def _plot_2d_radar_cam(colour_map_object,
max_colour_percentile,
figure_objects,
axes_object_matrices,
model_metadata_dict,
output_dir_name,
cam_matrix=None,
guided_cam_matrix=None,
significance_matrix=None,
full_storm_id_string=None,
storm_time_unix_sec=None):
"""Plots guided or unguided class-activation map for 2-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
C = number of radar channels
:param colour_map_object: See doc for `_plot_3d_radar_cam`.
:param max_colour_percentile: Same.
:param figure_objects: Same.
:param axes_object_matrices: Same.
:param model_metadata_dict: Same.
:param output_dir_name: Same.
:param cam_matrix: M-by-N numpy array of class activations.
:param guided_cam_matrix: M-by-N-by-C numpy array of guided-CAM output.
:param significance_matrix: See doc for `_plot_3d_radar_cam`.
:param full_storm_id_string: Same.
:param storm_time_unix_sec: Same.
"""
if cam_matrix is None:
quantity_string = 'max abs value'
else:
quantity_string = 'max activation'
pmm_flag = full_storm_id_string is None and storm_time_unix_sec is None
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
figure_index = 1 if conv_2d3d else 0
list_of_layer_operation_dicts = model_metadata_dict[
cnn.LAYER_OPERATIONS_KEY]
if list_of_layer_operation_dicts is None:
training_option_dict = model_metadata_dict[
cnn.TRAINING_OPTION_DICT_KEY]
radar_field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
num_channels = len(radar_field_names)
else:
num_channels = len(list_of_layer_operation_dicts)
if cam_matrix is None:
this_matrix = guided_cam_matrix
else:
this_matrix = numpy.expand_dims(cam_matrix, axis=-1)
this_matrix = numpy.repeat(this_matrix, repeats=num_channels, axis=-1)
max_contour_level = numpy.percentile(numpy.absolute(this_matrix),
max_colour_percentile)
if cam_matrix is None:
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_contour_level,
contour_interval=max_contour_level / HALF_NUM_CONTOURS,
row_major=False)
else:
cam_plotting.plot_many_2d_grids(
class_activation_matrix_3d=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
colour_map_object=colour_map_object,
max_contour_level=max_contour_level,
contour_interval=max_contour_level / NUM_CONTOURS,
row_major=False)
if significance_matrix is not None:
if cam_matrix is None:
this_matrix = significance_matrix
else:
this_matrix = numpy.expand_dims(significance_matrix, axis=-1)
this_matrix = numpy.repeat(this_matrix,
repeats=num_channels,
axis=-1)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
row_major=False)
this_title_string = figure_objects[figure_index]._suptitle
if this_title_string is not None:
this_title_string += ' ({0:s} = {1:.2e})'.format(
quantity_string, max_contour_level)
figure_objects[figure_index].suptitle(
this_title_string, fontsize=plot_input_examples.TITLE_FONT_SIZE)
output_file_name = plot_input_examples.metadata_to_file_name(
output_dir_name=output_dir_name,
is_sounding=False,
pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_unix_sec=storm_time_unix_sec,
radar_field_name='shear' if conv_2d3d else None)
print('Saving figure to: "{0:s}"...'.format(output_file_name))
figure_objects[figure_index].savefig(output_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(figure_objects[figure_index])
def _plot_2d_radar_saliency(
saliency_matrix, colour_map_object, max_colour_value, figure_objects,
axes_object_matrices, model_metadata_dict, output_dir_name,
significance_matrix=None, full_storm_id_string=None,
storm_time_unix_sec=None):
"""Plots saliency map for 2-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
C = number of radar channels
If this method is plotting a composite rather than single example (storm
object), `full_storm_id_string` and `storm_time_unix_sec` can be None.
:param saliency_matrix: M-by-N-by-C numpy array of saliency values.
:param colour_map_object: See documentation at top of file.
:param max_colour_value: Max value in colour scheme for saliency.
:param figure_objects: See doc for
`plot_input_examples._plot_2d_radar_scan`.
:param axes_object_matrices: Same.
:param model_metadata_dict: Dictionary returned by
`cnn.read_model_metadata`.
:param output_dir_name: Path to output directory. Figure(s) will be saved
here.
:param significance_matrix: M-by-N-by-H numpy array of Boolean flags,
indicating where differences with some other saliency map are
significant.
:param full_storm_id_string: Full storm ID.
:param storm_time_unix_sec: Storm time.
"""
pmm_flag = full_storm_id_string is None and storm_time_unix_sec is None
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
if conv_2d3d:
figure_index = 1
radar_field_name = 'shear'
else:
figure_index = 0
radar_field_name = None
list_of_layer_operation_dicts = model_metadata_dict[
cnn.LAYER_OPERATIONS_KEY]
if list_of_layer_operation_dicts is not None:
saliency_matrix = saliency_matrix[
..., plot_input_examples.LAYER_OP_INDICES_TO_KEEP
]
if significance_matrix is not None:
significance_matrix = significance_matrix[
..., plot_input_examples.LAYER_OP_INDICES_TO_KEEP
]
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(saliency_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / HALF_NUM_CONTOURS,
row_major=False)
if significance_matrix is not None:
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(significance_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
row_major=False)
this_title_object = figure_objects[figure_index]._suptitle
if this_title_object is not None:
this_title_string = '{0:s} (max abs saliency = {1:.2e})'.format(
this_title_object.get_text(), max_colour_value
)
figure_objects[figure_index].suptitle(
this_title_string, fontsize=plot_input_examples.TITLE_FONT_SIZE)
output_file_name = plot_input_examples.metadata_to_file_name(
output_dir_name=output_dir_name, is_sounding=False, pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_unix_sec=storm_time_unix_sec,
radar_field_name=radar_field_name)
print('Saving figure to: "{0:s}"...'.format(output_file_name))
figure_objects[figure_index].savefig(
output_file_name, dpi=FIGURE_RESOLUTION_DPI, pad_inches=0,
bbox_inches='tight'
)
pyplot.close(figure_objects[figure_index])
def _plot_3d_radar_cam(colour_map_object,
max_colour_percentile,
figure_objects,
axes_object_matrices,
model_metadata_dict,
output_dir_name,
cam_matrix=None,
guided_cam_matrix=None,
significance_matrix=None,
full_storm_id_string=None,
storm_time_unix_sec=None):
"""Plots guided or unguided class-activation map for 3-D radar data.
This method will plot either `cam_matrix` or `guided_cam_matrix`, not both.
M = number of rows in spatial grid
N = number of columns in spatial grid
H = number of heights in spatial grid
F = number of radar fields
If this method is plotting a composite rather than single example (storm
object), `full_storm_id_string` and `storm_time_unix_sec` can be None.
:param colour_map_object: See documentation at top of file.
:param max_colour_percentile: Same.
:param figure_objects: See doc for `plot_input_examples._plot_3d_example`.
:param axes_object_matrices: Same.
:param model_metadata_dict: Dictionary returned by
`cnn.read_model_metadata`.
:param output_dir_name: Path to output directory. Figure(s) will be saved
here.
:param cam_matrix: M-by-N-by-H numpy array of class activations.
:param guided_cam_matrix: M-by-N-by-H-by-F numpy array of guided-CAM output.
:param significance_matrix: Boolean numpy array with the same dimensions as
the array being plotted (`cam_matrix` or `guided_cam_matrix`),
indicating where differences with some other CAM are significant.
:param full_storm_id_string: Full storm ID.
:param storm_time_unix_sec: Storm time.
"""
if cam_matrix is None:
quantity_string = 'max abs value'
else:
quantity_string = 'max activation'
pmm_flag = full_storm_id_string is None and storm_time_unix_sec is None
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
if conv_2d3d:
loop_max = 1
radar_field_names = ['reflectivity']
else:
loop_max = len(figure_objects)
training_option_dict = model_metadata_dict[
cnn.TRAINING_OPTION_DICT_KEY]
radar_field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
for j in range(loop_max):
if cam_matrix is None:
this_matrix = guided_cam_matrix[..., j]
else:
this_matrix = cam_matrix
this_max_contour_level = numpy.percentile(numpy.absolute(this_matrix),
max_colour_percentile)
if cam_matrix is None:
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[j],
colour_map_object=colour_map_object,
max_absolute_contour_level=this_max_contour_level,
contour_interval=this_max_contour_level / HALF_NUM_CONTOURS)
else:
cam_plotting.plot_many_2d_grids(
class_activation_matrix_3d=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[j],
colour_map_object=colour_map_object,
max_contour_level=this_max_contour_level,
contour_interval=this_max_contour_level / NUM_CONTOURS)
if significance_matrix is not None:
if cam_matrix is None:
this_matrix = significance_matrix[..., j]
else:
this_matrix = significance_matrix
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(this_matrix, axis=0),
axes_object_matrix=axes_object_matrices[j])
this_title_string = figure_objects[j]._suptitle
if this_title_string is not None:
this_title_string += ' ({0:s} = {1:.2e})'.format(
quantity_string, this_max_contour_level)
figure_objects[j].suptitle(
this_title_string,
fontsize=plot_input_examples.TITLE_FONT_SIZE)
this_file_name = plot_input_examples.metadata_to_file_name(
output_dir_name=output_dir_name,
is_sounding=False,
pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_unix_sec=storm_time_unix_sec,
radar_field_name=radar_field_names[j])
print('Saving figure to: "{0:s}"...'.format(this_file_name))
figure_objects[j].savefig(this_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(figure_objects[j])
def _plot_one_composite(saliency_file_name, monte_carlo_file_name,
composite_name_abbrev, composite_name_verbose,
colour_map_object, max_colour_value, half_num_contours,
smoothing_radius_grid_cells, output_dir_name):
"""Plots saliency map for one composite.
:param saliency_file_name: Path to saliency file (will be read by
`saliency.read_file`).
:param monte_carlo_file_name: Path to Monte Carlo file (will be read by
`_read_monte_carlo_file`).
:param composite_name_abbrev: Abbrev composite name (will be used in file
names).
:param composite_name_verbose: Verbose composite name (will be used in
figure title).
:param colour_map_object: See documentation at top of file.
:param max_colour_value: Same.
:param half_num_contours: Same.
:param smoothing_radius_grid_cells: Same.
:param output_dir_name: Name of output directory (figures will be saved
here).
:return: main_figure_file_name: Path to main image file created by this
method.
"""
(mean_radar_matrices, mean_saliency_matrices, significance_matrices,
model_metadata_dict) = _read_one_composite(
saliency_file_name=saliency_file_name,
smoothing_radius_grid_cells=smoothing_radius_grid_cells,
monte_carlo_file_name=monte_carlo_file_name)
refl_heights_m_agl = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY][
trainval_io.RADAR_HEIGHTS_KEY]
num_refl_heights = len(refl_heights_m_agl)
handle_dict = plot_examples.plot_one_example(
list_of_predictor_matrices=mean_radar_matrices,
model_metadata_dict=model_metadata_dict,
pmm_flag=True,
plot_sounding=False,
allow_whitespace=True,
plot_panel_names=True,
panel_name_font_size=PANEL_NAME_FONT_SIZE,
add_titles=False,
label_colour_bars=True,
colour_bar_length=COLOUR_BAR_LENGTH,
colour_bar_font_size=COLOUR_BAR_FONT_SIZE,
num_panel_rows=num_refl_heights)
axes_object_matrices = handle_dict[plot_examples.RADAR_AXES_KEY]
this_saliency_matrix = numpy.flip(mean_saliency_matrices[0][0, ..., 0],
axis=0)
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=this_saliency_matrix,
axes_object_matrix=axes_object_matrices[0],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / half_num_contours,
row_major=True)
this_sig_matrix = numpy.flip(significance_matrices[0][0, ..., 0], axis=0)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=this_sig_matrix,
axes_object_matrix=axes_object_matrices[0],
marker_size=2,
row_major=True)
this_saliency_matrix = numpy.flip(mean_saliency_matrices[1][0, ...],
axis=0)
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=this_saliency_matrix,
axes_object_matrix=axes_object_matrices[1],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / half_num_contours,
row_major=False)
this_sig_matrix = numpy.flip(significance_matrices[1][0, ...], axis=0)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=this_sig_matrix,
axes_object_matrix=axes_object_matrices[1],
marker_size=2,
row_major=False)
refl_figure_object = handle_dict[plot_examples.RADAR_FIGURES_KEY][0]
refl_figure_file_name = '{0:s}/{1:s}_reflectivity.jpg'.format(
output_dir_name, composite_name_abbrev)
print('Saving figure to: "{0:s}"...'.format(refl_figure_file_name))
refl_figure_object.savefig(refl_figure_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(refl_figure_object)
shear_figure_object = handle_dict[plot_examples.RADAR_FIGURES_KEY][1]
shear_figure_file_name = '{0:s}/{1:s}_shear.jpg'.format(
output_dir_name, composite_name_abbrev)
print('Saving figure to: "{0:s}"...'.format(shear_figure_file_name))
shear_figure_object.savefig(shear_figure_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(shear_figure_object)
main_figure_file_name = '{0:s}/{1:s}_saliency.jpg'.format(
output_dir_name, composite_name_abbrev)
print('Concatenating panels to: "{0:s}"...'.format(main_figure_file_name))
imagemagick_utils.concatenate_images(
input_file_names=[refl_figure_file_name, shear_figure_file_name],
output_file_name=main_figure_file_name,
num_panel_rows=1,
num_panel_columns=2,
border_width_pixels=50,
extra_args_string='-gravity south')
imagemagick_utils.resize_image(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
output_size_pixels=CONCAT_FIGURE_SIZE_PX)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=TITLE_FONT_SIZE + 25)
_overlay_text(image_file_name=main_figure_file_name,
x_offset_from_center_px=0,
y_offset_from_top_px=0,
text_string=composite_name_verbose)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=10)
return main_figure_file_name
def _plot_3d_radar_saliency(
saliency_matrix, colour_map_object, max_colour_value, figure_objects,
axes_object_matrices, model_metadata_dict, output_dir_name,
significance_matrix=None, full_storm_id_string=None,
storm_time_unix_sec=None):
"""Plots saliency map for 3-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
H = number of heights in spatial grid
F = number of radar fields
If this method is plotting a composite rather than single example (storm
object), `full_storm_id_string` and `storm_time_unix_sec` can be None.
:param saliency_matrix: M-by-N-by-H-by-F numpy array of saliency values.
:param colour_map_object: See documentation at top of file.
:param max_colour_value: Max value in colour scheme for saliency.
:param figure_objects: See doc for
`plot_input_examples._plot_3d_radar_scan`.
:param axes_object_matrices: Same.
:param model_metadata_dict: Dictionary returned by
`cnn.read_model_metadata`.
:param output_dir_name: Path to output directory. Figure(s) will be saved
here.
:param significance_matrix: M-by-N-by-H-by-F numpy array of Boolean flags,
indicating where differences with some other saliency map are
significant.
:param full_storm_id_string: Full storm ID.
:param storm_time_unix_sec: Storm time.
"""
pmm_flag = full_storm_id_string is None and storm_time_unix_sec is None
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
if conv_2d3d:
loop_max = 1
radar_field_names = ['reflectivity']
else:
loop_max = len(figure_objects)
training_option_dict = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY]
radar_field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
for j in range(loop_max):
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(saliency_matrix[..., j], axis=0),
axes_object_matrix=axes_object_matrices[j],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / HALF_NUM_CONTOURS)
if significance_matrix is not None:
this_matrix = numpy.flip(significance_matrix[..., j], axis=0)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=this_matrix,
axes_object_matrix=axes_object_matrices[j]
)
this_title_object = figure_objects[j]._suptitle
if this_title_object is not None:
this_title_string = '{0:s} (max abs saliency = {1:.2e})'.format(
this_title_object.get_text(), max_colour_value
)
figure_objects[j].suptitle(
this_title_string, fontsize=plot_input_examples.TITLE_FONT_SIZE)
this_file_name = plot_input_examples.metadata_to_file_name(
output_dir_name=output_dir_name, is_sounding=False,
pmm_flag=pmm_flag, full_storm_id_string=full_storm_id_string,
storm_time_unix_sec=storm_time_unix_sec,
radar_field_name=radar_field_names[j]
)
print('Saving figure to: "{0:s}"...'.format(this_file_name))
figure_objects[j].savefig(
this_file_name, dpi=FIGURE_RESOLUTION_DPI, pad_inches=0,
bbox_inches='tight'
)
pyplot.close(figure_objects[j])
def _plot_2d_radar_saliency(
saliency_matrix, colour_map_object, max_colour_value, half_num_contours,
label_colour_bars, colour_bar_length, figure_objects,
axes_object_matrices, model_metadata_dict, output_dir_name,
significance_matrix=None, full_storm_id_string=None,
storm_time_unix_sec=None):
"""Plots saliency map for 2-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
C = number of radar channels
If this method is plotting a composite rather than single example (storm
object), `full_storm_id_string` and `storm_time_unix_sec` can be None.
:param saliency_matrix: M-by-N-by-C numpy array of saliency values.
:param colour_map_object: See documentation at top of file.
:param max_colour_value: Same.
:param half_num_contours: Same.
:param label_colour_bars: Same.
:param colour_bar_length: Same.
:param figure_objects: See doc for
`plot_input_examples._plot_2d_radar_scan`.
:param axes_object_matrices: Same.
:param model_metadata_dict: Dictionary returned by
`cnn.read_model_metadata`.
:param output_dir_name: Path to output directory. Figure(s) will be saved
here.
:param significance_matrix: M-by-N-by-H numpy array of Boolean flags,
indicating where differences with some other saliency map are
significant.
:param full_storm_id_string: Full storm ID.
:param storm_time_unix_sec: Storm time.
"""
if max_colour_value is None:
max_colour_value = numpy.percentile(
numpy.absolute(saliency_matrix), MAX_COLOUR_PERCENTILE
)
pmm_flag = full_storm_id_string is None and storm_time_unix_sec is None
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
if conv_2d3d:
figure_index = 1
radar_field_name = 'shear'
else:
figure_index = 0
radar_field_name = None
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(saliency_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / half_num_contours,
row_major=False)
if significance_matrix is not None:
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(significance_matrix, axis=0),
axes_object_matrix=axes_object_matrices[figure_index],
row_major=False)
colour_bar_object = plotting_utils.plot_linear_colour_bar(
axes_object_or_matrix=axes_object_matrices[figure_index],
data_matrix=saliency_matrix,
colour_map_object=colour_map_object, min_value=0.,
max_value=max_colour_value, orientation_string='horizontal',
fraction_of_axis_length=colour_bar_length / (1 + int(conv_2d3d)),
extend_min=False, extend_max=True, font_size=COLOUR_BAR_FONT_SIZE)
if label_colour_bars:
colour_bar_object.set_label(
'Absolute saliency', fontsize=COLOUR_BAR_FONT_SIZE)
output_file_name = plot_examples.metadata_to_file_name(
output_dir_name=output_dir_name, is_sounding=False, pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_unix_sec=storm_time_unix_sec,
radar_field_name=radar_field_name)
print('Saving figure to: "{0:s}"...'.format(output_file_name))
figure_objects[figure_index].savefig(
output_file_name, dpi=FIGURE_RESOLUTION_DPI, pad_inches=0,
bbox_inches='tight'
)
pyplot.close(figure_objects[figure_index])
def _plot_one_composite(saliency_file_name, monte_carlo_file_name,
composite_name_abbrev, composite_name_verbose,
colour_map_object, max_colour_value, half_num_contours,
smoothing_radius_grid_cells, output_dir_name):
"""Plots saliency map for one composite.
:param saliency_file_name: Path to saliency file (will be read by
`saliency.read_file`).
:param monte_carlo_file_name: Path to Monte Carlo file (will be read by
`_read_monte_carlo_file`).
:param composite_name_abbrev: Abbrev composite name (will be used in file
names).
:param composite_name_verbose: Verbose composite name (will be used in
figure title).
:param colour_map_object: See documentation at top of file.
:param max_colour_value: Max value in colour bar (may be NaN).
:param half_num_contours: See documentation at top of file.
:param smoothing_radius_grid_cells: Same.
:param output_dir_name: Name of output directory (figures will be saved
here).
:return: main_figure_file_name: Path to main image file created by this
method.
:return: max_colour_value: Same as input but cannot be None.
"""
(mean_radar_matrix, mean_saliency_matrix, significance_matrix,
model_metadata_dict) = _read_one_composite(
saliency_file_name=saliency_file_name,
smoothing_radius_grid_cells=smoothing_radius_grid_cells,
monte_carlo_file_name=monte_carlo_file_name)
training_option_dict = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY]
field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
num_fields = mean_radar_matrix.shape[-1]
num_heights = mean_radar_matrix.shape[-2]
if numpy.isnan(max_colour_value):
max_colour_value = numpy.percentile(
numpy.absolute(mean_saliency_matrix[0, ...]), 99.)
these_indices = numpy.where(
max_colour_value <= POSSIBLE_MAX_COLOUR_VALUES)[0]
if len(these_indices) == 0:
this_index = -1
else:
this_index = these_indices[0]
this_index -= 1
max_colour_value = POSSIBLE_MAX_COLOUR_VALUES[this_index]
handle_dict = plot_examples.plot_one_example(
list_of_predictor_matrices=[mean_radar_matrix],
model_metadata_dict=model_metadata_dict,
pmm_flag=True,
allow_whitespace=True,
plot_panel_names=True,
panel_name_font_size=PANEL_NAME_FONT_SIZE,
add_titles=False,
label_colour_bars=True,
colour_bar_length=COLOUR_BAR_LENGTH,
colour_bar_font_size=COLOUR_BAR_FONT_SIZE,
num_panel_rows=num_heights)
figure_objects = handle_dict[plot_examples.RADAR_FIGURES_KEY]
axes_object_matrices = handle_dict[plot_examples.RADAR_AXES_KEY]
for k in range(num_fields):
this_saliency_matrix = mean_saliency_matrix[0, ..., k]
saliency_plotting.plot_many_2d_grids_with_contours(
saliency_matrix_3d=numpy.flip(this_saliency_matrix, axis=0),
axes_object_matrix=axes_object_matrices[k],
colour_map_object=colour_map_object,
max_absolute_contour_level=max_colour_value,
contour_interval=max_colour_value / half_num_contours)
this_sig_matrix = significance_matrix[0, ..., k]
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(this_sig_matrix, axis=0),
axes_object_matrix=axes_object_matrices[k])
panel_file_names = [None] * num_fields
for k in range(num_fields):
panel_file_names[k] = '{0:s}/{1:s}_{2:s}.jpg'.format(
output_dir_name, composite_name_abbrev,
field_names[k].replace('_', '-'))
print('Saving figure to: "{0:s}"...'.format(panel_file_names[k]))
figure_objects[k].savefig(panel_file_names[k],
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(figure_objects[k])
main_figure_file_name = '{0:s}/{1:s}_saliency.jpg'.format(
output_dir_name, composite_name_abbrev)
print('Concatenating panels to: "{0:s}"...'.format(main_figure_file_name))
imagemagick_utils.concatenate_images(
input_file_names=panel_file_names,
output_file_name=main_figure_file_name,
num_panel_rows=1,
num_panel_columns=num_fields,
border_width_pixels=50)
imagemagick_utils.resize_image(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
output_size_pixels=CONCAT_FIGURE_SIZE_PX)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=TITLE_FONT_SIZE + 25)
_overlay_text(image_file_name=main_figure_file_name,
x_offset_from_center_px=0,
y_offset_from_top_px=0,
text_string=composite_name_verbose)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=10)
return main_figure_file_name, max_colour_value
def _plot_one_composite(gradcam_file_name, monte_carlo_file_name,
composite_name_abbrev, composite_name_verbose,
colour_map_object, min_colour_value, max_colour_value,
num_contours, smoothing_radius_grid_cells,
monte_carlo_max_fdr, output_dir_name):
"""Plots class-activation map for one composite.
:param gradcam_file_name: Path to input file (will be read by
`gradcam.read_file`).
:param monte_carlo_file_name: Path to Monte Carlo file (will be read by
`_read_monte_carlo_file`).
:param composite_name_abbrev: Abbrev composite name (will be used in file
names).
:param composite_name_verbose: Verbose composite name (will be used in
figure title).
:param colour_map_object: See documentation at top of file.
:param min_colour_value: Minimum value in colour bar (may be NaN).
:param max_colour_value: Max value in colour bar (may be NaN).
:param num_contours: See documentation at top of file.
:param smoothing_radius_grid_cells: Same.
:param monte_carlo_max_fdr: Same.
:param output_dir_name: Name of output directory (figures will be saved
here).
:return: main_figure_file_name: Path to main image file created by this
method.
:return: min_colour_value: Same as input but cannot be None.
:return: max_colour_value: Same as input but cannot be None.
"""
(mean_radar_matrix, mean_class_activn_matrix, significance_matrix,
model_metadata_dict) = _read_one_composite(
gradcam_file_name=gradcam_file_name,
smoothing_radius_grid_cells=smoothing_radius_grid_cells,
monte_carlo_file_name=monte_carlo_file_name,
monte_carlo_max_fdr=monte_carlo_max_fdr)
if numpy.isnan(min_colour_value) or numpy.isnan(max_colour_value):
min_colour_value_log10 = numpy.log10(
numpy.percentile(mean_class_activn_matrix, 1.))
max_colour_value_log10 = numpy.log10(
numpy.percentile(mean_class_activn_matrix, 99.))
min_colour_value_log10 = max([min_colour_value_log10, -2.])
max_colour_value_log10 = max([max_colour_value_log10, -1.])
min_colour_value_log10 = min([min_colour_value_log10, 1.])
max_colour_value_log10 = min([max_colour_value_log10, 2.])
min_colour_value = 10**min_colour_value_log10
max_colour_value = 10**max_colour_value_log10
else:
min_colour_value_log10 = numpy.log10(min_colour_value)
max_colour_value_log10 = numpy.log10(max_colour_value)
contour_interval_log10 = (
(max_colour_value_log10 - min_colour_value_log10) / (num_contours - 1))
mean_activn_matrix_log10 = numpy.log10(mean_class_activn_matrix)
training_option_dict = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY]
field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
num_fields = mean_radar_matrix.shape[-1]
num_heights = mean_radar_matrix.shape[-2]
handle_dict = plot_examples.plot_one_example(
list_of_predictor_matrices=[mean_radar_matrix],
model_metadata_dict=model_metadata_dict,
pmm_flag=True,
allow_whitespace=True,
plot_panel_names=True,
panel_name_font_size=PANEL_NAME_FONT_SIZE,
add_titles=False,
label_colour_bars=True,
colour_bar_length=COLOUR_BAR_LENGTH,
colour_bar_font_size=COLOUR_BAR_FONT_SIZE,
num_panel_rows=num_heights)
figure_objects = handle_dict[plot_examples.RADAR_FIGURES_KEY]
axes_object_matrices = handle_dict[plot_examples.RADAR_AXES_KEY]
for k in range(num_fields):
cam_plotting.plot_many_2d_grids(
class_activation_matrix_3d=numpy.flip(
mean_activn_matrix_log10[0, ...], axis=0),
axes_object_matrix=axes_object_matrices[k],
colour_map_object=colour_map_object,
min_contour_level=min_colour_value_log10,
max_contour_level=max_colour_value_log10,
contour_interval=contour_interval_log10)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(significance_matrix[0, ...],
axis=0),
axes_object_matrix=axes_object_matrices[k])
panel_file_names = [None] * num_fields
for k in range(num_fields):
panel_file_names[k] = '{0:s}/{1:s}_{2:s}.jpg'.format(
output_dir_name, composite_name_abbrev,
field_names[k].replace('_', '-'))
print('Saving figure to: "{0:s}"...'.format(panel_file_names[k]))
figure_objects[k].savefig(panel_file_names[k],
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(figure_objects[k])
main_figure_file_name = '{0:s}/{1:s}_gradcam.jpg'.format(
output_dir_name, composite_name_abbrev)
print('Concatenating panels to: "{0:s}"...'.format(main_figure_file_name))
imagemagick_utils.concatenate_images(
input_file_names=panel_file_names,
output_file_name=main_figure_file_name,
num_panel_rows=1,
num_panel_columns=num_fields,
border_width_pixels=50)
imagemagick_utils.resize_image(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
output_size_pixels=CONCAT_FIGURE_SIZE_PX)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=TITLE_FONT_SIZE + 25)
_overlay_text(image_file_name=main_figure_file_name,
x_offset_from_center_px=0,
y_offset_from_top_px=0,
text_string=composite_name_verbose)
imagemagick_utils.trim_whitespace(input_file_name=main_figure_file_name,
output_file_name=main_figure_file_name,
border_width_pixels=10)
return main_figure_file_name, min_colour_value, max_colour_value
def _plot_3d_radar_difference(difference_matrix,
colour_map_object,
max_colour_percentile,
model_metadata_dict,
backwards_opt_dict,
output_dir_name,
example_index=None,
significance_matrix=None):
"""Plots difference (after minus before optimization) for 3-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
H = number of heights in spatial grid
F = number of fields
:param difference_matrix: M-by-N-by-H-by-F numpy array of differences (after
minus before optimization).
:param colour_map_object: See documentation at top of file.
:param max_colour_percentile: Same.
:param model_metadata_dict: Dictionary returned by
`cnn.read_model_metadata`.
:param backwards_opt_dict: Dictionary returned by
`backwards_optimization.read_standard_file` or
`backwards_optimization.read_pmm_file`, containing metadata.
:param output_dir_name: Name of output directory. Figure(s) will be saved
here.
:param example_index: This method will plot only the [i]th example, where
i = `example_index`. This will be used to find metadata for the given
example in `backwards_opt_dict`. If `backwards_opt_dict` contains PMM
(probability-matched means), leave this argument alone.
:param significance_matrix: M-by-N-by-H-by-F numpy array of Boolean flags,
indicating where these differences are significantly different than
differences from another backwards optimization.
"""
training_option_dict = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY]
radar_heights_m_agl = training_option_dict[trainval_io.RADAR_HEIGHTS_KEY]
num_heights = len(radar_heights_m_agl)
num_panel_rows = int(numpy.floor(numpy.sqrt(num_heights)))
pmm_flag = backwards_opt.MEAN_FINAL_ACTIVATION_KEY in backwards_opt_dict
if pmm_flag:
initial_activation = backwards_opt_dict[
backwards_opt.MEAN_INITIAL_ACTIVATION_KEY]
final_activation = backwards_opt_dict[
backwards_opt.MEAN_FINAL_ACTIVATION_KEY]
full_storm_id_string = None
storm_time_string = None
else:
initial_activation = backwards_opt_dict[
backwards_opt.INITIAL_ACTIVATIONS_KEY][example_index]
final_activation = backwards_opt_dict[
backwards_opt.FINAL_ACTIVATIONS_KEY][example_index]
full_storm_id_string = backwards_opt_dict[
backwards_opt.FULL_IDS_KEY][example_index]
storm_time_string = time_conversion.unix_sec_to_string(
backwards_opt_dict[backwards_opt.STORM_TIMES_KEY][example_index],
plot_input_examples.TIME_FORMAT)
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
if conv_2d3d:
radar_field_names = [radar_utils.REFL_NAME]
else:
radar_field_names = training_option_dict[trainval_io.RADAR_FIELDS_KEY]
num_fields = len(radar_field_names)
for j in range(num_fields):
this_max_colour_value = numpy.percentile(
numpy.absolute(difference_matrix[..., j]), max_colour_percentile)
this_colour_norm_object = matplotlib.colors.Normalize(
vmin=-1 * this_max_colour_value,
vmax=this_max_colour_value,
clip=False)
# TODO(thunderhoser): Deal with change of units.
this_figure_object, this_axes_object_matrix = (
radar_plotting.plot_3d_grid_without_coords(
field_matrix=numpy.flip(difference_matrix[..., j], axis=0),
field_name=radar_field_names[j],
grid_point_heights_metres=radar_heights_m_agl,
ground_relative=True,
num_panel_rows=num_panel_rows,
font_size=FONT_SIZE_SANS_COLOUR_BARS,
colour_map_object=colour_map_object,
colour_norm_object=this_colour_norm_object))
if significance_matrix is not None:
this_matrix = numpy.flip(significance_matrix[..., j], axis=0)
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=this_matrix,
axes_object_matrix=this_axes_object_matrix)
plotting_utils.plot_colour_bar(
axes_object_or_matrix=this_axes_object_matrix,
data_matrix=difference_matrix[..., j],
colour_map_object=colour_map_object,
colour_norm_object=this_colour_norm_object,
orientation_string='horizontal',
extend_min=True,
extend_max=True)
if pmm_flag:
this_title_string = 'PMM'
else:
this_title_string = 'Storm "{0:s}" at {1:s}'.format(
full_storm_id_string, storm_time_string)
this_title_string += (
'; {0:s}; activation from {1:.2e} to {2:.2e}').format(
radar_field_names[j], initial_activation, final_activation)
this_figure_object.suptitle(this_title_string,
fontsize=TITLE_FONT_SIZE)
this_file_name = plot_input_examples.metadata_to_radar_fig_file_name(
output_dir_name=output_dir_name,
pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_string=storm_time_string,
radar_field_name=radar_field_names[j])
print('Saving figure to: "{0:s}"...'.format(this_file_name))
this_figure_object.savefig(this_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(this_figure_object)
def _plot_2d_radar_difference(difference_matrix,
colour_map_object,
max_colour_percentile,
model_metadata_dict,
backwards_opt_dict,
output_dir_name,
example_index=None,
significance_matrix=None):
"""Plots difference (after minus before optimization) for 2-D radar data.
M = number of rows in spatial grid
N = number of columns in spatial grid
C = number of channels
:param difference_matrix: M-by-N-by-C numpy array of differences (after
minus before optimization).
:param colour_map_object: See doc for `_plot_3d_radar_difference`.
:param max_colour_percentile: Same.
:param model_metadata_dict: Same.
:param backwards_opt_dict: Same.
:param output_dir_name: Same.
:param example_index: Same.
:param significance_matrix: M-by-N-by-C numpy array of Boolean flags,
indicating where these differences are significantly different than
differences from another backwards optimization.
"""
pmm_flag = backwards_opt.MEAN_FINAL_ACTIVATION_KEY in backwards_opt_dict
if pmm_flag:
initial_activation = backwards_opt_dict[
backwards_opt.MEAN_INITIAL_ACTIVATION_KEY]
final_activation = backwards_opt_dict[
backwards_opt.MEAN_FINAL_ACTIVATION_KEY]
full_storm_id_string = None
storm_time_string = None
else:
initial_activation = backwards_opt_dict[
backwards_opt.INITIAL_ACTIVATIONS_KEY][example_index]
final_activation = backwards_opt_dict[
backwards_opt.FINAL_ACTIVATIONS_KEY][example_index]
full_storm_id_string = backwards_opt_dict[
backwards_opt.FULL_IDS_KEY][example_index]
storm_time_string = time_conversion.unix_sec_to_string(
backwards_opt_dict[backwards_opt.STORM_TIMES_KEY][example_index],
plot_input_examples.TIME_FORMAT)
conv_2d3d = model_metadata_dict[cnn.CONV_2D3D_KEY]
training_option_dict = model_metadata_dict[cnn.TRAINING_OPTION_DICT_KEY]
if conv_2d3d:
num_fields = len(training_option_dict[trainval_io.RADAR_FIELDS_KEY])
radar_heights_m_agl = numpy.full(num_fields,
radar_utils.SHEAR_HEIGHT_M_ASL,
dtype=int)
else:
radar_heights_m_agl = training_option_dict[
trainval_io.RADAR_HEIGHTS_KEY]
list_of_layer_operation_dicts = model_metadata_dict[
cnn.LAYER_OPERATIONS_KEY]
if list_of_layer_operation_dicts is None:
field_name_by_panel = training_option_dict[
trainval_io.RADAR_FIELDS_KEY]
panel_names = radar_plotting.radar_fields_and_heights_to_panel_names(
field_names=field_name_by_panel, heights_m_agl=radar_heights_m_agl)
else:
field_name_by_panel, panel_names = (
radar_plotting.layer_ops_to_field_and_panel_names(
list_of_layer_operation_dicts=list_of_layer_operation_dicts))
num_panels = len(field_name_by_panel)
plot_cbar_by_panel = numpy.full(num_panels, True, dtype=bool)
cmap_object_by_panel = [colour_map_object] * num_panels
cnorm_object_by_panel = [None] * num_panels
for j in range(num_panels):
this_max_colour_value = numpy.percentile(
numpy.absolute(difference_matrix[..., j]), max_colour_percentile)
cnorm_object_by_panel[j] = matplotlib.colors.Normalize(
vmin=-1 * this_max_colour_value,
vmax=this_max_colour_value,
clip=False)
num_panel_rows = int(numpy.floor(numpy.sqrt(num_panels)))
figure_object, axes_object_matrix = (
radar_plotting.plot_many_2d_grids_without_coords(
field_matrix=numpy.flip(difference_matrix, axis=0),
field_name_by_panel=field_name_by_panel,
num_panel_rows=num_panel_rows,
panel_names=panel_names,
row_major=False,
colour_map_object_by_panel=cmap_object_by_panel,
colour_norm_object_by_panel=cnorm_object_by_panel,
plot_colour_bar_by_panel=plot_cbar_by_panel,
font_size=FONT_SIZE_WITH_COLOUR_BARS))
if significance_matrix is not None:
significance_plotting.plot_many_2d_grids_without_coords(
significance_matrix=numpy.flip(significance_matrix, axis=0),
axes_object_matrix=axes_object_matrix,
row_major=False)
if pmm_flag:
this_title_string = 'PMM'
else:
this_title_string = 'Storm "{0:s}" at {1:s}'.format(
full_storm_id_string, storm_time_string)
this_title_string += '; activation from {0:.2e} to {1:.2e}'.format(
initial_activation, final_activation)
figure_object.suptitle(this_title_string, fontsize=TITLE_FONT_SIZE)
output_file_name = plot_input_examples.metadata_to_radar_fig_file_name(
output_dir_name=output_dir_name,
pmm_flag=pmm_flag,
full_storm_id_string=full_storm_id_string,
storm_time_string=storm_time_string,
radar_field_name='shear' if conv_2d3d else None)
print('Saving figure to: "{0:s}"...'.format(output_file_name))
figure_object.savefig(output_file_name,
dpi=FIGURE_RESOLUTION_DPI,
pad_inches=0,
bbox_inches='tight')
pyplot.close(figure_object)