def optimize(self, X, C, y, kernelNames, regions, reml=True, maxiter=100):
methodName = ('REML' if reml else 'ML')
if self.verbose:
print 'Finding MKLMM', methodName, 'parameters for', len(
regions), 'regions with lengths:', [
np.sum(r) for r in regions
]
#prepare initial values for sig2e and for fixed effects
hyp0_sig2e, hyp0_fixedEffects = self.getInitialHyps(X, C, y)
#build kernel and train a model
t0 = time.time()
kernel, hyp0_kernels = self.buildKernel(X, kernelNames, regions,
y.var())
hyp0 = np.concatenate((hyp0_sig2e, hyp0_fixedEffects, hyp0_kernels))
args = (kernel, C, y, reml)
funcToSolve = self.infExact_scipy
# # #check gradient correctness
# # if (len(hyp0) < 10):
# # self.optimization_counter=0
# # likFunc = lambda hyp: funcToSolve(hyp, kernel, C, y, reml)[0]
# # gradFunc = lambda hyp: funcToSolve(hyp, kernel, C, y, reml)[1]
# # err = optimize.check_grad(likFunc, gradFunc, hyp0)
# # print 'gradient error:', err
if self.verbose: print 'Beginning Optimization'
self.optimization_counter = 0
optObj = gpUtils.minimize(hyp0, funcToSolve, -maxiter, *args)
if (not optObj.success):
print 'Optimization status:', optObj.status
print 'optimization message:', optObj.message
raise Exception('Optimization failed with message: ' +
optObj.message)
sig2e = np.exp(2 * optObj.x[0])
fixedEffects = optObj.x[1:C.shape[1] + 1]
hyp_kernels = optObj.x[C.shape[1] + 1:]
kernelObj = kernel
if self.verbose:
print 'done in %0.2f' % (time.time() - t0), 'seconds'
print 'sig2e:', sig2e
print 'Fixed effects:', fixedEffects
if (hyp_kernels.shape[0] < 18): print 'kernel params:', hyp_kernels
return kernelObj, hyp_kernels, sig2e, fixedEffects
def optimize(self, X, C, y, kernelNames, regions, reml=True, maxiter=100):
methodName = ('REML' if reml else 'ML')
if self.verbose: print 'Finding MKLMM', methodName, 'parameters for', len(regions), 'regions with lengths:', [np.sum(r) for r in regions]
#prepare initial values for sig2e and for fixed effects
hyp0_sig2e, hyp0_fixedEffects = self.getInitialHyps(X, C, y)
#build kernel and train a model
t0 = time.time()
kernel, hyp0_kernels = self.buildKernel(X, kernelNames, regions, y.var())
hyp0 = np.concatenate((hyp0_sig2e, hyp0_fixedEffects, hyp0_kernels))
args = (kernel, C, y, reml)
funcToSolve = self.infExact_scipy
# # #check gradient correctness
# # if (len(hyp0) < 10):
# # self.optimization_counter=0
# # likFunc = lambda hyp: funcToSolve(hyp, kernel, C, y, reml)[0]
# # gradFunc = lambda hyp: funcToSolve(hyp, kernel, C, y, reml)[1]
# # err = optimize.check_grad(likFunc, gradFunc, hyp0)
# # print 'gradient error:', err
if self.verbose: print 'Beginning Optimization'
self.optimization_counter=0
optObj = gpUtils.minimize(hyp0, funcToSolve, -maxiter, *args)
if (not optObj.success):
print 'Optimization status:', optObj.status
print 'optimization message:', optObj.message
raise Exception('Optimization failed with message: ' + optObj.message)
sig2e = np.exp(2*optObj.x[0])
fixedEffects = optObj.x[1:C.shape[1]+1]
hyp_kernels = optObj.x[C.shape[1]+1:]
kernelObj = kernel
if self.verbose:
print 'done in %0.2f'%(time.time()-t0), 'seconds'
print 'sig2e:', sig2e
print 'Fixed effects:', fixedEffects
if (hyp_kernels.shape[0] < 18): print 'kernel params:', hyp_kernels
return kernelObj, hyp_kernels, sig2e, fixedEffects
def buildKernelAdapt(self, X, C, y, regions, reml=True, maxiter=100):
#prepare initial values for sig2e and for fixed effects
hyp0_sig2e, hyp0_fixedEffects = self.getInitialHyps(X, C, y)
bestKernelNames = []
kernelsListAll = []
hyp_kernels = []
funcToSolve = self.infExact_scipy
yVar = y.var()
for r_i, r in enumerate(regions):
#if (r_i == 0): kernelsToTry = ['lin']
#else:
# kernelsToTry = ['lin', 'poly2_lin', 'rbf_lin', 'nn_lin']
kernelsToTry = ['lin', 'poly2_lin', 'rbf_lin', 'nn_lin']
if self.verbose:
print
print 'selecting a kernel for region', r_i, 'with', r.sum(), 'SNPs'
#add linear kernel
X_lastRegion = X[:, r]
linKernel = kernels.linearKernel(X_lastRegion)
kernelsListAll.append(kernels.ScaledKernel(linKernel))
kernelsListAll.append(None)
bestFun = np.inf
bestKernelName = None
best_hyp0 = None
bestKernel = None
bestPval = np.inf
#iterate over every possible kernel
for kernelToTry in kernelsToTry:
hyp0 = [0.5*np.log(0.5*yVar)]
if self.verbose: print 'Testing kernel:', kernelToTry
#create the kernel
if (kernelToTry == 'lin'):
kernel = None
df = None
elif (kernelToTry == 'rbf_lin'):
kernel = kernels.RBFKernel(X_lastRegion)
hyp0.append(np.log(1.0)) #ell
df = 2
elif (kernelToTry == 'nn_lin'):
kernel = kernels.NNKernel(X_lastRegion)
hyp0.append(np.log(1.0)) #ell
df = 2
elif (kernelToTry == 'poly2_lin'):
kernel = kernels.Poly2KernelHomo(linKernel)
df = 1
else:
raise Exception('unrecognized kernel name')
if (kernel is not None):
#scale the kernel
kernel = kernels.ScaledKernel(kernel)
hyp0.append(0.5*np.log(0.5*yVar)) #scaling hyp
#add the kernel as the final kernel in the kernels list
kernelsListAll[-1] = kernel
sumKernel = kernels.SumKernel(kernelsListAll)
else:
sumKernel = kernels.SumKernel(kernelsListAll[:-1])
#test log likelihood obtained with this kernel for this region
args = (sumKernel, C, y, reml)
self.optimization_counter=0
hyp0_all = np.concatenate((hyp0_sig2e, hyp0_fixedEffects, hyp_kernels+hyp0))
optObj = gpUtils.minimize(hyp0_all, funcToSolve, -maxiter, *args)
if (not optObj.success):
print 'Optimization status:', optObj.status
print 'optimization message:', optObj.message
raise Exception('optimization failed')
print 'final LL: %0.5e'%(-optObj.fun)
if (kernelToTry == 'lin'):
linLL = -optObj.fun
pVal = 1.0
else:
llDiff = -optObj.fun - linLL
if (llDiff < 0): pVal = 1.0
else: pVal = 0.5*stats.chi2(df).sf(llDiff)
print 'llDiff: %0.5e'%llDiff, 'pVal:%0.5e'%pVal
if (kernelToTry == 'lin' or (pVal < bestPval and (len(kernelsToTry)==1 or pVal < 0.05/(len(kernelsToTry)-1)))):
bestOptObj = optObj
bestPval = pVal
bestKernelName = kernelToTry
best_hyp0 = hyp0
best_sumKernel = sumKernel
bestKernel = kernel
if (bestKernel is not None): kernelsListAll[-1] = bestKernel
else: kernelsListAll = kernelsListAll[:-1]
hyp_kernels += best_hyp0
bestKernelNames.append(bestKernelName)
if self.verbose: print 'selected kernel:', bestKernelName
if self.verbose:
print 'selected kernels:', bestKernelNames
print
return bestKernelNames
def buildKernelAdapt(self, X, C, y, regions, reml=True, maxiter=100):
#prepare initial values for sig2e and for fixed effects
hyp0_sig2e, hyp0_fixedEffects = self.getInitialHyps(X, C, y)
bestKernelNames = []
kernelsListAll = []
hyp_kernels = []
funcToSolve = self.infExact_scipy
yVar = y.var()
for r_i, r in enumerate(regions):
#if (r_i == 0): kernelsToTry = ['lin']
#else:
# kernelsToTry = ['lin', 'poly2_lin', 'rbf_lin', 'nn_lin']
kernelsToTry = ['lin', 'poly2_lin', 'rbf_lin', 'nn_lin']
if self.verbose:
print
print 'selecting a kernel for region', r_i, 'with', r.sum(
), 'SNPs'
#add linear kernel
X_lastRegion = X[:, r]
linKernel = kernels.linearKernel(X_lastRegion)
kernelsListAll.append(kernels.ScaledKernel(linKernel))
kernelsListAll.append(None)
bestFun = np.inf
bestKernelName = None
best_hyp0 = None
bestKernel = None
bestPval = np.inf
#iterate over every possible kernel
for kernelToTry in kernelsToTry:
hyp0 = [0.5 * np.log(0.5 * yVar)]
if self.verbose: print 'Testing kernel:', kernelToTry
#create the kernel
if (kernelToTry == 'lin'):
kernel = None
df = None
elif (kernelToTry == 'rbf_lin'):
kernel = kernels.RBFKernel(X_lastRegion)
hyp0.append(np.log(1.0)) #ell
df = 2
elif (kernelToTry == 'nn_lin'):
kernel = kernels.NNKernel(X_lastRegion)
hyp0.append(np.log(1.0)) #ell
df = 2
elif (kernelToTry == 'poly2_lin'):
kernel = kernels.Poly2KernelHomo(linKernel)
df = 1
else:
raise Exception('unrecognized kernel name')
if (kernel is not None):
#scale the kernel
kernel = kernels.ScaledKernel(kernel)
hyp0.append(0.5 * np.log(0.5 * yVar)) #scaling hyp
#add the kernel as the final kernel in the kernels list
kernelsListAll[-1] = kernel
sumKernel = kernels.SumKernel(kernelsListAll)
else:
sumKernel = kernels.SumKernel(kernelsListAll[:-1])
#test log likelihood obtained with this kernel for this region
args = (sumKernel, C, y, reml)
self.optimization_counter = 0
hyp0_all = np.concatenate(
(hyp0_sig2e, hyp0_fixedEffects, hyp_kernels + hyp0))
optObj = gpUtils.minimize(hyp0_all, funcToSolve, -maxiter,
*args)
if (not optObj.success):
print 'Optimization status:', optObj.status
print 'optimization message:', optObj.message
raise Exception('optimization failed')
print 'final LL: %0.5e' % (-optObj.fun)
if (kernelToTry == 'lin'):
linLL = -optObj.fun
pVal = 1.0
else:
llDiff = -optObj.fun - linLL
if (llDiff < 0): pVal = 1.0
else: pVal = 0.5 * stats.chi2(df).sf(llDiff)
print 'llDiff: %0.5e' % llDiff, 'pVal:%0.5e' % pVal
if (kernelToTry == 'lin'
or (pVal < bestPval and
(len(kernelsToTry) == 1 or pVal < 0.05 /
(len(kernelsToTry) - 1)))):
bestOptObj = optObj
bestPval = pVal
bestKernelName = kernelToTry
best_hyp0 = hyp0
best_sumKernel = sumKernel
bestKernel = kernel
if (bestKernel is not None): kernelsListAll[-1] = bestKernel
else: kernelsListAll = kernelsListAll[:-1]
hyp_kernels += best_hyp0
bestKernelNames.append(bestKernelName)
if self.verbose: print 'selected kernel:', bestKernelName
if self.verbose:
print 'selected kernels:', bestKernelNames
print
return bestKernelNames
