123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596 |
- import numpy as np
- from sklearn.gaussian_process import GaussianProcessRegressor
- # Different kernels available in sklearn:
- from sklearn.gaussian_process.kernels import RBF, ConstantKernel, Matern, WhiteKernel, ExpSineSquared, RationalQuadratic
- from matplotlib import pyplot as plt
- plt.rcParams.update({
- 'axes.labelsize': 16,
- 'xtick.labelsize': 14,
- 'ytick.labelsize': 14
- })
- # Set random seed for same results
- np.random.seed(12345)
- # Load data
- inFileName = 'DATA/original_histograms/mass_mm_higgs_Background.npz'
- with np.load(inFileName) as data:
- bin_centers = data['bin_centers']
- bin_values = np.log(data['bin_values'])
- bin_errors = data['bin_errors'] / data['bin_values'] # How does the error scale when taking f(x) = ln(x)?
- # Mask
- signal_region = (119, 131)
- mask = (bin_centers < signal_region[0]) | (bin_centers > signal_region[1])
- bin_centers_masked = bin_centers[mask]
- bin_values_masked = bin_values[mask]
- bin_errors_masked = bin_errors[mask]
- # Set hyper-parameter bounds for ConstantKernel
- nEvts = np.max(bin_values)
- const0 = 1.
- const_low = 1e-1
- const_hi = 1e3
- # Set hyper-parameter bounds for RBF kernel
- RBF0 = 1.
- RBF_low = 1e-1
- RBF_high = 1e2
- # A) Define kernel: ConstantKernel * RBF:
- kernel_RBF = ConstantKernel(const0, constant_value_bounds=(const_low, const_hi)) * RBF(RBF0, length_scale_bounds=(RBF_low, RBF_high))
- # B) Define kernel: ConstantKernel * Matern:
- kernel_Matern = ConstantKernel(const0, constant_value_bounds=(const_low, const_hi)) * Matern(RBF0, length_scale_bounds=(RBF_low, RBF_high), nu=1.5)
- # Transform x data into 2d vector!
- X = np.atleast_2d(bin_centers_masked).T # true datapoints
- X_to_predict = np.atleast_2d(np.linspace(110, 160, 1000)).T # what to predict
- y = bin_values_masked
- # Initialize Gaussian Process Regressor: !!! alpha = bin_errors, 2*bin_errors or bin_errors**2? Your task to figure out!!!
- gp = GaussianProcessRegressor(kernel=kernel_RBF, n_restarts_optimizer=1, alpha=bin_errors_masked**2)
- # Fit on X with values y
- gp.fit(X, y)
- print('Final kernel combination:\n', gp.kernel_)
- # Predict
- y_pred, sigma = gp.predict(X_to_predict, return_std=True)
- y_pred_sparse, sigma_sparse = gp.predict(np.atleast_2d(bin_centers).T, return_std=True)
- fig, axes = plt.subplot_mosaic([['main'],['main'],['main'],['ratio']], sharex=True, figsize=(8, 8))
- # Main pad
- axes['main'].set_title('Example GPR with RBF kernel', fontsize=20, fontweight='bold')
- axes['main'].fill_between(X_to_predict.ravel(), y_pred - sigma, y_pred + sigma)
- axes['main'].scatter(bin_centers_masked, bin_values_masked, color='r', linewidth=0.5, marker='o', s=25, label='Data')
- axes['main'].scatter(bin_centers[~mask], bin_values[~mask], color='g', marker='+', s=100, label='Blinded data (not used in the fit)')
- axes['main'].plot(X_to_predict, y_pred, color='k', label='GPR Prediction')
- axes['main'].set_ylabel('ln(events/bin)', fontsize=16)
- axes['main'].set_ylim((8, 12))
- axes['main'].legend(fontsize=16)
- # Ratio pad
- axes['ratio'].errorbar(bin_centers, bin_values/y_pred_sparse, yerr=sigma_sparse, color='k', linewidth=0., elinewidth=0.5, marker='.')
- axes['ratio'].axhline(1, c='k', lw=1, alpha=0.7)
- axes['ratio'].set_xlabel(r'$m_{\mu\mu}$ [GeV]', fontsize=16)
- axes['ratio'].set_ylabel('Data/Pred.', fontsize=16)
- # Make ratio plot labels symmetric around 1.
- max = np.max(np.abs(axes['ratio'].get_yticks() - 1.)) / 1.5
- axes['ratio'].set_ylim((1. - max, 1. + max))
- axes['ratio'].grid()
- # Make an inner plot
- axes['main'].plot([113, 119], [9.9, 10.9], 'k--')
- axes['main'].plot([138, 131], [9.9, 10.3], 'k--')
- ax_inner = fig.add_axes([0.2125, 0.3625, 0.4, 0.25])
- ax_inner.set_title('zoom-in', fontsize=20)
- ax_inner.scatter(bin_centers[~mask], bin_values[~mask], color='g', marker='+', s=100)
- ax_inner.plot(X_to_predict, y_pred, color='k')
- ax_inner.set_xlim(signal_region)
- ax_inner.set_ylim((10.3, 11))
- ax_inner.grid()
- plt.tight_layout()
- plt.show()
- plt.savefig('GPR_simple.pdf')
|