def test_buildQSigma():
tol = 1e-5
dataFilename = os.path.join(os.path.dirname(__file__), "data/get_full_from_lowplusdiag.mat")
mat = loadmat(dataFilename)
nLatents = mat['q_sqrt'].shape[1]
q_sqrt = [torch.from_numpy(mat['q_sqrt'][(0,i)]).permute(2,0,1) for i in range(nLatents)]
q_diag = [torch.from_numpy(mat['q_diag'][(0,i)]).permute(2,0,1) for i in range(nLatents)]
q_sigma = [torch.from_numpy(mat['q_sigma'][(0,i)]).permute(2,0,1) for i in range(nLatents)]
qMu0 = [[] for i in range(nLatents)]
params0 = {"qMu0": qMu0, "qSVec0": q_sqrt, "qSDiag0": q_diag}
qU = SVPosteriorOnIndPoints()
qU.setInitialParams(initialParams=params0)
qSigma = qU.buildQSigma();
error = torch.tensor([(qSigma[k]-q_sigma[k]).norm() for k in range(len(qSigma))]).sum()
assert(error<tol)
def test_computeMeansAndVars_assocTimes():
tol = 5e-6
dataFilename = os.path.join(os.path.dirname(__file__), "data/Estep_Objective_PointProcess_svGPFA.mat")
mat = loadmat(dataFilename)
nLatents = mat["Z"].shape[0]
nTrials = mat["Z"][0,0].shape[2]
qMu0 = [torch.from_numpy(mat["q_mu"][(0,i)]).type(torch.DoubleTensor).permute(2,0,1) for i in range(nLatents)]
qSVec0 = [torch.from_numpy(mat["q_sqrt"][(0,i)]).type(torch.DoubleTensor).permute(2,0,1) for i in range(nLatents)]
qSDiag0 = [torch.from_numpy(mat["q_diag"][(0,i)]).type(torch.DoubleTensor).permute(2,0,1) for i in range(nLatents)]
Z0 = [torch.from_numpy(mat["Z"][(i,0)]).type(torch.DoubleTensor).permute(2,0,1) for i in range(nLatents)]
Y = [torch.from_numpy(mat["Y"][tr,0]).type(torch.DoubleTensor) for tr in range(nTrials)]
mu_k = [torch.from_numpy(mat["mu_k_Spikes"][0,tr]).type(torch.DoubleTensor) for tr in range(nTrials)]
var_k = [torch.from_numpy(mat["var_k_Spikes"][0,tr]).type(torch.DoubleTensor) for tr in range(nTrials)]
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0,k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k,0][0]),
float(hprs[k,0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0,k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k,0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s"%(kernelNames[k]))
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS()
indPointsLocsAndTimesKMS = IndPointsLocsAndAssocTimesKMS()
qK = SVPosteriorOnLatentsAssocTimes(svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=
indPointsLocsAndTimesKMS)
quParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
kmsParams0 = {"kernelsParams0": kernelsParams0,
"inducingPointsLocs0": Z0}
qU.setInitialParams(initialParams=quParams0)
indPointsLocsKMS.setKernels(kernels=kernels)
indPointsLocsKMS.setInitialParams(initialParams=kmsParams0)
indPointsLocsKMS.buildKernelsMatrices()
indPointsLocsAndTimesKMS.setKernels(kernels=kernels)
indPointsLocsAndTimesKMS.setInitialParams(initialParams=kmsParams0)
indPointsLocsAndTimesKMS.setTimes(times=Y)
indPointsLocsAndTimesKMS.buildKernelsMatrices()
qKMu, qKVar = qK.computeMeansAndVars()
for tr in range(nTrials):
qKMuError = math.sqrt(((mu_k[tr]-qKMu[tr])**2).mean())
assert(qKMuError<tol)
qKVarError = math.sqrt(((var_k[tr]-qKVar[tr])**2).mean())
assert(qKVarError<tol)
def test_evalSumAcrossTrialsAndNeurons_pointProcessExpLink():
tol = 3e-4
yNonStackedFilename = os.path.join(os.path.dirname(__file__),
"data/YNonStacked.mat")
dataFilename = os.path.join(
os.path.dirname(__file__),
"data/Estep_Objective_PointProcess_svGPFA.mat")
mat = loadmat(dataFilename)
nLatents = len(mat['Z'])
nTrials = mat['Z'][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat['q_mu'][(0, i)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat['q_sqrt'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat['q_diag'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
t = torch.from_numpy(mat['ttQuad']).type(torch.DoubleTensor).permute(
2, 0, 1)
Z0 = [
torch.from_numpy(mat['Z'][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
C0 = torch.from_numpy(mat["C"]).type(torch.DoubleTensor)
b0 = torch.from_numpy(mat["b"]).type(torch.DoubleTensor).squeeze()
hermQuadPoints = torch.from_numpy(mat['xxHerm']).type(torch.DoubleTensor)
hermQuadWeights = torch.from_numpy(mat['wwHerm']).type(torch.DoubleTensor)
legQuadPoints = torch.from_numpy(mat['ttQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
legQuadWeights = torch.from_numpy(mat['wwQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
Elik = torch.from_numpy(mat['Elik'])
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
yMat = loadmat(yNonStackedFilename)
YNonStacked_tmp = yMat['YNonStacked']
nNeurons = YNonStacked_tmp[0, 0].shape[0]
YNonStacked = [[[] for n in range(nNeurons)] for r in range(nTrials)]
for r in range(nTrials):
for n in range(nNeurons):
YNonStacked[r][n] = torch.from_numpy(
YNonStacked_tmp[r, 0][n, 0][:, 0]).type(torch.DoubleTensor)
linkFunction = torch.exp
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
qHParams0 = {"C0": C0, "d0": b0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
initialParams = {
"svPosteriorOnIndPoints": qUParams0,
"kernelsMatricesStore": kmsParams0,
"svEmbedding": qHParams0
}
quadParams = {
"legQuadPoints": legQuadPoints,
"legQuadWeights": legQuadWeights
}
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS()
indPointsLocsAndAllTimesKMS = IndPointsLocsAndAllTimesKMS()
indPointsLocsAndAssocTimesKMS = IndPointsLocsAndAssocTimesKMS()
qKAllTimes = SVPosteriorOnLatentsAllTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAllTimesKMS)
qKAssocTimes = SVPosteriorOnLatentsAssocTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAssocTimesKMS)
qHAllTimes = LinearSVEmbeddingAllTimes(svPosteriorOnLatents=qKAllTimes)
qHAssocTimes = LinearSVEmbeddingAssocTimes(
svPosteriorOnLatents=qKAssocTimes)
eLL = PointProcessELLExpLink(svEmbeddingAllTimes=qHAllTimes,
svEmbeddingAssocTimes=qHAssocTimes)
eLL.setKernels(kernels=kernels)
eLL.setInitialParams(initialParams=initialParams)
eLL.setMeasurements(measurements=YNonStacked)
eLL.setQuadParams(quadParams=quadParams)
eLL.buildKernelsMatrices()
sELL = eLL.evalSumAcrossTrialsAndNeurons()
sELLerror = abs(sELL - Elik)
assert (sELLerror < tol)
def test_evalSumAcrossLatentsTrials():
tol = 1e-5
dataFilename = os.path.join(
os.path.dirname(__file__),
"data/Estep_Objective_PointProcess_svGPFA.mat")
mat = loadmat(dataFilename)
nLatents = mat['q_sqrt'].shape[1]
nTrials = mat['Z'][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat['q_mu'][(0, i)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat['q_sqrt'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat['q_diag'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
Z0 = [
torch.from_numpy(mat['Z'][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
matKLDiv = torch.from_numpy(mat['KLd'])
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
indPointsLocsKMS = IndPointsLocsKMS()
qU = SVPosteriorOnIndPoints()
klDiv = KLDivergence(indPointsLocsKMS=indPointsLocsKMS,
svPosteriorOnIndPoints=qU)
qU.setInitialParams(initialParams=qUParams0)
indPointsLocsKMS.setKernels(kernels=kernels)
indPointsLocsKMS.setInitialParams(initialParams=kmsParams0)
indPointsLocsKMS.buildKernelsMatrices()
klDivEval = klDiv.evalSumAcrossLatentsAndTrials()
klError = abs(matKLDiv - klDivEval)
assert (klError < tol)
def test_computeMeansAndVars_allTimes():
tol = 5e-6
dataFilename = os.path.join(
os.path.dirname(__file__),
"data/Estep_Objective_PointProcess_svGPFA.mat")
mat = loadmat(dataFilename)
nLatents = mat["Z"].shape[0]
nTrials = mat["Z"][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat["q_mu"][(0, i)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat["q_sqrt"][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat["q_diag"][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
t = torch.from_numpy(mat["ttQuad"]).type(torch.DoubleTensor).permute(
2, 0, 1)
Z0 = [
torch.from_numpy(mat["Z"][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
Y = [
torch.from_numpy(mat["Y"][tr, 0]).type(torch.DoubleTensor)
for tr in range(nTrials)
]
C0 = torch.from_numpy(mat["C"]).type(torch.DoubleTensor)
b0 = torch.from_numpy(mat["b"]).type(torch.DoubleTensor)
mu_h = torch.from_numpy(mat["mu_h_Quad"]).type(torch.DoubleTensor).permute(
2, 0, 1)
var_h = torch.from_numpy(mat["var_h_Quad"]).type(
torch.DoubleTensor).permute(2, 0, 1)
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS()
indPointsLocsAndAllTimesKMS = IndPointsLocsAndAllTimesKMS()
qK = SVPosteriorOnLatentsAllTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAllTimesKMS)
qH = LinearSVEmbeddingAllTimes(svPosteriorOnLatents=qK)
qH.setKernels(kernels=kernels)
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
qHParams0 = {"C0": C0, "d0": b0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
initialParams = {
"svPosteriorOnIndPoints": qUParams0,
"kernelsMatricesStore": kmsParams0,
"svEmbedding": qHParams0
}
qH.setInitialParams(initialParams=initialParams)
qH.setTimes(times=t)
qH.buildKernelsMatrices()
qHMu, qHVar = qH.computeMeansAndVars()
qHMuError = math.sqrt(((mu_h - qHMu)**2).mean())
assert (qHMuError < tol)
qHVarError = math.sqrt(((var_h - qHVar)**2).mean())
assert (qHVarError < tol)
def test_computeMeansAndVars_assocTimes():
tol = 5e-6
dataFilename = os.path.join(
os.path.dirname(__file__),
"data/Estep_Objective_PointProcess_svGPFA.mat")
mat = loadmat(dataFilename)
nLatents = mat["Z"].shape[0]
nTrials = mat["Z"][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat["q_mu"][(0, i)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat["q_sqrt"][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat["q_diag"][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
t = torch.from_numpy(mat["ttQuad"]).type(torch.DoubleTensor).permute(
2, 0, 1)
Z0 = [
torch.from_numpy(mat["Z"][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
Y = [
torch.from_numpy(mat["Y"][tr, 0]).type(torch.DoubleTensor)
for tr in range(nTrials)
]
C0 = torch.from_numpy(mat["C"]).type(torch.DoubleTensor)
b0 = torch.from_numpy(mat["b"]).type(torch.DoubleTensor)
mu_h = [
torch.from_numpy(mat["mu_h_Spikes"][0, i]).type(
torch.DoubleTensor).squeeze() for i in range(nTrials)
]
var_h = [
torch.from_numpy(mat["var_h_Spikes"][0, i]).type(
torch.DoubleTensor).squeeze() for i in range(nTrials)
]
index = [
torch.from_numpy(mat["index"][i, 0][:, 0]).type(torch.ByteTensor) - 1
for i in range(nTrials)
]
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS()
indPointsLocsAndAssocTimesKMS = IndPointsLocsAndAssocTimesKMS()
qK = SVPosteriorOnLatentsAssocTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAssocTimesKMS)
qH = LinearSVEmbeddingAssocTimes(svPosteriorOnLatents=qK)
qH.setKernels(kernels=kernels)
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
qHParams0 = {"C0": C0, "d0": b0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
initialParams = {
"svPosteriorOnIndPoints": qUParams0,
"kernelsMatricesStore": kmsParams0,
"svEmbedding": qHParams0
}
qH.setInitialParams(initialParams=initialParams)
qH.setKernels(kernels=kernels)
qH.setTimes(times=Y)
qH.setNeuronForSpikeIndex(neuronForSpikeIndex=index)
# begin patches because we are not using SVPosteriorOnLatentsAssocTimes in
# conjunction with SVPosteriorOnLatentsAllTimes
qU.setInitialParams(initialParams=qUParams0)
indPointsLocsKMS.setKernels(kernels=kernels)
indPointsLocsKMS.setInitialParams(initialParams=kmsParams0)
indPointsLocsKMS.setKernels(kernels=kernels)
indPointsLocsKMS.buildKernelsMatrices()
# end patches because we are not using SVPosteriorOnLatentsAssocTimes in
# conjunction with SVPosteriorOnLatentsAllTimes
qH.buildKernelsMatrices()
qHMu, qHVar = qH.computeMeansAndVars()
for i in range(len(mu_h)):
qHMuError = math.sqrt(torch.sum(
(mu_h[i] - qHMu[i])**2)) / mu_h[i].shape[0]
assert (qHMuError < tol)
qHVarError = math.sqrt(torch.sum((var_h[i]-qHVar[i])**2))/\
var_h[i].shape[0]
assert (qHVarError < tol)
def test_maximize_pointProcess():
tol = 1e-5
yNonStackedFilename = os.path.join(os.path.dirname(__file__),
"data/YNonStacked.mat")
dataFilename = os.path.join(os.path.dirname(__file__),
"data/variationalEM.mat")
# yNonStackedFilename = os.path.expanduser("~/tmp/svGPFA/ci/data/YNonStacked.mat")
# dataFilename = os.path.expanduser("~/tmp/svGPFA/ci/data/variationalEM.mat")
mat = loadmat(dataFilename)
nLatents = len(mat['Z0'])
nTrials = mat['Z0'][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat['q_mu0'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat['q_sqrt0'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat['q_diag0'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
Z0 = [
torch.from_numpy(mat['Z0'][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
C0 = torch.from_numpy(mat["C0"]).type(torch.DoubleTensor)
b0 = torch.from_numpy(mat["b0"]).type(torch.DoubleTensor).squeeze()
legQuadPoints = torch.from_numpy(mat['ttQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
legQuadWeights = torch.from_numpy(mat['wwQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
yMat = loadmat(yNonStackedFilename)
YNonStacked_tmp = yMat['YNonStacked']
nNeurons = YNonStacked_tmp[0, 0].shape[0]
YNonStacked = [[[] for n in range(nNeurons)] for r in range(nTrials)]
for r in range(nTrials):
for n in range(nNeurons):
YNonStacked[r][n] = torch.from_numpy(
YNonStacked_tmp[r, 0][n, 0][:, 0]).type(torch.DoubleTensor)
linkFunction = torch.exp
kernelNames = mat["kernelNames"]
hprs = mat["hprs0"]
leasLowerBound = mat['lowerBound'][0, 0]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
indPointsLocsKMSEpsilon = 1e-2
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS(epsilon=indPointsLocsKMSEpsilon)
indPointsLocsAndAllTimesKMS = IndPointsLocsAndAllTimesKMS()
indPointsLocsAndAssocTimesKMS = IndPointsLocsAndAssocTimesKMS()
qKAllTimes = SVPosteriorOnLatentsAllTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAllTimesKMS)
qKAssocTimes = SVPosteriorOnLatentsAssocTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAssocTimesKMS)
qHAllTimes = LinearSVEmbeddingAllTimes(svPosteriorOnLatents=qKAllTimes)
qHAssocTimes = LinearSVEmbeddingAssocTimes(
svPosteriorOnLatents=qKAssocTimes)
eLL = PointProcessELLExpLink(svEmbeddingAllTimes=qHAllTimes,
svEmbeddingAssocTimes=qHAssocTimes)
klDiv = KLDivergence(indPointsLocsKMS=indPointsLocsKMS,
svPosteriorOnIndPoints=qU)
svlb = SVLowerBound(eLL=eLL, klDiv=klDiv)
svlb.setKernels(kernels=kernels)
svEM = SVEM()
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
qHParams0 = {"C0": C0, "d0": b0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
initialParams = {
"svPosteriorOnIndPoints": qUParams0,
"kernelsMatricesStore": kmsParams0,
"svEmbedding": qHParams0
}
quadParams = {
"legQuadPoints": legQuadPoints,
"legQuadWeights": legQuadWeights
}
'''
optimParams = {"emMaxNIter":20,
#
"eStepMaxNIter":100,
"eStepTol":1e-3,
"eStepLR":1e-3,
"eStepNIterDisplay":10,
#
"mStepModelParamsMaxNIter":100,
"mStepModelParamsTol":1e-3,
"mStepModelParamsLR":1e-3,
"mStepModelParamsNIterDisplay":10,
#
"mStepKernelParamsMaxNIter":100,
"mStepKernelParamsTol":1e-3,
"mStepKernelParamsLR":1e-5,
"mStepKernelParamsNIterDisplay":10,
#
"mStepIndPointsMaxNIter":100,
"mStepIndPointsParamsTol":1e-3,
"mStepIndPointsLR":1e-3,
"mStepIndPointsNIterDisplay":10}
'''
optimParams = {
"emMaxNIter": 3,
#
"eStepEstimate": True,
"eStepMaxNIter": 20,
"eStepTol": 1e-2,
"eStepLR": 1e-2,
"eStepLineSearchFn": "strong_wolfe",
"eStepNIterDisplay": 1,
#
"mStepModelParamsEstimate": True,
"mStepModelParamsMaxNIter": 20,
"mStepModelParamsTol": 1e-2,
"mStepModelParamsLR": 1e-3,
"mStepModelParamsLineSearchFn": "strong_wolfe",
"mStepModelParamsNIterDisplay": 1,
#
"mStepKernelParamsEstimate": True,
"mStepKernelParamsMaxNIter": 20,
"mStepKernelParamsTol": 1e-2,
"mStepKernelParamsLR": 1e-4,
"mStepKernelParamsLineSearchFn": "strong_wolfe",
"mStepKernelParamsNIterDisplay": 1,
#
"mStepIndPointsEstimate": True,
"mStepIndPointsMaxNIter": 20,
"mStepIndPointsTol": 1e-2,
"mStepIndPointsLR": 1e-3,
"mStepIndPointsLineSearchFn": "strong_wolfe",
"mStepIndPointsNIterDisplay": 1,
#
"verbose": True
}
lowerBoundHist, elapsedTimeHist = svEM.maximize(
model=svlb,
measurements=YNonStacked,
initialParams=initialParams,
quadParams=quadParams,
optimParams=optimParams,
plotLatentsEstimates=False)
assert (lowerBoundHist[-1] > leasLowerBound)
def test_mStepIndPoints_pointProcess():
tol = 1e-5
yNonStackedFilename = os.path.join(os.path.dirname(__file__),
"data/YNonStacked.mat")
dataFilename = os.path.join(os.path.dirname(__file__),
"data/inducingPointsMstep_all.mat")
mat = loadmat(dataFilename)
nLatents = len(mat['Z0'])
nTrials = mat['Z0'][0, 0].shape[2]
qMu0 = [
torch.from_numpy(mat['q_mu'][(0, i)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
qSVec0 = [
torch.from_numpy(mat['q_sqrt'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
qSDiag0 = [
torch.from_numpy(mat['q_diag'][(0, i)]).type(
torch.DoubleTensor).permute(2, 0, 1) for i in range(nLatents)
]
Z0 = [
torch.from_numpy(mat['Z0'][(i, 0)]).type(torch.DoubleTensor).permute(
2, 0, 1) for i in range(nLatents)
]
C0 = torch.from_numpy(mat["C"]).type(torch.DoubleTensor)
b0 = torch.from_numpy(mat["b"]).type(torch.DoubleTensor).squeeze()
nLowerBound = mat['nLowerBound'][0, 0]
legQuadPoints = torch.from_numpy(mat['ttQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
legQuadWeights = torch.from_numpy(mat['wwQuad']).type(
torch.DoubleTensor).permute(2, 0, 1)
yMat = loadmat(yNonStackedFilename)
YNonStacked_tmp = yMat['YNonStacked']
nNeurons = YNonStacked_tmp[0, 0].shape[0]
YNonStacked = [[[] for n in range(nNeurons)] for r in range(nTrials)]
for r in range(nTrials):
for n in range(nNeurons):
YNonStacked[r][n] = torch.from_numpy(
YNonStacked_tmp[r, 0][n, 0][:, 0]).type(torch.DoubleTensor)
linkFunction = torch.exp
kernelNames = mat["kernelNames"]
hprs = mat["hprs"]
kernels = [[None] for k in range(nLatents)]
kernelsParams0 = [[None] for k in range(nLatents)]
for k in range(nLatents):
if np.char.equal(kernelNames[0, k][0], "PeriodicKernel"):
kernels[k] = PeriodicKernel(scale=1.0)
kernelsParams0[k] = torch.tensor(
[float(hprs[k, 0][0]),
float(hprs[k, 0][1])],
dtype=torch.double)
elif np.char.equal(kernelNames[0, k][0], "rbfKernel"):
kernels[k] = ExponentialQuadraticKernel(scale=1.0)
kernelsParams0[k] = torch.tensor([float(hprs[k, 0][0])],
dtype=torch.double)
else:
raise ValueError("Invalid kernel name: %s" % (kernelNames[k]))
indPointsLocsKMSEpsilon = 1e-5
qU = SVPosteriorOnIndPoints()
indPointsLocsKMS = IndPointsLocsKMS(epsilon=indPointsLocsKMSEpsilon)
indPointsLocsAndAllTimesKMS = IndPointsLocsAndAllTimesKMS()
indPointsLocsAndAssocTimesKMS = IndPointsLocsAndAssocTimesKMS()
qKAllTimes = SVPosteriorOnLatentsAllTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAllTimesKMS)
qKAssocTimes = SVPosteriorOnLatentsAssocTimes(
svPosteriorOnIndPoints=qU,
indPointsLocsKMS=indPointsLocsKMS,
indPointsLocsAndTimesKMS=indPointsLocsAndAssocTimesKMS)
qHAllTimes = LinearSVEmbeddingAllTimes(svPosteriorOnLatents=qKAllTimes)
qHAssocTimes = LinearSVEmbeddingAssocTimes(
svPosteriorOnLatents=qKAssocTimes)
eLL = PointProcessELLExpLink(svEmbeddingAllTimes=qHAllTimes,
svEmbeddingAssocTimes=qHAssocTimes)
klDiv = KLDivergence(indPointsLocsKMS=indPointsLocsKMS,
svPosteriorOnIndPoints=qU)
svlb = SVLowerBound(eLL=eLL, klDiv=klDiv)
svEM = SVEM()
qUParams0 = {"qMu0": qMu0, "qSVec0": qSVec0, "qSDiag0": qSDiag0}
qHParams0 = {"C0": C0, "d0": b0}
kmsParams0 = {"kernelsParams0": kernelsParams0, "inducingPointsLocs0": Z0}
initialParams = {
"svPosteriorOnIndPoints": qUParams0,
"kernelsMatricesStore": kmsParams0,
"svEmbedding": qHParams0
}
quadParams = {
"legQuadPoints": legQuadPoints,
"legQuadWeights": legQuadWeights
}
svlb.setKernels(kernels=kernels)
svlb.setInitialParams(initialParams=initialParams)
svlb.setMeasurements(measurements=YNonStacked)
svlb.setQuadParams(quadParams=quadParams)
svlb.buildKernelsMatrices()
res = svEM._mStepIndPoints(model=svlb,
maxNIter=10,
tol=1e-3,
lr=1e-3,
lineSearchFn="strong_wolfe",
verbose=True,
nIterDisplay=1)
assert (res["lowerBound"] > (-nLowerBound))