Запуск скрипта под профилировщиком, когда сам скрипт импортирует множество других модулей, один из которых очень медленный из-за известной функции - PullRequest
Новая
       20

Запуск скрипта под профилировщиком, когда сам скрипт импортирует множество других модулей, один из которых очень медленный из-за известной функции

0 голосов
/ 02 ноября 2019

Вот компоновка симуляций, которые я запускаю 

----main directory
                  -----output (directory)
                  -----halo (directory)
                  -----my_script.py
                  -----settings_centroid.py
                  -----simulation (directory)
                        -----halo_dark (directory)
                        -----halo_analysis (directory)
                        -----gizmo (directory)
                                     -----gizmo_plot.py
                                     .
                                     .
                                     .

My my_script.py (работает в главном каталоге): 

.
.
.
from simulation import gizmo
import settings_centroid
settings_centroid.init()
.
.
.
os.system('> output/{}/Info/{}/{}/redshift_{:.3f}/all_subhalo_properties_gas.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.rotation_status, settings_centroid.redshift_z))
.
.
.
gizmo.plot.Image.plot_image(...)

Мой settings_centroid.py скрипт:

.
.
.
def init():
    global ....    
.
.
.

Мой gizmo_plot.py: 

.
.
.
class ImageClass(ut.io.SayClass):

    def plot_image():

        dimen_label = {0: 'x', 1: 'y', 2: 'z'}

        if dimensions_select is None or not len(dimensions_select):
            dimensions_select = dimensions_plot

        if np.isscalar(distances_max):
            distances_max = [distances_max for dimen_i in
                             range(part[species_name]['position'].shape[1])]
        distances_max = np.array(distances_max, dtype=np.float64)

        position_limits = []
        for dimen_i in range(distances_max.shape[0]):
            position_limits.append([-distances_max[dimen_i], distances_max[dimen_i]])
        position_limits = np.array(position_limits)

        if part_indices is None or not len(part_indices):
            part_indices = ut.array.get_arange(part[species_name]['position'].shape[0])

        if property_select:
            part_indices = ut.catalog.get_indices_catalog(
                part[species_name], property_select, part_indices)

        if subsample_factor is not None and subsample_factor > 1:
            part_indices = part_indices[::subsample_factor]

        positions = np.array(part[species_name]['position'][part_indices])
        mass_array = np.array(part[species_name]['mass'][part_indices])                     
        velocity_array = np.array(part[species_name]['velocity'][part_indices])            
        if species_name == 'gas':
            HI_fraction_array = np.array(part[species_name]['hydrogen.neutral.fraction'])   

        weights = None
        if weight_name:
            weights = part[species_name].prop(weight_name, part_indices)

        center_position = ut.particle.parse_property(part, 'center_position', center_position)

        if center_position is not None and len(center_position):
            # re-orient to input center
            positions -= center_position
            positions *= part.snapshot['scalefactor']

            if rotation is not None:
                # rotate image
                if rotation is True:
                    # rotate according to principal axes
                    if (len(part[species_name].host_rotation_tensors) and
                            len(part[species_name].host_rotation_tensors[0])):
                        # rotate to align with stored principal axes
                        rotation_tensor = part[species_name].host_rotation_tensors[0]
                    else:
                        # compute principal axes using all particles originally within image limits
                        masks = (positions[:, dimensions_select[0]] <= distances_max[0])
                        for dimen_i in dimensions_select:
                            masks *= (
                                (positions[:, dimen_i] >= -distances_max[dimen_i]) *
                                (positions[:, dimen_i] <= distances_max[dimen_i])
                            )
                        rotation_tensor = ut.coordinate.get_principal_axes(
                            positions[masks], weights[masks])[0]
                elif len(rotation):
                    # use input rotation vectors
                    rotation_tensor = np.asarray(rotation)
                    if (np.ndim(rotation_tensor) != 2 or
                            rotation_tensor.shape[0] != positions.shape[1] or
                            rotation_tensor.shape[1] != positions.shape[1]):
                        raise ValueError('wrong shape for rotation = {}'.format(rotation))
                else:
                    raise ValueError('cannot parse rotation = {}'.format(rotation))

                positions = ut.coordinate.get_coordinates_rotated(positions, rotation_tensor)

            # keep only particles within distance limits and with speeds less than 500km/sec compared to the parent halo.
            masks1 = (positions[:, dimensions_select[0]] <= distances_max[0])   #part[species_name]['position'][part_indices]
            for dimen_i in dimensions_select:
                masks2 = (np.abs(part[species_name]['velocity'][:, dimen_i] - settings_centroid.HCV[dimen_i]) < 500)
                masks_part = masks1 * masks2
                masks_part *= (
                              (positions[:, dimen_i] >= -distances_max[dimen_i]) *
                              (positions[:, dimen_i] <= distances_max[dimen_i])
                              )

            positions = positions[masks_part]
            mass_array = mass_array[masks_part]
            velocity_array = velocity_array[masks_part]
            if species_name == 'gas':
                HI_fraction_array = HI_fraction_array[masks_part]
            else:
                HI_fraction_array = None


            if weights is not None:
                weights = weights[masks_part]
        else:
            raise ValueError('need to input center position')

        if distance_bin_width is not None and distance_bin_width > 0:
            position_bin_number = int(
                np.round(2 * np.max(distances_max[dimensions_plot]) / distance_bin_width))
        elif distance_bin_number is not None and distance_bin_number > 0:
            position_bin_number = 2 * distance_bin_number
        else:
            raise ValueError('need to input either distance bin width or bin number')

        #radiuss_array, positions_array, masss_array = [], [], []
        if hal is not None:
            # compile halos
            if hal_indices is None or not len(hal_indices):
                hal_indices = ut.array.get_arange(hal['mass.200m'])

            if 0 not in hal_indices:
                hal_indices = np.concatenate([[0], hal_indices])

            hal_positions = np.array(hal[hal_position_kind][hal_indices])
            if center_position is not None and len(center_position):
                hal_positions -= center_position
            hal_positions *= hal.snapshot['scalefactor']
            hal_radiuss = hal[hal_radius_kind][hal_indices]
            hal_masss = hal['mass.200m'][hal_indices]                                   
            hal_ids = hal['id'][hal_indices]                                            
            hal_distances = np.linalg.norm(hal['host.distance'], axis=1)[hal_indices]   
            hal_star_masss = hal['star.mass'][hal_indices]                      
            hal_star_sizes = hal['star.radius.90'][hal_indices]                 

            # initialize masks

            masks = (hal_positions[:, dimensions_select[0]] <= distances_max[0])
            for dimen_i in dimensions_select:
                masks *= (
                    (hal_positions[:, dimen_i] >= -distances_max[dimen_i]) *
                    (hal_positions[:, dimen_i] <= distances_max[dimen_i])
                )

            hal_radiuss = hal_radiuss[masks]
            hal_positions = hal_positions[masks]
            hal_masss = hal_masss[masks]                         
            hal_ids = hal_ids[masks]                             
            hal_distances = hal_distances[masks]            
            hal_star_masss = hal_star_masss[masks]                
            hal_star_sizes = hal_star_sizes[masks]               

            halo_ids = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_ids.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_masses = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_masses.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_radii = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_radii.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_positions = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_positions.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_distances = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_distances.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_star_masss = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_star_masses.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)
            halo_star_sizes = 'output/{}/Info/{}/halo_catalog_{:.3f}/halo_star_sizes.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)

            np.savetxt(halo_ids, hal_ids, fmt='%.0f')
            np.savetxt(halo_masses, hal_masss, fmt='%.3e')
            np.savetxt(halo_radii, hal_radiuss, fmt='%.3e')
            np.savetxt(halo_positions, hal_positions, fmt='%.3e')
            np.savetxt(halo_distances, hal_distances, fmt='%.3e')
            np.savetxt(halo_star_masss, hal_star_masss, fmt='%.3e')
            np.savetxt(halo_star_sizes, hal_star_sizes, fmt='%.3e')

    def get_histogram(...):
        if '3d' in image_kind:
            # calculate maximum local density along projected dimension
            hist_valuess, (hist_xs, hist_ys, hist_zs) = np.histogramdd(positions, position_bin_number, position_limits, weights=weights, normed=False,)

            # convert to 3-d density
            hist_valuess /= (np.diff(hist_xs)[0] * np.diff(hist_ys)[0] * np.diff(hist_zs)[0])

        else:
            # project along single dimension
            hist_valuess, hist_xs, hist_ys = np.histogram2d(positions[:, dimensions_plot[0]], positions[:, dimensions_plot[1]], position_bin_number, position_limits[dimensions_plot], weights=weights, normed=False,)

            # convert to surface density
            hist_valuess /= np.diff(hist_xs)[0] * np.diff(hist_ys)[0]

            # convert to number density
           if use_column_units:
               hist_valuess *= ut.basic.constant.hydrogen_per_sun * ut.basic.constant.kpc_per_cm ** 2
               lls_number = np.sum((hist_valuess > 1e17) * (hist_valuess < 2e20))
               dla_number = np.sum(hist_valuess > 2e20)

               LLS, DLA = lls_number, dla_number
               self.say('Number of grids:      LLS = {:.0f}, \t DLA = {:.0f}'.format(lls_number, dla_number))


               # Counting absorber grid number in each subhalo
               if return_halo_info:
                   subhalos_gas = 'output/{}/Info/{}/{}/redshift_{:.3f}/all_subhalo_properties_gas.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.rotation_status, settings_centroid.redshift_z)

                   hal_positions_data = np.loadtxt(r'output/{}/Info/{}/halo_catalog_{:.3f}/halo_positions.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))
                   hal_radiuss_data = np.loadtxt(r'output/{}/Info/{}/halo_catalog_{:.3f}/halo_radii.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))
                   hal_masss_data = np.loadtxt(r'output/{}/Info/{}/halo_catalog_{:.3f}/halo_masses.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))
                   hal_ids_data = np.loadtxt(r'output/{}/Info/{}/halo_catalog_{:.3f}/halo_ids.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))
                   hal_distances_data = np.loadtxt(r'output/{}/Info/{}/halo_catalog_{:.3f}/halo_distances.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))

                   totals_gas = {}
                   sub_circle_catalog_gas  = []
                   enclosing_circles_gas = {}
                   for hal_id, hal_position, hal_radius, hal_mass, hal_distance in zip(hal_ids_data, hal_positions_data, hal_radiuss_data, hal_masss_data, hal_distances_data): 
                       if ((hal_distance <= settings_centroid.distance_max) and (log10(hal_mass) >= settings_centroid.low_mass_cutoff)):
                           hal_gas_mass = sum(settings_centroid.part_HI_mass[i]*settings_centroid.part_HI_fraction[i] for i in np.where(settings_centroid.part_HI_fraction > 0)[0] if (np.linalg.norm(settings_centroid.part_HI_position[i] - hal_position) * settings_centroid.scale_factor <= hal_radius))

                           if (hal_gas_mass/hal_mass > 1.0e-8):
                               enclosing_circles_gas[hal_id] = float(settings_centroid.trunc_digits(log10(hal_mass), 4))

                   # choose all subhalos' IDs enclosing the DLA pixel
                   enclosing_circles = list(enclosing_circles_gas.keys())
                   sub_circle_catalog_gas += [(enclosing_circles_gas[i], 1) for i in enclosing_circles]

                   # add up all special grids in each sub-circle when looping over all grids
                   for key, value in sub_circle_catalog_gas:
                       totals_gas[key] = totals_gas.get(key, 0) + value
                   totals_gas = collections.OrderedDict(sorted(totals_gas.items()))
                   totals_gas = list(totals_gas.items())

                   with open(subhalos_gas, "a") as smallest_local_subhalos:
                       print('{}'.format(totals_gas), file=smallest_local_subhalos)
                       smallest_local_subhalos.close()

                   os.system('> output/{}/Info/{}/halo_catalog_{:.3f}/halo_positions.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))
                   os.system('> output/{}/Info/{}/halo_catalog_{:.3f}/halo_radii.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)) 
                   os.system('> output/{}/Info/{}/halo_catalog_{:.3f}/halo_masses.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z)) 
                   os.system('> output/{}/Info/{}/halo_catalog_{:.3f}/halo_ids.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))  
                   os.system('> output/{}/Info/{}/halo_catalog_{:.3f}/halo_distances.txt'.format(settings_centroid.halo_size, settings_centroid.halo_name, settings_centroid.redshift_z))


       masks = (hist_valuess > 0)
       self.say('histogram min, med, max = {:.3e}, {:.3e}, {:.3e}'.format(hist_valuess[masks].min(), np.median(hist_valuess[masks]), hist_valuess[masks].max()))
       hist_limits = np.array([hist_valuess[masks].min(), hist_valuess[masks].max()])
       return  hist_valuess, hist_xs, hist_ys, hist_limits
.
.
.

Кажется, что при запуске my_script.py на получение результатов уходит целая вечность. Методом проб и ошибок кажется, что медленная часть происходит внутри if ... if вложенного цикла под функцией, называемой get_histogram() внутри gizmo_plot.py module. Но мне нужно запустить my_script.py под профилировщиком, чтобы точно найти медленную линию. Не могли бы вы помочь мне, как это можно сделать путем вывода профиля в другой файл? В частности, как найти именно медленную линию функции?

Вывод процедуры, предложенной Уилксом ниже: 

my_script.prof% sort cumulative
my_script.prof% stats 5
Mon Nov  4 14:32:25 2019    my_script.prof

         76741270081 function calls (76741240862 primitive calls) in 107707.564 seconds

   Ordered by: cumulative time
   List reduced from 4432 to 5 due to restriction <5>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
    814/1    0.373    0.000 107707.595 107707.595 {built-in method builtins.exec}
        1    0.348    0.348 107707.331 107707.331 my_script.py:1(<module>)
        2  312.439  156.219 105178.281 52589.140 gizmo_plot.py:220(plot_image)
        2    5.966    2.983 104209.307 52104.654 gizmo_plot.py:658(get_histogram)
       69    0.001    0.000 104168.283 1509.685 {built-in method builtins.sum}

Вторая проверка: 

my_script.prof% sort time
my_script.prof% stats 10
Mon Nov  4 14:32:25 2019    my_script.prof

         76741270081 function calls (76741240862 primitive calls) in 107707.564 seconds

   Ordered by: internal time
   List reduced from 4432 to 10 due to restriction <10>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
9592513999 41813.829    0.000 80076.700    0.000 linalg.py:2203(norm)
       69 24091.619  349.154 104168.282 1509.685 gizmo_plot.py:726(<genexpr>)
9592514405 9783.770    0.000 9783.770    0.000 {built-in method numpy.core.multiarray.dot}
9592514634 8158.522    0.000 11181.488    0.000 numeric.py:433(asarray)
9592514062 7065.503    0.000 7065.503    0.000 {method 'ravel' of 'numpy.ndarray' objects}
9592513998 5393.512    0.000 7708.295    0.000 linalg.py:113(isComplexType)
19185030173/19185030020 4839.445    0.000 4839.831    0.000 {built-in method builtins.issubclass}
9592517050 3069.880    0.000 3072.494    0.000 {built-in method numpy.core.multiarray.array}
     11/3  605.531   55.048  605.532  201.844 gizmo_io.py:190(prop)
      376  530.257    1.410  530.655    1.411 dataset.py:634(read_direct)

1 Ответ

1 голос
/ 03 ноября 2019

Запустите профилировщик как python3 -m cProfile -o my_script.prof my_script.py. Это должно создать my_script.prof после завершения вашего скрипта. Вы можете загрузить этот файл .prof с помощью python3 -m pstats my_script.prof.

Добро пожаловать на сайт PullRequest, где вы можете задавать вопросы и получать ответы от других членов сообщества.

Похожие темы

...