesis':
expctddLdX = R.T @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW @ X))
else:
expctddLdX = R @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW.T @ X))
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target.angles = nn.init.constant_(target.angles, val=angle)
target.mus = torch.tensor(mus, dtype=datatype)
# Actual values
torch.autograd.set_detect_anomaly(True)
Z = target.forward(X)
target.zero_grad()
Z.backward(dLdZ)
actualdLdX = X.grad
actualdLdW = target.angles.grad
# Evaluation
self.assertTrue(
torch.allclose(actualdLdX, expctddLdX, rtol=rtol, atol=atol))
self.assertTrue(
torch.allclose(actualdLdW, expctddLdW, rtol=rtol, atol=atol))
def testGradCheckNxNRandAngMus(self, mode, ncols, npoints):
# Configuration
datatype = torch.double
nPoints = npoints
nAngs = int(nPoints * (nPoints - 1) / 2.)
mus = (-1)**torch.randint(high=2, size=(nPoints, ))
angs = 2. * math.pi * torch.randn(nAngs, dtype=datatype)
# Expcted values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=True)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target.angles.data = angs
target.mus = mus
torch.autograd.set_detect_anomaly(True)
Z = target.forward(X)
target.zero_grad()
# Evaluation
self.assertTrue(torch.autograd.gradcheck(target, (X, )))
def testInstantiationWithInvalidMode(self):
mode = 'Invalid'
# Instantiation of target class
with self.assertRaises(InvalidMode):
target = OrthonormalTransform(mode=mode)
def testBackword8x8RandAngMusPdAng7(self, mode, ncols):
datatype = torch.double
rtol, atol = 1e-4, 1e-7
# Configuration
#mode = 'Synthesis'
nPoints = 8
#ncols = 2
mus = [1, -1, 1, -1, 1, -1, 1, -1]
angs0 = 2 * math.pi * torch.rand(28, dtype=datatype)
angs1 = angs0.clone()
angs2 = angs0.clone()
pdAng = 7
delta = 1e-4
angs1[pdAng] = angs0[pdAng] - delta / 2.
angs2[pdAng] = angs0[pdAng] + delta / 2.
# Expcted values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=False)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
# Instantiation of target class
target0 = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target0.angles.data = angs0
target0.mus = mus
target1 = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target1.angles.data = angs1
target1.mus = mus
target2 = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target2.angles.data = angs2
target2.mus = mus
# Expctd values
if mode == 'Analysis':
bwmode = 'Synthesis'
else:
bwmode = 'Analysis'
backprop = OrthonormalTransform(n=nPoints, dtype=datatype, mode=bwmode)
backprop.angles.data = angs0
backprop.mus = mus
torch.autograd.set_detect_anomaly(True)
dZdW = (target2.forward(X) - target1.forward(X)) / delta # ~ d(R*X)/dW
expctddLdW = torch.sum(dLdZ * dZdW) # ~ dLdW
# Actual values
X.detach()
X.requires_grad = True
Z = target0.forward(X)
target0.zero_grad()
#print(torch.autograd.gradcheck(target0,(X,angs0)))
Z.backward(dLdZ)
actualdLdW = target0.angles.grad[pdAng]
# Evaluation
self.assertTrue(
torch.allclose(actualdLdW, expctddLdW, rtol=rtol, atol=atol))
def testBackward4x4RandAngPdAng1(self, mode, ncols):
datatype = torch.double
rtol, atol = 1e-2, 1e-5
# Configuration
#mode = 'Synthesis'
nPoints = 4
#ncols = 2
mus = [-1, -1, -1, -1]
angs = 2. * math.pi * torch.randn(6, dtype=datatype)
pdAng = 1
delta = 1e-4
# Expcted values
X = torch.randn(nPoints, ncols, dtype=datatype, requires_grad=True)
dLdZ = torch.randn(nPoints, ncols, dtype=datatype)
R = torch.as_tensor(
torch.tensor(mus).view(-1,1) * \
torch.tensor(
[ [1, 0, 0, 0. ],
[0, 1, 0, 0. ],
[0, 0, math.cos(angs[5]), -math.sin(angs[5]) ],
[0, 0, math.sin(angs[5]), math.cos(angs[5]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[4]), 0, -math.sin(angs[4]) ],
[0, 0, 1, 0 ],
[0, math.sin(angs[4]), 0, math.cos(angs[4]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[3]), -math.sin(angs[3]), 0 ],
[0, math.sin(angs[3]), math.cos(angs[3]), 0 ],
[0, 0, 0, 1 ] ]
) @ torch.tensor(
[ [ math.cos(angs[2]), 0, 0, -math.sin(angs[2]) ],
[0, 1, 0, 0 ],
[0, 0, 1, 0 ],
[ math.sin(angs[2]), 0, 0, math.cos(angs[2]) ] ]
) @ torch.tensor(
[ [math.cos(angs[1]), 0, -math.sin(angs[1]), 0 ],
[0, 1, 0, 0 ],
[math.sin(angs[1]), 0, math.cos(angs[1]), 0 ],
[0, 0, 0, 1 ] ]
) @ torch.tensor(
[ [ math.cos(angs[0]), -math.sin(angs[0]), 0, 0 ],
[ math.sin(angs[0]), math.cos(angs[0]), 0, 0 ],
[ 0, 0, 1, 0 ],
[ 0, 0, 0, 1 ] ]
),dtype=datatype)
dRdW = torch.as_tensor(
(1./delta) * torch.tensor(mus).view(-1,1) * \
torch.tensor(
[ [1, 0, 0, 0. ],
[0, 1, 0., 0. ],
[0, 0, math.cos(angs[5]), -math.sin(angs[5]) ],
[0., 0, math.sin(angs[5]), math.cos(angs[5]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[4]), 0, -math.sin(angs[4]) ],
[0, 0, 1, 0 ],
[0, math.sin(angs[4]), 0, math.cos(angs[4]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[3]), -math.sin(angs[3]), 0 ],
[0, math.sin(angs[3]), math.cos(angs[3]), 0 ],
[0, 0, 0, 1 ] ]
) @ torch.tensor(
[ [ math.cos(angs[2]), 0, 0, -math.sin(angs[2]) ],
[0, 1, 0, 0 ],
[0, 0, 1, 0 ],
[ math.sin(angs[2]), 0, 0, math.cos(angs[2]) ] ]
) @ (
torch.tensor(
[ [math.cos(angs[1]+delta/2.), 0, -math.sin(angs[1]+delta/2.), 0 ],
[0, 1, 0, 0 ],
[math.sin(angs[1]+delta/2.), 0, math.cos(angs[1]+delta/2.), 0 ],
[0, 0, 0, 1 ] ] ) - \
torch.tensor(
[ [math.cos(angs[1]-delta/2.), 0, -math.sin(angs[1]-delta/2.), 0 ],
[0, 1, 0, 0 ],
[math.sin(angs[1]-delta/2.), 0, math.cos(angs[1]-delta/2.), 0 ],
[0, 0, 0, 1 ] ] )
) @ torch.tensor(
[ [ math.cos(angs[0]), -math.sin(angs[0]), 0, 0 ],
[ math.sin(angs[0]), math.cos(angs[0]), 0, 0 ],
[ 0, 0, 1, 0 ],
[ 0, 0, 0, 1 ] ]
),dtype=datatype)
if mode != 'Synthesis':
expctddLdX = R.T @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW @ X))
else:
expctddLdX = R @ dLdZ # = dZdX @ dLdZ
expctddLdW = torch.sum(dLdZ * (dRdW.T @ X))
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target.angles.data = angs
target.mus = mus
# Actual values
torch.autograd.set_detect_anomaly(True)
Z = target.forward(X)
target.zero_grad()
Z.backward(dLdZ)
actualdLdX = X.grad
actualdLdW = target.angles.grad[pdAng]
# Evaluation
self.assertTrue(
torch.allclose(actualdLdX, expctddLdX, rtol=rtol, atol=atol))
self.assertTrue(
torch.allclose(actualdLdW, expctddLdW, rtol=rtol, atol=atol))
def testForward4x4RandAngs(self, datatype, mode, ncols):
rtol, atol = 1e-4, 1e-7
# Configuration
#mode = 'Synthesis'
nPoints = 4
#ncols = 2
mus = [-1, 1, -1, 1]
angs = 2. * math.pi * torch.randn(6, dtype=datatype)
# Expcted values
X = torch.randn(nPoints, ncols, dtype=datatype)
R = torch.as_tensor(
torch.tensor(mus).view(-1,1) * \
torch.tensor(
[ [1, 0, 0, 0. ],
[0, 1, 0, 0. ],
[0, 0, math.cos(angs[5]), -math.sin(angs[5]) ],
[0, 0, math.sin(angs[5]), math.cos(angs[5]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[4]), 0, -math.sin(angs[4]) ],
[0, 0, 1, 0 ],
[0, math.sin(angs[4]), 0, math.cos(angs[4]) ] ]
) @ torch.tensor(
[ [1, 0, 0, 0 ],
[0, math.cos(angs[3]), -math.sin(angs[3]), 0 ],
[0, math.sin(angs[3]), math.cos(angs[3]), 0 ],
[0, 0, 0, 1 ] ]
) @ torch.tensor(
[ [ math.cos(angs[2]), 0, 0, -math.sin(angs[2]) ],
[0, 1, 0, 0 ],
[0, 0, 1, 0 ],
[ math.sin(angs[2]), 0, 0, math.cos(angs[2]) ] ]
) @ torch.tensor(
[ [math.cos(angs[1]), 0, -math.sin(angs[1]), 0 ],
[0, 1, 0, 0 ],
[math.sin(angs[1]), 0, math.cos(angs[1]), 0 ],
[0, 0, 0, 1 ] ]
) @ torch.tensor(
[ [ math.cos(angs[0]), -math.sin(angs[0]), 0, 0 ],
[ math.sin(angs[0]), math.cos(angs[0]), 0, 0 ],
[ 0, 0, 1, 0 ],
[ 0, 0, 0, 1 ] ]
),dtype=datatype)
if mode != 'Synthesis':
expctdZ = R @ X
else:
expctdZ = R.T @ X
# Instantiation of target class
target = OrthonormalTransform(n=nPoints, dtype=datatype, mode=mode)
target.angles.data = angs
target.mus = mus
# Actual values
with torch.no_grad():
actualZ = target.forward(X)
# Evaluation
self.assertTrue(torch.allclose(actualZ, expctdZ, rtol=rtol, atol=atol))
def testSetInvalidMus(self):
mus = 2
with self.assertRaises(InvalidMus):
target = OrthonormalTransform()
target.mus = mus
def testInstantiationWithInvalidMus(self):
mus = 2
with self.assertRaises(InvalidMus):
target = OrthonormalTransform(mus=mus)
def testSetInvalidMode(self):
mode = 'Invalid'
with self.assertRaises(InvalidMode):
target = OrthonormalTransform()
target.mode = 'InvalidMode'
class NsoltFinalRotation2dLayer(nn.Module):
"""
NSOLTFINALROTATION2DLAYER
コンポーネント別に入力(nComponents):
nSamples x nRows x nCols x nChs
コンポーネント別に出力(nComponents):
nSamples x nRows x nCols x nDecs
Requirements: Python 3.7.x, PyTorch 1.7.x
Copyright (c) 2020-2021, Shogo MURAMATSU
All rights reserved.
Contact address: Shogo MURAMATSU,
Faculty of Engineering, Niigata University,
8050 2-no-cho Ikarashi, Nishi-ku,
Niigata, 950-2181, JAPAN
http://msiplab.eng.niigata-u.ac.jp/
"""
def __init__(self,
number_of_channels=[],
decimation_factor=[],
no_dc_leakage=False,
name=''):
super(NsoltFinalRotation2dLayer, self).__init__()
self.name = name
self.number_of_channels = number_of_channels
self.decimation_factor = decimation_factor
self.description = "NSOLT final rotation " \
+ "(ps,pa) = (" \
+ str(self.number_of_channels[0]) + "," \
+ str(self.number_of_channels[1]) + "), " \
+ "(mv,mh) = (" \
+ str(self.decimation_factor[Direction.VERTICAL]) + "," \
+ str(self.decimation_factor[Direction.HORIZONTAL]) + ")"
# Instantiation of orthormal transforms
ps, pa = self.number_of_channels
self.orthTransW0T = OrthonormalTransform(n=ps, mode='Synthesis')
self.orthTransW0T.angles = nn.init.zeros_(self.orthTransW0T.angles)
self.orthTransU0T = OrthonormalTransform(n=pa, mode='Synthesis')
self.orthTransU0T.angles = nn.init.zeros_(self.orthTransU0T.angles)
# No DC leakage
self.no_dc_leakage = no_dc_leakage
def forward(self, X):
nSamples = X.size(dim=0)
nrows = X.size(dim=1)
ncols = X.size(dim=2)
ps, pa = self.number_of_channels
stride = self.decimation_factor
nDecs = stride[0] * stride[1] # math.prod(stride)
# No DC leackage
if self.no_dc_leakage:
self.orthTransW0T.mus[0] = 1
self.orthTransW0T.angles.data[:ps-1] = \
torch.zeros(ps-1,dtype=self.orthTransW0T.angles.data.dtype)
# Process
Ys = X[:, :, :, :ps].view(-1, ps).T
Ya = X[:, :, :, ps:].view(-1, pa).T
ms = int(math.ceil(nDecs / 2.))
ma = int(math.floor(nDecs / 2.))
Zsa = torch.cat((self.orthTransW0T.forward(Ys)[:ms, :],
self.orthTransU0T.forward(Ya)[:ma, :]),
dim=0)
return Zsa.T.view(nSamples, nrows, ncols, nDecs)