def test_mean_poisson_likelihoood_gaussian(): prediction = np.array([1, 1, 1], dtype="float") spe = 0.5 small_mean_likelihood = mean_poisson_likelihood_gaussian( prediction, spe, 0) large_mean_likelihood = mean_poisson_likelihood_gaussian( prediction, spe, 1) assert small_mean_likelihood < large_mean_likelihood
def get_likelihood(self, source_x, source_y, core_x, core_y, energy, x_max_scale, goodness_of_fit=False): """Get the likelihood that the image predicted at the given test position matches the camera image. Parameters ---------- source_x: float Source position of shower in the nominal system (in deg) source_y: float Source position of shower in the nominal system (in deg) core_x: float Core position of shower in tilted telescope system (in m) core_y: float Core position of shower in tilted telescope system (in m) energy: float Shower energy (in TeV) x_max_scale: float Scaling factor applied to geometrically calculated Xmax goodness_of_fit: boolean Determines whether expected likelihood should be subtracted from result Returns ------- float: Likelihood the model represents the camera image at this position """ # First we add units back onto everything. Currently not # handled very well, maybe in future we could just put # everything in the correct units when loading in the class # and ignore them from then on zenith = (np.pi / 2) - self.array_direction.alt.to(u.rad).value azimuth = self.array_direction.az # Geometrically calculate the depth of maximum given this test position x_max = self.get_shower_max(source_x, source_y, core_x, core_y, zenith) x_max *= x_max_scale # Calculate expected Xmax given this energy x_max_exp = guess_shower_depth(energy) # / np.cos(20*u.deg) # Convert to binning of Xmax x_max_bin = x_max - x_max_exp # Check for range if x_max_bin > 200: x_max_bin = 200 if x_max_bin < -100: x_max_bin = -100 # Calculate impact distance for all telescopes impact = np.sqrt(np.power(self.tel_pos_x - core_x, 2) + np.power(self.tel_pos_y - core_y, 2)) # And the expected rotation angle phi = np.arctan2((self.tel_pos_x - core_x), (self.tel_pos_y - core_y)) * u.rad # Rotate and translate all pixels such that they match the # template orientation pix_y_rot, pix_x_rot = self.rotate_translate( self.pixel_x, self.pixel_y, source_x, source_y, phi ) # In the interpolator class we can gain speed advantages by using masked arrays # so we need to make sure here everything is masked prediction = ma.zeros(self.image.shape) prediction.mask = ma.getmask(self.image) time_gradients = np.zeros((self.image.shape[0],2)) # Loop over all telescope types and get prediction for tel_type in np.unique(self.tel_types).tolist(): type_mask = self.tel_types == tel_type prediction[type_mask] = \ self.image_prediction(tel_type, energy * np.ones_like(impact[type_mask]), impact[type_mask], x_max_bin * np.ones_like(impact[type_mask]), pix_x_rot[type_mask] * (180 / math.pi) * -1, pix_y_rot[type_mask] * (180 / math.pi)) if self.use_time_gradient: time_gradients[type_mask] = \ self.predict_time(tel_type, energy * np.ones_like(impact[type_mask]), impact[type_mask], x_max_bin * np.ones_like(impact[type_mask])) if self.use_time_gradient: time_mask = np.logical_and(np.invert(ma.getmask(self.image)), self.time > 0) weight = np.sqrt(self.image) * time_mask rv = norm() sx = pix_x_rot * weight sxx = pix_x_rot * pix_x_rot * weight sy = self.time * weight sxy = self.time * pix_x_rot * weight d = weight.sum(axis=1) * sxx.sum(axis=1) - sx.sum(axis=1) * sx.sum(axis=1) time_fit = (weight.sum(axis=1) * sxy.sum(axis=1) - sx.sum(axis=1) * sy.sum( axis=1)) / d time_fit /= -1 * (180 / math.pi) chi2 = -2 * np.log(rv.pdf((time_fit - time_gradients.T[0])/ time_gradients.T[1])) # Likelihood function will break if we find a NaN or a 0 prediction[np.isnan(prediction)] = 1e-8 prediction[prediction < 1e-8] = 1e-8 prediction *= self.template_scale # Get likelihood that the prediction matched the camera image like = poisson_likelihood_gaussian(self.image, prediction, self.spe, self.ped) like[np.isnan(like)] = 1e9 like *= np.invert(ma.getmask(self.image)) like = ma.MaskedArray(like, mask=ma.getmask(self.image)) array_like = like if goodness_of_fit: return np.sum(like - mean_poisson_likelihood_gaussian(prediction, self.spe, self.ped)) prior_pen = 0 # Add prior penalities if we have them array_like += 1e-8 if "energy" in self.priors: prior_pen += energy_prior(energy, index=-1) if "xmax" in self.priors: prior_pen += xmax_prior(energy, x_max) array_like += prior_pen / float(len(array_like)) if self.array_return: array_like = array_like.ravel() return array_like[np.invert(ma.getmask(array_like))] final_sum = array_like.sum() if self.use_time_gradient: final_sum += chi2.sum() #* np.sum(ma.getmask(self.image)) return final_sum
def get_likelihood(self, source_x, source_y, core_x, core_y, energy, x_max_scale, goodness_of_fit=False): """Get the likelihood that the image predicted at the given test position matches the camera image. Parameters ---------- source_x: float Source position of shower in the nominal system (in deg) source_y: float Source position of shower in the nominal system (in deg) core_x: float Core position of shower in tilted telescope system (in m) core_y: float Core position of shower in tilted telescope system (in m) energy: float Shower energy (in TeV) x_max_scale: float Scaling factor applied to geometrically calculated Xmax goodness_of_fit: boolean Determines whether expected likelihood should be subtracted from result Returns ------- float: Likelihood the model represents the camera image at this position """ # First we add units back onto everything. Currently not # handled very well, maybe in future we could just put # everything in the correct units when loading in the class # and ignore them from then on zenith = (np.pi / 2) - self.array_direction.alt.to(u.rad).value azimuth = self.array_direction.az # Geometrically calculate the depth of maximum given this test position x_max = self.get_shower_max(source_x, source_y, core_x, core_y, zenith) x_max *= x_max_scale # Calculate expected Xmax given this energy x_max_exp = guess_shower_depth(energy) # / np.cos(20*u.deg) # Convert to binning of Xmax x_max_bin = x_max - x_max_exp # Check for range if x_max_bin > 200: x_max_bin = 200 if x_max_bin < -100: x_max_bin = -100 # Calculate impact distance for all telescopes impact = np.sqrt(np.power(self.tel_pos_x - core_x, 2) + np.power(self.tel_pos_y - core_y, 2)) # And the expected rotation angle phi = np.arctan2((self.tel_pos_x - core_x), (self.tel_pos_y - core_y)) * u.rad # Rotate and translate all pixels such that they match the # template orientation pix_y_rot, pix_x_rot = self.rotate_translate( self.pixel_x, self.pixel_y, source_x, source_y, phi ) # In the interpolator class we can gain speed advantages by using masked arrays # so we need to make sure here everything is masked prediction = ma.zeros(self.image.shape) prediction.mask = ma.getmask(self.image) time_gradients = np.zeros((self.image.shape[0],2)) # Loop over all telescope types and get prediction for tel_type in np.unique(self.tel_types).tolist(): type_mask = self.tel_types == tel_type prediction[type_mask] = \ self.image_prediction(tel_type, energy * np.ones_like(impact[type_mask]), impact[type_mask], x_max_bin * np.ones_like(impact[type_mask]), pix_x_rot[type_mask] * (180 / math.pi) * -1, pix_y_rot[type_mask] * (180 / math.pi)) if self.use_time_gradient: time_gradients[type_mask] = \ self.predict_time(tel_type, energy * np.ones_like(impact[type_mask]), impact[type_mask], x_max_bin * np.ones_like(impact[type_mask])) if self.use_time_gradient: time_mask = np.logical_and(np.invert(ma.getmask(self.image)), self.time > 0) weight = np.sqrt(self.image) * time_mask rv = norm() sx = pix_x_rot * weight sxx = pix_x_rot * pix_x_rot * weight sy = self.time * weight sxy = self.time * pix_x_rot * weight d = weight.sum(axis=1) * sxx.sum(axis=1) - sx.sum(axis=1) * sx.sum(axis=1) time_fit = (weight.sum(axis=1) * sxy.sum(axis=1) - sx.sum(axis=1) * sy.sum( axis=1)) / d time_fit /= -1 * (180 / math.pi) chi2 = -2 * np.log(rv.pdf((time_fit - time_gradients.T[0])/ time_gradients.T[1])) # Likelihood function will break if we find a NaN or a 0 prediction[np.isnan(prediction)] = 1e-8 prediction[prediction < 1e-8] = 1e-8 prediction *= self.template_scale # Get likelihood that the prediction matched the camera image like = poisson_likelihood_gaussian(self.image, prediction, self.spe, self.ped) like[np.isnan(like)] = 1e9 like *= np.invert(ma.getmask(self.image)) like = ma.MaskedArray(like, mask=ma.getmask(self.image)) array_like = like if goodness_of_fit: return np.sum(like - mean_poisson_likelihood_gaussian(prediction, self.spe, self.ped)) prior_pen = 0 # Add prior penalities if we have them array_like += 1e-8 if "energy" in self.priors: prior_pen += energy_prior(energy, index=-1) if "xmax" in self.priors: prior_pen += xmax_prior(energy, x_max) array_like += prior_pen / float(len(array_like)) if self.array_return: array_like = array_like.ravel() return array_like[np.invert(ma.getmask(array_like))] final_sum = array_like.sum() if self.use_time_gradient: final_sum += chi2.sum() #* np.sum(ma.getmask(self.image)) return final_sum