我们从Python开源项目中，提取了以下18个代码示例，用于说明如何使用scipy.optimize.fmin_cg()。
def train(self): """ train the network """ from scipy import optimize def costFunction(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return cost def costFunctionGradient(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return grad print print "Training %s..." % self.search() print args = (self._input, self._indicatorFunction) initialParams = self.initialise() params = optimize.fmin_cg(costFunction, x0=initialParams, fprime=costFunctionGradient, \ args=args, maxiter = self._maxiter) self._trainedParams = params
def fit(self, X): from scipy import optimize def objective(W, *args): X, pad = args Obj, DeltaW = self.objective(X, W) return Obj def objectiveG(W, *args): X, pad = args #print type(X) Obj, DeltaW = self.objective(X, W) return DeltaW n, m = np.shape(X) self.n = n args = (X, 1) initW = np.ravel(np.random.rand(self.k, n), order="F") optW = optimize.fmin_cg(objective, x0=initW, fprime=objectiveG, \ args=args, maxiter=self.maxiter, \ callback=self.callback) self.trainedW = optW
def conjugate_gradient(x0, f, f_prime, hessian=None): all_x_i = [x0[0]] all_y_i = [x0[1]] all_f_i = [f(x0)] def store(X): x, y = X all_x_i.append(x) all_y_i.append(y) all_f_i.append(f(X)) optimize.fmin_cg(f, x0, f_prime, callback=store, gtol=1e-12) return all_x_i, all_y_i, all_f_i
def train(self): """ train the network """ from scipy import optimize def costFunction(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return cost def costFunctionGradient(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return grad print() print("Training %s..." % self.search()) print() args = (self._input, self._indicatorFunction) initialParams = self.initialise() params = optimize.fmin_cg(costFunction, x0=initialParams, fprime=costFunctionGradient, \ args=args, maxiter = self._maxiter) self._trainedParams = params
def __call__(self, net, input, target): from scipy.optimize import fmin_cg if 'disp' not in self.kwargs: self.kwargs['disp'] = 0 x = fmin_cg(self.fcn, self.x.copy(), fprime=self.grad, callback=self.step, **self.kwargs) self.x[:] = x return None
def refine(solution, qs, refine_cycle): A_refined = solution.A.copy() def _fun(x, *argv): asx, bsx, csx, asy, bsy, csy, asz, bsz, csz = x h, k, l, qx, qy, qz = argv r1 = (asx*h + bsx*k + csx*l - qx) r2 = (asy*h + bsy*k + csy*l - qy) r3 = (asz*h + bsz*k + csz*l - qz) return r1**2. + r2**2. + r3**2. def _gradient(x, *argv): asx, bsx, csx, asy, bsy, csy, asz, bsz, csz = x h, k, l, qx, qy, qz = argv r1 = (asx*h + bsx*k + csx*l - qx) r2 = (asy*h + bsy*k + csy*l - qy) r3 = (asz*h + bsz*k + csz*l - qz) g_asx, g_bsx, g_csx = 2.*h*r1, 2.*k*r1, 2.*l*r1 g_asy, g_bsy, g_csy = 2.*h*r2, 2.*k*r2, 2.*l*r2 g_asz, g_bsz, g_csz = 2.*h*r3, 2.*k*r3, 2.*l*r3 return np.asarray((g_asx, g_bsx, g_csx, g_asy, g_bsy, g_csy, g_asz, g_bsz, g_csz)) rhkls = solution.rhkls pair_ids = solution.pair_ids for i in range(refine_cycle): for j in range(len(pair_ids)): # refine by each reflection pair_id = pair_ids[j] x0 = A_refined.reshape((-1)) rhkl = rhkls[pair_id,:] q = qs[pair_id,:] args = (rhkl[0], rhkl[1], rhkl[2], q[0], q[1], q[2]) res = fmin_cg(_fun, x0, fprime=_gradient, args=args, disp=0) A_refined = res.reshape((3,3)) eXYZs = np.abs(A_refined.dot(rhkls.T) - qs.T).T dists = norm(eXYZs, axis=1) pair_dist = dists[pair_ids].mean() if pair_dist < solution.pair_dist: solution.A_refined = A_refined solution.pair_dist_refined = pair_dist solution.hkls_refined = np.linalg.inv(A_refined).dot(qs.T).T solution.rhkls_refined = np.rint(solution.hkls_refined) solution.ehkls_refined = np.abs(solution.hkls_refined - solution.rhkls_refined) else: solution.A_refined = solution.A.copy() solution.pair_dist_refined = solution.pair_dist solution.hkls_refined = solution.hkls.copy() solution.rhkls_refined = solution.rhkls.copy() solution.ehkls_refined = solution.ehkls.copy() return solution
def train(self, retry=3): """ train the network """ from scipy import optimize def costFunction(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return cost def costFunctionGradient(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return grad print print "Training %s..." % self.search() print try: targets = self._indicatorFunction except AttributeError: targets = self._targets # attempting to handle the case where the the optimization raises the following warning: # Warning: Desired accuracy not necessarily achieved due to percision loss. # # This seems to mean that the function was unable to calculate a gradient(?) could mean we # have randomly initialised the parmaters in a bad (flat) region of the parameter space. This # seems to be supported by the fact that it usually happpens after 1 iteration. Trying to solve # it by catching the warning and allowing the network to retry with a new randomly initialised set # of parameters a number of times. # Seems the warning is just a print statement not a real warning, so going to have to capture the # the print statement and search for the string and control flow based on that. args = (self._input, targets) counter = 0 output = [] # list to pass to Capturing to store print statements while counter <= retry: if counter > 0: print print "Training %s (Attempt: %d)..." % (self.search(), counter+1) print initialParams = self.initialise() with Capturing(output) as output: params = optimize.fmin_cg(costFunction, x0=initialParams, fprime=costFunctionGradient, \ args=args, maxiter = self._maxiter) print output[counter] if "Warning" in output[counter]: if counter == retry: print "Optimisation has failed %d times. Aborting training!" % (counter+1) print "Optimisation failed on attempt number %d!" % (counter+1) counter += 1 continue self._trainedParams = params break
def train(self, retry=3): """ train the network """ from scipy import optimize def costFunction(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return cost def costFunctionGradient(params, *args): input, targets = args cost, grad = self.costFunction(params, input, targets) return grad print() print("Training %s..." % self.search()) print() try: targets = self._indicatorFunction except AttributeError: targets = self._targets # attempting to handle the case where the the optimization raises the following warning: # Warning: Desired accuracy not necessarily achieved due to percision loss. # # This seems to mean that the function was unable to calculate a gradient(?) could mean we # have randomly initialised the parmaters in a bad (flat) region of the parameter space. This # seems to be supported by the fact that it usually happpens after 1 iteration. Trying to solve # it by catching the warning and allowing the network to retry with a new randomly initialised set # of parameters a number of times. # Seems the warning is just a print statement not a real warning, so going to have to capture the # the print statement and search for the string and control flow based on that. args = (self._input, targets) counter = 0 output = [] # list to pass to Capturing to store print statements while counter <= retry: if counter > 0: print() print("Training %s (Attempt: %d)..." % (self.search(), counter+1)) print() initialParams = self.initialise() with Capturing(output) as output: params = optimize.fmin_cg(costFunction, x0=initialParams, fprime=costFunctionGradient, \ args=args, maxiter = self._maxiter) print(output[counter]) if "Warning" in output[counter]: if counter == retry: print("Optimisation has failed %d times. Aborting training!" % (counter+1)) print("Optimisation failed on attempt number %d!" % (counter+1)) counter += 1 continue self._trainedParams = params break
