Esempio n. 1
0
    def fit_origin_minimise(self, seed=[0, 0, 1], test_function=None):
        ''' fits the origin of the gamma with a minimisation procedure
            this function expects that get_great_circles has been run already
            a seed should be given otherwise it defaults to "straight up"
            supperted functions to minimise are an M-estimator and the
            negative sum of the angles to all normal vectors of the
            great circles

            Parameters:
            -----------
            seed : length-3 array
                starting point of the minimisation
            test_function : member function if this class
                either _n_angle_sum or _MEst (or own implementation...)
                defaults to _n_angle_sum if none is given
                _n_angle_sum seemingly superior to _MEst

            Returns:
            --------
            direction : length-3 numpy array as dimensionless quantity
                best fit for the origin of the gamma from
                the minimisation process
        '''

        if test_function is None: test_function = self._n_angle_sum
        '''
            using the sum of the cosines of each direction with
            every otherdirection as weight;
            don't use the product -- with many steep angles, the product will
            become too small and the weight (and the whole fit) useless
        '''
        weights = [np.sum([linalg.length(np.cross(A.norm, B.norm))
                           for A in self.circles.values()]
                          ) * B.weight
                   for B in self.circles.values()]

        ''' minimising the test function '''
        self.fit_result_origin = minimize(test_function, seed,
                                          args=(self.circles, weights),
                                          method='BFGS',
                                          options={'disp': False}
                                          )

        return np.array(linalg.normalise(self.fit_result_origin.x))*u.dimless
Esempio n. 2
0
    def fit_origin_minimise(self, seed=(0, 0, 1), test_function=neg_angle_sum):
        ''' Fits the origin of the gamma with a minimisation procedure this
        function expects that
        :func:`get_great_circles<ctapipe.reco.FitGammaHillas.get_great_circles>`
        has been run already. A seed should be given otherwise it defaults to
        "straight up" supperted functions to minimise are an M-estimator and the
        negative sum of the angles to all normal vectors of the great
        circles

        Parameters
        -----------
        seed : length-3 array
            starting point of the minimisation
        test_function : function object, optional (default: neg_angle_sum)
            either neg_angle_sum or MEst (or own implementation...)
            neg_angle_sum seemingly superior to MEst

        Returns
        -------
        direction : length-3 numpy array as dimensionless quantity
            best fit for the origin of the gamma from
            the minimisation process

        '''

        # using the sum of the cosines of each direction with every
        # other direction as weight; don't use the product -- with many
        # steep angles, the product will become too small and the
        # weight (and the whole fit) useless

        weights = [np.sum([linalg.length(np.cross(A.norm, B.norm))
                           for A in self.circles.values()]
                          ) * B.weight
                   for B in self.circles.values()]

        # minimising the test function
        self.fit_result_origin = minimize(test_function, seed,
                                          args=(self.circles, weights),
                                          method='BFGS',
                                          options={'disp': False}
                                          )

        return np.array(linalg.normalise(self.fit_result_origin.x)) * u.dimless
Esempio n. 3
0
def MEst(origin, circles, weights):
    '''calculates the M-Estimator: a modified χ² that becomes
    asymptotically linear for high values and is therefore less
    sensitive to outliers

    the test is performed to maximise the angles between the fit
    direction and the all the normal vectors of the great circles

    .. math::

      M_{Est} = \sum_i{ 2*\sqrt{1 + \chi^2} - 2}

    .. note::

      seemingly inferior to negative sum of sin(angle)...


    Parameters
    -----------
    origin : length-3 array
        direction vector of the gamma's origin used as seed
    circles : GreatCircle array
        collection of great circles created from the camera images
    weights : array
        list of weights for each image/great circle

    Returns
    -------
    MEstimator : float


    '''

    sin_ang = np.array([linalg.length(np.cross(origin, circ.norm))
                        for circ in circles.values()])
    return -2 * np.sum(weights * np.sqrt((1 + np.square(sin_ang))) - 2)
    return -np.sum(weights * np.square(sin_ang))

    ang = np.array([linalg.angle(origin, circ.norm)
                    for circ in circles.values()])
    return np.sum(weights * np.sqrt(2. + (ang - np.pi / 2.) ** 2))
Esempio n. 4
0
    def _MEst(self, origin, circles, weights):
        '''
            calculates the M-Estimator:
            a modified chi2 that becomes asymptotically linear for high values
            and is therefore less sensitive to outliers

            the test is performed to maximise the angles between the
            fit direction and the all the normal vectors of the great circles

            Parameters:
            -----------
            origin : length-3 array
                direction vector of the gamma's origin used as seed
            circles : GreatCircle array
                collection of great circles created from the camera images
            weights : array
                list of weights for each image/great circle

            Returns:
            --------
            MEstimator : float


            Algorithm:
            ----------
            M-Est = sum[  weight * sqrt( 2 * chi**2 ) ]


            Note:
            -----
            seemingly inferior to negative sum of sin(angle)...
        '''

        sin_ang = np.array([linalg.length(np.cross(origin, circ.norm))
                            for circ in circles.values()])
        return np.sum(weights*np.sqrt(2. + (sin_ang-np.pi/2.)**2))

        ang = np.array([linalg.angle(origin, circ.norm)
                        for circ in circles.values()])
        ang[ang > np.pi/2.] = np.pi-ang[ang > np.pi/2]
        return np.sum(weights*np.sqrt(2. + (ang-np.pi/2.)**2))
Esempio n. 5
0
def n_angle_sum(origin, circles, weights):
    '''calculates the negative sum of the angle between the fit direction
    and all the normal vectors of the great circles

    Parameters
    -----------
    origin : length-3 array
        direction vector of the gamma's origin used as seed
    circles : GreatCircle array
        collection of great circles created from the camera images
    weights : array
        list of weights for each image/great circle

    Returns
    --------
    n_sum_angles : float
        negative of the sum of the angles between the test direction
        and all normal vectors of the given great circles

    '''

    sin_ang = np.array([linalg.length(np.cross(origin, circ.norm))
                        for circ in circles.values()])
    return -np.sum(weights * sin_ang)
Esempio n. 6
0
        source = hessio_event_source(filename,
                                     allowed_tels=allowed_tels,
                                     max_events=args.max_events)

        # loop that cleans and parametrises the images and performs the reconstruction
        for (event, hillas_dict, n_tels,
             tot_signal, max_signals, pos_fit, dir_fit, h_max,
             err_est_pos, err_est_dir) in preper.prepare_event(source):

            n_faint = 0
            for tel_id in hillas_dict.keys():
                cam_id = event.inst.subarray.tel[tel_id].camera.cam_id
                moments = hillas_dict[tel_id]
                tel_pos = np.array(event.inst.tel_pos[tel_id][:2]) * u.m
                impact_dist = linalg.length(tel_pos - pos_fit)

                if moments.size > pe_thersh:
                    n_faint += 1

                feature_events[cam_id]["impact_dist"] = impact_dist / dist_unit
                feature_events[cam_id]["sum_signal_evt"] = tot_signal
                feature_events[cam_id]["max_signal_cam"] = max_signals[tel_id]
                feature_events[cam_id]["sum_signal_cam"] = moments.size
                feature_events[cam_id]["N_LST"] = n_tels["LST"]
                feature_events[cam_id]["N_MST"] = n_tels["MST"]
                feature_events[cam_id]["N_SST"] = n_tels["SST"]
                feature_events[cam_id]["width"] = moments.width / dist_unit
                feature_events[cam_id]["length"] = moments.length / dist_unit
                feature_events[cam_id]["skewness"] = moments.skewness
                feature_events[cam_id]["kurtosis"] = moments.kurtosis
Esempio n. 7
0
def main():

    # your favourite units here
    energy_unit = u.TeV
    angle_unit = u.deg
    dist_unit = u.m

    parser = make_argparser()
    parser.add_argument('-o', '--outfile', type=str,
                        help="if given, write output file with reconstruction results")
    parser.add_argument('--plot_c', action='store_true',
                        help="plot camera-wise displays")
    group = parser.add_mutually_exclusive_group()
    group.add_argument('--proton', action='store_true',
                       help="do protons instead of gammas")
    group.add_argument('--electron', action='store_true',
                       help="do electrons instead of gammas")

    args = parser.parse_args()

    if args.infile_list:
        filenamelist = []
        for f in args.infile_list:
            filenamelist += glob("{}/{}".format(args.indir, f))
    elif args.proton:
        filenamelist = glob("{}/proton/*gz".format(args.indir))
        channel = "proton"
    elif args.electron:
        filenamelist = glob("{}/electron/*gz".format(args.indir))
        channel = "electron"
    elif args.gamma:
        filenamelist = glob("{}/gamma/*gz".format(args.indir))
        channel = "gamma"
    else:
        raise ValueError("don't know which input to use...")
    filenamelist.sort()

    if not filenamelist:
        print("no files found; check indir: {}".format(args.indir))
        exit(-1)
    else:
        print("found {} files".format(len(filenamelist)))

    tel_phi = {}
    tel_theta = {}

    # keeping track of events and where they were rejected
    Eventcutflow = CutFlow("EventCutFlow")
    Imagecutflow = CutFlow("ImageCutFlow")

    # takes care of image cleaning
    cleaner = ImageCleaner(mode=args.mode, cutflow=Imagecutflow,
                           wavelet_options=args.raw,
                           skip_edge_events=args.skip_edge_events, island_cleaning=True)

    # the class that does the shower reconstruction
    shower_reco = HillasReconstructor()

    shower_max_estimator = ShowerMaxEstimator("paranal")

    preper = EventPreparer(cleaner=cleaner,
                           hillas_parameters=hillas_parameters, shower_reco=shower_reco,
                           event_cutflow=Eventcutflow, image_cutflow=Imagecutflow,
                           # event/image cuts:
                           allowed_cam_ids=[],  # means: all
                           min_ntel=3,
                           min_charge=args.min_charge,
                           min_pixel=3)

    # a signal handler to abort the event loop but still do the post-processing
    signal_handler = SignalHandler()
    signal.signal(signal.SIGINT, signal_handler)

    try:
        # this class defines the reconstruction parameters to keep track of
        class RecoEvent(tb.IsDescription):
            NTels_trigg = tb.Int16Col(dflt=1, pos=0)
            NTels_clean = tb.Int16Col(dflt=1, pos=1)
            EnMC = tb.Float32Col(dflt=1, pos=2)
            xi = tb.Float32Col(dflt=1, pos=3)
            DeltaR = tb.Float32Col(dflt=1, pos=4)
            ErrEstPos = tb.Float32Col(dflt=1, pos=5)
            ErrEstDir = tb.Float32Col(dflt=1, pos=6)
            h_max = tb.Float32Col(dflt=1, pos=7)

        reco_outfile = tb.open_file(
            args.outfile, mode="w",
            # if we don't want to write the event list to disk, need to add more arguments
            **({} if args.store else {"driver": "H5FD_CORE",
                                      "driver_core_backing_store": False}))
        reco_table = reco_outfile.create_table("/", "reco_event", RecoEvent)
        reco_event = reco_table.row
    except:
        reco_event = RecoEvent()
        print("no pytables installed?")

    # ##        #######   #######  ########
    # ##       ##     ## ##     ## ##     ##
    # ##       ##     ## ##     ## ##     ##
    # ##       ##     ## ##     ## ########
    # ##       ##     ## ##     ## ##
    # ##       ##     ## ##     ## ##
    # ########  #######   #######  ##

    cam_id_map = {}

    # define here which telescopes to loop over
    allowed_tels = None
    # allowed_tels = prod3b_tel_ids("L+F+D")
    for i, filename in enumerate(filenamelist[:args.last]):
        print("file: {i} filename = {filename}".format(i=i, filename=filename))

        source = hessio_event_source(filename,
                                     allowed_tels=allowed_tels,
                                     max_events=args.max_events)

        # loop that cleans and parametrises the images and performs the reconstruction
        for (event, hillas_dict, n_tels,
             tot_signal, max_signal, pos_fit, dir_fit, h_max,
             err_est_pos, err_est_dir) in preper.prepare_event(source):

            shower = event.mc

            org_alt = u.Quantity(shower.alt).to(u.deg)
            org_az = u.Quantity(shower.az).to(u.deg)
            if org_az > 180 * u.deg:
                org_az -= 360 * u.deg

            org_the = alt_to_theta(org_alt)
            org_phi = az_to_phi(org_az)
            if org_phi > 180 * u.deg:
                org_phi -= 360 * u.deg
            if org_phi < -180 * u.deg:
                org_phi += 360 * u.deg

            shower_org = linalg.set_phi_theta(org_phi, org_the)
            shower_core = convert_astropy_array([shower.core_x, shower.core_y])

            xi = linalg.angle(dir_fit, shower_org).to(angle_unit)
            diff = linalg.length(pos_fit[:2] - shower_core)

            # print some performance
            print()
            print("xi = {:4.3f}".format(xi))
            print("pos = {:4.3f}".format(diff))
            print("h_max reco: {:4.3f}".format(h_max.to(u.km)))
            print("err_est_dir: {:4.3f}".format(err_est_dir.to(angle_unit)))
            print("err_est_pos: {:4.3f}".format(err_est_pos))

            try:
                # store the reconstruction data in the PyTable
                reco_event["NTels_trigg"] = n_tels["tot"]
                reco_event["NTels_clean"] = len(shower_reco.circles)
                reco_event["EnMC"] = event.mc.energy / energy_unit
                reco_event["xi"] = xi / angle_unit
                reco_event["DeltaR"] = diff / dist_unit
                reco_event["ErrEstPos"] = err_est_pos / dist_unit
                reco_event["ErrEstDir"] = err_est_dir / angle_unit
                reco_event["h_max"] = h_max / dist_unit
                reco_event.append()
                reco_table.flush()

                print()
                print("xi res (68-percentile) = {:4.3f} {}"
                      .format(np.percentile(reco_table.cols.xi, 68), angle_unit))
                print("core res (68-percentile) = {:4.3f} {}"
                      .format(np.percentile(reco_table.cols.DeltaR, 68), dist_unit))
                print("h_max (median) = {:4.3f} {}"
                      .format(np.percentile(reco_table.cols.h_max, 50), dist_unit))

            except NoPyTables:
                pass

            if args.plot_c:
                from mpl_toolkits.mplot3d import Axes3D
                fig = plt.figure()
                ax = fig.gca(projection='3d')
                for c in shower_reco.circles.values():
                    points = [c.pos + t * c.a * u.km for t in np.linspace(0, 15, 3)]
                    ax.plot(*np.array(points).T, linewidth=np.sqrt(c.weight) / 10)
                    ax.scatter(*c.pos[:, None].value, s=np.sqrt(c.weight))
                plt.xlabel("x")
                plt.ylabel("y")
                plt.pause(.1)

                # this plots
                # • the MC shower core
                # • the reconstructed shower core
                # • the used telescopes
                # • and the trace of the Hillas plane on the ground
                plt.figure()
                for tel_id, c in shower_reco.circles.items():
                    plt.scatter(c.pos[0], c.pos[1], s=np.sqrt(c.weight))
                    plt.gca().annotate(tel_id, (c.pos[0].value, c.pos[1].value))
                    plt.plot([c.pos[0].value-500*c.norm[1], c.pos[0].value+500*c.norm[1]],
                             [c.pos[1].value+500*c.norm[0], c.pos[1].value-500*c.norm[0]],
                             linewidth=np.sqrt(c.weight)/10)
                plt.scatter(*pos_fit[:2], c="black", marker="*", label="fitted")
                plt.scatter(*shower_core[:2], c="black", marker="P", label="MC")
                plt.legend()
                plt.xlabel("x")
                plt.ylabel("y")
                plt.xlim(-1400, 1400)
                plt.ylim(-1400, 1400)
                plt.show()

            if signal_handler.stop: break
        if signal_handler.stop: break

    print("\n" + "="*35 + "\n")
    print("xi res (68-percentile) = {:4.3f} {}"
          .format(np.percentile(reco_table.cols.xi, 68), angle_unit))
    print("core res (68-percentile) = {:4.3f} {}"
          .format(np.percentile(reco_table.cols.DeltaR, 68), dist_unit))
    print("h_max (median) = {:4.3f} {}"
          .format(np.percentile(reco_table.cols.h_max, 50), dist_unit))

    # print the cutflows for telescopes and camera images
    print("\n\n")
    Eventcutflow("min2Tels trig")
    print()
    Imagecutflow(sort_column=1)

    # if we don't want to plot anything, we can exit now
    if not args.plot:
        return

    # ########  ##        #######  ########  ######
    # ##     ## ##       ##     ##    ##    ##    ##
    # ##     ## ##       ##     ##    ##    ##
    # ########  ##       ##     ##    ##     ######
    # ##        ##       ##     ##    ##          ##
    # ##        ##       ##     ##    ##    ##    ##
    # ##        ########  #######     ##     ######

    plt.figure()
    plt.hist(reco_table.cols.h_max, bins=np.linspace(000, 15000, 51, True))
    plt.title(channel)
    plt.xlabel("h_max reco")
    plt.pause(.1)

    figure = plt.figure()
    xi_edges = np.linspace(0, 5, 20)
    plt.hist(reco_table.cols.xi, bins=xi_edges, log=True)
    plt.xlabel(r"$\xi$ / deg")
    if args.write:
        save_fig('{}/reco_xi_{}'.format(args.plots_dir, args.mode))
    plt.pause(.1)

    plt.figure()
    plt.hist(reco_table.cols.ErrEstDir[:],
             bins=np.linspace(0, 20, 50))
    plt.title(channel)
    plt.xlabel("beta")
    plt.pause(.1)

    plt.figure()
    plt.hist(np.log10(reco_table.cols.xi[:] / reco_table.cols.ErrEstDir[:]), bins=50)
    plt.title(channel)
    plt.xlabel("log_10(xi / beta)")
    plt.pause(.1)

    # convert the xi-list into a dict with the number of used telescopes as keys
    xi_vs_tel = {}
    for xi, ntel in zip(reco_table.cols.xi, reco_table.cols.NTels_clean):
        if ntel not in xi_vs_tel:
            xi_vs_tel[ntel] = [xi]
        else:
            xi_vs_tel[ntel].append(xi)

    print(args.mode)
    for ntel, xis in sorted(xi_vs_tel.items()):
        print("NTel: {} -- median xi: {}".format(ntel, np.median(xis)))
        # print("histogram:", np.histogram(xis, bins=xi_edges))

    # create a list of energy bin-edges and -centres for violin plots
    Energy_edges = np.linspace(2, 8, 13)
    Energy_centres = (Energy_edges[1:]+Energy_edges[:-1])/2.

    # convert the xi-list in to an energy-binned dict with the bin centre as keys
    xi_vs_energy = {}
    for en, xi in zip(reco_table.cols.EnMC, reco_table.cols.xi):

        # get the bin number this event belongs into
        sbin = np.digitize(np.log10(en), Energy_edges)-1

        # the central value of the bin is the key for the dictionary
        if Energy_centres[sbin] not in xi_vs_energy:
            xi_vs_energy[Energy_centres[sbin]]  = [xi]
        else:
            xi_vs_energy[Energy_centres[sbin]] += [xi]

    # plotting the angular error as violin plots with binning in
    # number of telescopes and shower energy
    figure = plt.figure()
    plt.subplot(211)
    plt.violinplot([np.log10(a) for a in xi_vs_tel.values()],
                   [a for a in xi_vs_tel.keys()],
                   points=60, widths=.75, showextrema=False, showmedians=True)
    plt.xlabel("Number of Telescopes")
    plt.ylabel(r"log($\xi$ / deg)")
    plt.ylim(-3, 2)
    plt.grid()

    plt.subplot(212)
    plt.violinplot([np.log10(a) for a in xi_vs_energy.values()],
                   [a for a in xi_vs_energy.keys()],
                   points=60, widths=(Energy_edges[1] - Energy_edges[0]) / 1.5,
                   showextrema=False, showmedians=True)
    plt.xlabel(r"log(Energy / GeV)")
    plt.ylabel(r"log($\xi$ / deg)")
    plt.ylim(-3, 2)
    plt.grid()

    plt.tight_layout()
    if args.write:
        save_fig('{}/reco_xi_vs_E_NTel_{}'.format(args.plots_dir, args.mode))

    plt.pause(.1)

    # convert the diffs-list into a dict with the number of used telescopes as keys
    diff_vs_tel = {}
    for diff, ntel in zip(reco_table.cols.DeltaR, reco_table.cols.NTels_clean):
        if ntel not in diff_vs_tel:
            diff_vs_tel[ntel] = [diff]
        else:
            diff_vs_tel[ntel].append(diff)

    # convert the diffs-list in to an energy-binned dict with the bin centre as keys
    diff_vs_energy = {}
    for en, diff in zip(reco_table.cols.EnMC, reco_table.cols.DeltaR):

        # get the bin number this event belongs into
        sbin = np.digitize(np.log10(en), Energy_edges) - 1

        # the central value of the bin is the key for the dictionary
        if Energy_centres[sbin] not in diff_vs_energy:
            diff_vs_energy[Energy_centres[sbin]]  = [diff]
        else:
            diff_vs_energy[Energy_centres[sbin]] += [diff]

    # plotting the core position error as violin plots with binning in
    # number of telescopes an shower energy
    plt.figure()
    plt.subplot(211)
    plt.violinplot([np.log10(a) for a in diff_vs_tel.values()],
                   [a for a in diff_vs_tel.keys()],
                   points=60, widths=.75, showextrema=False, showmedians=True)
    plt.xlabel("Number of Telescopes")
    plt.ylabel(r"log($\Delta R$ / m)")
    plt.grid()

    plt.subplot(212)
    plt.violinplot([np.log10(a) for a in diff_vs_energy.values()],
                   [a for a in diff_vs_energy.keys()],
                   points=60, widths=(Energy_edges[1] - Energy_edges[0]) / 1.5,
                   showextrema=False, showmedians=True)
    plt.xlabel(r"log(Energy / GeV)")
    plt.ylabel(r"log($\Delta R$ / m)")
    plt.grid()

    plt.tight_layout()
    if args.write:
        save_fig('{}/reco_dist_vs_E_NTel_{}'.format(args.plots_dir, args.mode))
    plt.show()
def main():

    # your favourite units here
    energy_unit = u.TeV
    angle_unit = u.deg
    dist_unit = u.m

    agree_threshold = .5
    min_tel = 3

    parser = make_argparser()
    parser.add_argument('--classifier', type=str,
                        default='data/classifier_pickle/classifier'
                                '_{mode}_{cam_id}_{classifier}.pkl')
    parser.add_argument('--regressor', type=str,
                        default='data/classifier_pickle/regressor'
                                '_{mode}_{cam_id}_{regressor}.pkl')
    parser.add_argument('-o', '--outfile', type=str, default="",
                        help="location to write the classified events to.")
    parser.add_argument('--wave_dir', type=str, default=None,
                        help="directory where to find mr_filter. "
                             "if not set look in $PATH")
    parser.add_argument('--wave_temp_dir', type=str, default='/tmp/', help="directory "
                        "where mr_filter to store the temporary fits files")

    group = parser.add_mutually_exclusive_group()
    group.add_argument('--proton', action='store_true',
                       help="do protons instead of gammas")
    group.add_argument('--electron', action='store_true',
                       help="do electrons instead of gammas")

    args = parser.parse_args()

    if args.infile_list:
        filenamelist = []
        for f in args.infile_list:
            filenamelist += glob("{}/{}".format(args.indir, f))
        filenamelist.sort()
    elif args.proton:
        filenamelist = sorted(glob("{}/proton/*gz".format(args.indir)))
    elif args.electron:
        filenamelist = glob("{}/electron/*gz".format(args.indir))
        channel = "electron"
    else:
        filenamelist = sorted(glob("{}/gamma/*gz".format(args.indir)))

    if not filenamelist:
        print("no files found; check indir: {}".format(args.indir))
        exit(-1)

    # keeping track of events and where they were rejected
    Eventcutflow = CutFlow("EventCutFlow")
    Imagecutflow = CutFlow("ImageCutFlow")

    # takes care of image cleaning
    cleaner = ImageCleaner(mode=args.mode, cutflow=Imagecutflow,
                           wavelet_options=args.raw,
                           skip_edge_events=False, island_cleaning=True)

    # the class that does the shower reconstruction
    shower_reco = HillasReconstructor()

    preper = EventPreparer(
                cleaner=cleaner, hillas_parameters=hillas_parameters,
                shower_reco=shower_reco,
                event_cutflow=Eventcutflow, image_cutflow=Imagecutflow,
                # event/image cuts:
                allowed_cam_ids=[],
                min_ntel=2, min_charge=args.min_charge, min_pixel=3)

    # wrapper for the scikit-learn classifier
    classifier = EventClassifier.load(
                    args.classifier.format(**{
                            "mode": args.mode,
                            "wave_args": "mixed",
                            "classifier": 'RandomForestClassifier',
                            "cam_id": "{cam_id}"}),
                    cam_id_list=args.cam_ids)

    # wrapper for the scikit-learn regressor
    regressor = EnergyRegressor.load(
                    args.regressor.format(**{
                            "mode": args.mode,
                            "wave_args": "mixed",
                            "regressor": "RandomForestRegressor",
                            "cam_id": "{cam_id}"}),
                    cam_id_list=args.cam_ids)

    ClassifierFeatures = namedtuple(
        "ClassifierFeatures", (
            "impact_dist",
            "sum_signal_evt",
            "max_signal_cam",
            "sum_signal_cam",
            "N_LST",
            "N_MST",
            "N_SST",
            "width",
            "length",
            "skewness",
            "kurtosis",
            "h_max",
            "err_est_pos",
            "err_est_dir"))

    EnergyFeatures = namedtuple(
        "EnergyFeatures", (
            "impact_dist",
            "sum_signal_evt",
            "max_signal_cam",
            "sum_signal_cam",
            "N_LST",
            "N_MST",
            "N_SST",
            "width",
            "length",
            "skewness",
            "kurtosis",
            "h_max",
            "err_est_pos",
            "err_est_dir"))

    # catch ctr-c signal to exit current loop and still display results
    signal_handler = SignalHandler()
    signal.signal(signal.SIGINT, signal_handler)

    # this class defines the reconstruction parameters to keep track of
    class RecoEvent(tb.IsDescription):
        Run_ID = tb.Int16Col(dflt=-1, pos=0)
        Event_ID = tb.Int16Col(dflt=-1, pos=1)
        NTels_trig = tb.Int16Col(dflt=0, pos=0)
        NTels_reco = tb.Int16Col(dflt=0, pos=1)
        NTels_reco_lst = tb.Int16Col(dflt=0, pos=2)
        NTels_reco_mst = tb.Int16Col(dflt=0, pos=3)
        NTels_reco_sst = tb.Int16Col(dflt=0, pos=4)
        MC_Energy = tb.Float32Col(dflt=np.nan, pos=5)
        reco_Energy = tb.Float32Col(dflt=np.nan, pos=6)
        reco_phi = tb.Float32Col(dflt=np.nan, pos=7)
        reco_theta = tb.Float32Col(dflt=np.nan, pos=8)
        off_angle = tb.Float32Col(dflt=np.nan, pos=9)
        xi = tb.Float32Col(dflt=np.nan, pos=10)
        DeltaR = tb.Float32Col(dflt=np.nan, pos=11)
        ErrEstPos = tb.Float32Col(dflt=np.nan, pos=12)
        ErrEstDir = tb.Float32Col(dflt=np.nan, pos=13)
        gammaness = tb.Float32Col(dflt=np.nan, pos=14)

    channel = "gamma" if "gamma" in " ".join(filenamelist) else "proton"
    reco_outfile = tb.open_file(
            mode="w",
            # if no outfile name is given (i.e. don't to write the event list to disk),
            # need specify two "driver" arguments
            **({"filename": args.outfile} if args.outfile else
               {"filename": "no_outfile.h5",
                "driver": "H5FD_CORE", "driver_core_backing_store": False}))

    reco_table = reco_outfile.create_table("/", "reco_events", RecoEvent)
    reco_event = reco_table.row

    allowed_tels = None  # all telescopes
    allowed_tels = prod3b_tel_ids("L+N+D")
    for i, filename in enumerate(filenamelist[:args.last]):
        # print(f"file: {i} filename = {filename}")

        source = hessio_event_source(filename,
                                     allowed_tels=allowed_tels,
                                     max_events=args.max_events)

        # loop that cleans and parametrises the images and performs the reconstruction
        for (event, hillas_dict, n_tels,
             tot_signal, max_signals, pos_fit, dir_fit, h_max,
             err_est_pos, err_est_dir) in preper.prepare_event(source, True):

            # now prepare the features for the classifier
            cls_features_evt = {}
            reg_features_evt = {}
            if hillas_dict is not None:
              for tel_id in hillas_dict.keys():
                Imagecutflow.count("pre-features")

                tel_pos = np.array(event.inst.tel_pos[tel_id][:2]) * u.m

                moments = hillas_dict[tel_id]

                impact_dist = linalg.length(tel_pos - pos_fit)
                cls_features_tel = ClassifierFeatures(
                    impact_dist=impact_dist / u.m,
                    sum_signal_evt=tot_signal,
                    max_signal_cam=max_signals[tel_id],
                    sum_signal_cam=moments.size,
                    N_LST=n_tels["LST"],
                    N_MST=n_tels["MST"],
                    N_SST=n_tels["SST"],
                    width=moments.width / u.m,
                    length=moments.length / u.m,
                    skewness=moments.skewness,
                    kurtosis=moments.kurtosis,
                    h_max=h_max / u.m,
                    err_est_pos=err_est_pos / u.m,
                    err_est_dir=err_est_dir / u.deg
                )

                reg_features_tel = EnergyFeatures(
                    impact_dist=impact_dist / u.m,
                    sum_signal_evt=tot_signal,
                    max_signal_cam=max_signals[tel_id],
                    sum_signal_cam=moments.size,
                    N_LST=n_tels["LST"],
                    N_MST=n_tels["MST"],
                    N_SST=n_tels["SST"],
                    width=moments.width / u.m,
                    length=moments.length / u.m,
                    skewness=moments.skewness,
                    kurtosis=moments.kurtosis,
                    h_max=h_max / u.m,
                    err_est_pos=err_est_pos / u.m,
                    err_est_dir=err_est_dir / u.deg
                )

                if np.isnan(cls_features_tel).any() or np.isnan(reg_features_tel).any():
                    continue

                Imagecutflow.count("features nan")

                cam_id = event.inst.subarray.tel[tel_id].camera.cam_id

                try:
                    reg_features_evt[cam_id] += [reg_features_tel]
                    cls_features_evt[cam_id] += [cls_features_tel]
                except KeyError:
                    reg_features_evt[cam_id] = [reg_features_tel]
                    cls_features_evt[cam_id] = [cls_features_tel]

            # save basic event infos
            reco_event["MC_Energy"] = event.mc.energy.to(energy_unit).value
            reco_event["Event_ID"] = event.r1.event_id
            reco_event["Run_ID"] = event.r1.run_id

            if cls_features_evt and reg_features_evt:

                predict_energ = regressor.predict_by_event([reg_features_evt])["mean"][0]
                predict_proba = classifier.predict_proba_by_event([cls_features_evt])
                gammaness = predict_proba[0, 0]

                # the MC direction of origin of the simulated particle
                shower = event.mc
                shower_core = np.array([shower.core_x / u.m, shower.core_y / u.m]) * u.m
                shower_org = linalg.set_phi_theta(shower.az + 90 * u.deg,
                                                  90. * u.deg - shower.alt)

                # and how the reconstructed direction compares to that
                xi = linalg.angle(dir_fit, shower_org)
                phi, theta = linalg.get_phi_theta(dir_fit)
                phi = (phi if phi > 0 else phi + 360 * u.deg)

                DeltaR = linalg.length(pos_fit[:2] - shower_core)

                # TODO: replace with actual array pointing direction
                array_pointing = linalg.set_phi_theta(0 * u.deg, 20. * u.deg)
                # angular offset between the reconstructed direction and the array
                # pointing
                off_angle = linalg.angle(dir_fit, array_pointing)

                reco_event["NTels_trig"] = len(event.dl0.tels_with_data)
                reco_event["NTels_reco"] = len(hillas_dict)
                reco_event["NTels_reco_lst"] = n_tels["LST"]
                reco_event["NTels_reco_mst"] = n_tels["MST"]
                reco_event["NTels_reco_sst"] = n_tels["SST"]
                reco_event["reco_Energy"] = predict_energ.to(energy_unit).value
                reco_event["reco_phi"] = phi / angle_unit
                reco_event["reco_theta"] = theta / angle_unit
                reco_event["off_angle"] = off_angle / angle_unit
                reco_event["xi"] = xi / angle_unit
                reco_event["DeltaR"] = DeltaR / dist_unit
                reco_event["ErrEstPos"] = err_est_pos / dist_unit
                reco_event["ErrEstDir"] = err_est_dir / angle_unit
                reco_event["gammaness"] = gammaness
                reco_event.append()
                reco_table.flush()

            if signal_handler.stop:
                break
        if signal_handler.stop:
            break

    try:
        print()
        Eventcutflow()
        print()
        Imagecutflow()

        # do some simple event selection
        # and print the corresponding selection efficiency
        N_selected = len([x for x in reco_table.where(
            """(NTels_reco > min_tel) & (gammaness > agree_threshold)""")])
        N_total = len(reco_table)
        print("\nfraction selected events:")
        print("{} / {} = {} %".format(N_selected, N_total, N_selected / N_total * 100))

    except ZeroDivisionError:
        pass

    print("\nlength filenamelist:", len(filenamelist[:args.last]))

    # do some plotting if so desired
    if args.plot:
        gammaness = [x['gammaness'] for x in reco_table]
        NTels_rec = [x['NTels_reco'] for x in reco_table]
        NTel_bins = np.arange(np.min(NTels_rec), np.max(NTels_rec) + 2) - .5

        NTels_rec_lst = [x['NTels_reco_lst'] for x in reco_table]
        NTels_rec_mst = [x['NTels_reco_mst'] for x in reco_table]
        NTels_rec_sst = [x['NTels_reco_sst'] for x in reco_table]

        reco_energy = np.array([x['reco_Energy'] for x in reco_table])
        mc_energy = np.array([x['MC_Energy'] for x in reco_table])

        fig = plt.figure(figsize=(15, 5))
        plt.suptitle(" ** ".join([args.mode, "protons" if args.proton else "gamma"]))
        plt.subplots_adjust(left=0.05, right=0.97, hspace=0.39, wspace=0.2)

        ax = plt.subplot(131)
        histo = np.histogram2d(NTels_rec, gammaness,
                               bins=(NTel_bins, np.linspace(0, 1, 11)))[0].T
        histo_normed = histo / histo.max(axis=0)
        im = ax.imshow(histo_normed, interpolation='none', origin='lower',
                       aspect='auto',
                       # extent=(*NTel_bins[[0, -1]], 0, 1),
                       cmap=plt.cm.inferno)
        ax.set_xlabel("NTels")
        ax.set_ylabel("drifted gammaness")
        plt.title("Total Number of Telescopes")

        # next subplot

        ax = plt.subplot(132)
        histo = np.histogram2d(NTels_rec_sst, gammaness,
                               bins=(NTel_bins, np.linspace(0, 1, 11)))[0].T
        histo_normed = histo / histo.max(axis=0)
        im = ax.imshow(histo_normed, interpolation='none', origin='lower',
                       aspect='auto',
                       # extent=(*NTel_bins[[0, -1]], 0, 1),
                       cmap=plt.cm.inferno)
        ax.set_xlabel("NTels")
        plt.setp(ax.get_yticklabels(), visible=False)
        plt.title("Number of SSTs")

        # next subplot

        ax = plt.subplot(133)
        histo = np.histogram2d(NTels_rec_mst, gammaness,
                               bins=(NTel_bins, np.linspace(0, 1, 11)))[0].T
        histo_normed = histo / histo.max(axis=0)
        im = ax.imshow(histo_normed, interpolation='none', origin='lower',
                       aspect='auto',
                       # extent=(*NTel_bins[[0, -1]], 0, 1),
                       cmap=plt.cm.inferno)
        cb = fig.colorbar(im, ax=ax)
        ax.set_xlabel("NTels")
        plt.setp(ax.get_yticklabels(), visible=False)
        plt.title("Number of MSTs")

        plt.subplots_adjust(wspace=0.05)

        # plot the energy migration matrix
        plt.figure()
        plt.hist2d(np.log10(reco_energy), np.log10(mc_energy), bins=20,
                   cmap=plt.cm.inferno)
        plt.xlabel("E_MC / TeV")
        plt.ylabel("E_rec / TeV")
        plt.colorbar()

        plt.show()
Esempio n. 9
0
    def fit_core_crosses(self):
        r"""calculates the core position as the least linear square solution
        of an (over-constrained) equation system

        Notes
        -----
        The basis is the "trace" of each telescope's `GreatCircle` which
        can be determined by the telescope's position P=(Px, Py) and
        the circle's normal vector, projected to the ground n=(nx,
        ny), so that for every r=(x, y) on the trace

        :math:`\vec n \cdot \vec r = \vec n \cdot \vec P` ,

        :math:`n_x \cdot x + n_y \cdot y = d`

        In a perfect world, the traces of all telescopes cross in the
        shower's point of impact. This means that there is one common
        point (x, y) for every telescope, so we can write in matrix
        form:

        .. math::
            :label: fullmatrix

            \begin{pmatrix}
                nx_1  &  ny_1  \\
                \vdots & \vdots \\
                nx_n  &  ny_n
            \end{pmatrix}
                \cdot (x, y) =
            \begin{pmatrix}
                d_1  \\
                \vdots \\
                d_n
            \end{pmatrix}



        or :math:`\boldsymbol{A} \cdot \vec r = \vec D` .

        Since we do not live in a perfect world and there probably is
        no point r that fulfils this equation system, it is solved by
        the method of least linear square:

        .. math::
            :label: rchisqr

            \vec{r}_{\chi^2} = (\boldsymbol{A}^\text{T} \cdot \boldsymbol{A})^{-1}
            \boldsymbol{A}^\text{T} \cdot \vec D


        :math:`\vec{r}_{\chi^2}` minimises the squared difference of


        .. math::

            \vec D - \boldsymbol{A} \cdot \vec r


        Weights are applied to every line of equation :eq:`fullmatrix`
        as stored in circle.weight (assuming they have been set in
        `get_great_circles` or elsewhere).

        Returns
        -------
        r_chisqr: numpy.ndarray(2)
            the minimum :math:`\chi^2` solution for the shower impact position
        pos_uncert: astropy length quantity
            error estimate on the reconstructed core position

        """

        A = np.zeros((len(self.circles), 2))
        D = np.zeros(len(self.circles))
        for i, circle in enumerate(self.circles.values()):
            # apply weight from circle and from the tilt of the circle
            # towards the horizontal plane: simply projecting
            # circle.norm to the ground gives higher weight to planes
            # perpendicular to the ground and less to those that have
            # a steeper angle
            A[i] = circle.weight * circle.norm[:2]
            # since A[i] is used in the dot-product, no need to multiply the
            # weight here
            D[i] = np.dot(A[i], circle.pos[:2])

        # the math from equation (2) would look like this:
        # ATA = np.dot(A.T, A)
        # ATAinv = np.linalg.inv(ATA)
        # ATAinvAT = np.dot(ATAinv, A.T)
        # return np.dot(ATAinvAT, D) * unit

        # instead used directly the numpy implementation
        # speed is the same, just handles already "SingularMatrixError"
        pos = np.linalg.lstsq(A, D)[0] * u.m

        weighted_sum_dist = np.sum([np.dot(pos[:2] - c.pos[:2], c.norm[:2]) * c.weight
                                    for c in self.circles.values()]) * pos.unit
        norm_sum_dist = np.sum([c.weight * linalg.length(c.norm[:2])
                                for c in self.circles.values()])
        pos_uncert = abs(weighted_sum_dist / norm_sum_dist)

        return pos, pos_uncert
Esempio n. 10
0
        # loop that cleans and parametrises the images and performs the reconstruction
        for (event, hillas_dict, n_tels, tot_signal, max_signals, pos_fit,
             dir_fit, h_max, err_est_pos,
             err_est_dir) in preper.prepare_event(source):

            # now prepare the features for the classifier
            cls_features_evt = {}
            reg_features_evt = {}
            for tel_id in hillas_dict.keys():
                Imagecutflow.count("pre-features")

                tel_pos = np.array(event.inst.tel_pos[tel_id][:2]) * u.m

                moments = hillas_dict[tel_id]

                impact_dist = linalg.length(tel_pos - pos_fit)

                reg_features_tel = EnergyFeatures(
                    impact_dist=(impact_dist / u.m).si,
                    sum_signal_evt=tot_signal,
                    max_signal_cam=max_signals[tel_id],
                    sum_signal_cam=moments.size,
                    N_LST=n_tels["LST"],
                    N_MST=n_tels["MST"],
                    N_SST=n_tels["SST"],
                    width=(moments.width / u.m).si,
                    length=(moments.length / u.m).si,
                    skewness=moments.skewness,
                    kurtosis=moments.kurtosis,
                    h_max=(h_max / u.m).si,
                    err_est_pos=(err_est_pos / u.m).si,