def main():
swire_names, swire_coords, _ = pipeline.generate_swire_features(
overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(
swire_coords, swire_labels, overwrite=False)
cids = list(
pipeline.cross_identify_all(swire_names,
swire_coords,
swire_labels,
swire_test_sets,
swire_labels[:, 0],
field='cdfs'))
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_swire_norris = {}
key_to_atlas = {}
for row in table:
name = row['Component Name (Franzen)']
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
if not swire or not swire.startswith('SWIRE') or not name:
continue
atlas_to_swire_norris[name] = swire
print(
'Labeller\tClassifier\tQuadrant\tDataset\tn_correct\tn_total\tn_skipped\tAccuracy'
)
for cid in cids:
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES['RGZ & Norris'],
cid.quadrant].nonzero()[0]
n_total = 0
n_correct = 0
n_skipped = 0
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_norris:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_norris[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
n_total += 1
print(cid.labeller,
cid.classifier,
cid.quadrant,
'{:<20}'.format(cid.dataset_name),
n_correct,
n_total,
n_skipped,
'{:.02%}'.format(n_correct / n_total),
sep='\t')
def main():
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False)
cids = list(pipeline.cross_identify_all(swire_names, swire_coords, swire_labels, swire_test_sets, swire_labels[:, 0], field='cdfs'))
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_swire_norris = {}
key_to_atlas = {}
for row in table:
name = row['Component Name (Franzen)']
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
if not swire or not swire.startswith('SWIRE') or not name:
continue
atlas_to_swire_norris[name] = swire
print('Labeller\tClassifier\tQuadrant\tDataset\tn_correct\tn_total\tn_skipped\tAccuracy')
for cid in cids:
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES['RGZ & Norris'], cid.quadrant].nonzero()[0]
n_total = 0
n_correct = 0
n_skipped = 0
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_norris:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_norris[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
n_total += 1
print(cid.labeller, cid.classifier, cid.quadrant, '{:<20}'.format(cid.dataset_name),
n_correct, n_total, n_skipped, '{:.02%}'.format(n_correct / n_total),
sep='\t')
def print_table(field='cdfs'):
titlemap = {
'RGZ & Norris & compact': 'Compact',
'RGZ & Norris & resolved': 'Resolved',
'RGZ & Norris': 'All',
'RGZ & compact': 'Compact',
'RGZ & resolved': 'Resolved',
'RGZ': 'All',
}
lr_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'LogisticRegression_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'LogisticRegression_rgz_{}_predictions'.format(field)))
rf_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'RandomForestClassifier_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'RandomForestClassifier_rgz_{}_predictions'.format(field)))
cnn_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_rgz_{}_predictions'.format(field)))
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False, field=field)
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False, field=field)
swire_names = numpy.array(swire_names)
swire_coords = numpy.array(swire_coords)
predictions_map = collections.defaultdict(dict) # SWIRE -> predictor -> probability
swire_coords_map = {}
swire_expert_map = {}
swire_rgz_map = {}
known_predictors = set()
for classifier, predictions_ in [['LR', lr_predictions], ['CNN', cnn_predictions], ['RF', rf_predictions]]:
for predictions in predictions_:
dataset_name = predictions.dataset_name
labeller = predictions.labeller
if labeller == 'rgz' and 'Norris' in dataset_name:
labeller = 'RGZ N'
continue
labeller = labeller.title() if labeller == 'norris' else labeller.upper()
predictor_name = '{}({} / {})'.format(classifier, labeller, titlemap[dataset_name])
if field == 'cdfs':
swire_names_ = swire_names[swire_test_sets[:, pipeline.SET_NAMES['RGZ'], predictions.quadrant]]
swire_coords_ = swire_coords[swire_test_sets[:, pipeline.SET_NAMES['RGZ'], predictions.quadrant]]
swire_labels_ = swire_labels[swire_test_sets[:, pipeline.SET_NAMES['RGZ'], predictions.quadrant]]
else:
swire_names_ = swire_names[swire_test_sets[:, 0, 0]]
swire_coords_ = swire_coords[swire_test_sets[:, 0, 0]]
swire_labels_ = swire_labels[swire_test_sets[:, 0, 0]]
assert predictions.probabilities.shape[0] == len(swire_names_), \
'expected {}, got {}'.format(predictions.probabilities.shape[0], len(swire_names_))
for name, coords, prediction, label in zip(swire_names_, swire_coords_, predictions.probabilities, swire_labels_):
predictions_map[name][predictor_name] = prediction
swire_coords_map[name] = coords
swire_expert_map[name] = label[0]
swire_rgz_map[name] = label[1]
known_predictors.add(predictor_name)
known_predictors = sorted(known_predictors)
swires = sorted(predictions_map)
ras = []
decs = []
is_expert_host = []
is_rgz_host = []
predictor_columns = collections.defaultdict(list)
for swire in swires:
for predictor in known_predictors:
predictor_columns[predictor].append(predictions_map[swire].get(predictor, ''))
ras.append(swire_coords_map[swire][0])
decs.append(swire_coords_map[swire][1])
is_expert_host.append(['no', 'yes'][swire_expert_map[swire]])
is_rgz_host.append(['no', 'yes'][swire_rgz_map[swire]])
table = astropy.table.Table(
data=[swires, ras, decs, is_expert_host, is_rgz_host] + [predictor_columns[p] for p in known_predictors],
names=['SWIRE', 'RA', 'Dec', 'Expert host', 'RGZ host'] + known_predictors)
table.write('/Users/alger/data/Crowdastro/predicted_swire_table_{}_21_03_18.csv'.format(field), format='csv')
for p in known_predictors:
table[p].format = '{:.4f}'
table.write('/Users/alger/data/Crowdastro/predicted_swire_table_{}_21_03_18.tex'.format(field), format='latex')
def main(examples=None, classifier='CNN', labeller='Norris'):
# Load SWIRE stuff.
swire_names, swire_coords, swire_features = pipeline.generate_swire_features(overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False)
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, overwrite=False)
swire_tree = KDTree(swire_coords)
swire_name_to_index = {n: i for i, n in enumerate(swire_names)}
# Load ATLAS coords.
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_coords = {}
atlas_to_swire_coords = {}
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
atlas_to_coords[name] = row['Component RA (Franzen)'], row['Component DEC (Franzen)']
index = swire_name_to_index.get(row['Source SWIRE (Norris)'] or '')
if index:
atlas_to_swire_coords[name] = swire_coords[index]
ir_stretch = astropy.visualization.LogStretch(0.001)
if examples is None:
examples = examples_incorrect.get_examples()
examples = examples[labeller, classifier, 'All']
for example in examples:
print('Plotting {}'.format(example))
predictor_name = '{}_{}'.format(classifier, labeller)
cid = example[2]
# Load FITS stuff.
try:
radio_fits = astropy.io.fits.open(CDFS_PATH + cid + '_radio.fits')
except FileNotFoundError:
if example[1]: # Has Zooniverse ID
print('{} not in RGZ'.format(cid))
continue
ir_fits = astropy.io.fits.open(CDFS_PATH + cid + '_ir.fits')
wcs = astropy.wcs.WCS(radio_fits[0].header)
# Compute info for contour levels. (also from Enno Middelberg)
median = numpy.median(radio_fits[0].data)
mad = numpy.median(numpy.abs(radio_fits[0].data - median))
sigma = mad / mad2sigma
# Set up the plot.
fig = plt.figure()
ax = astropy.visualization.wcsaxes.WCSAxes(
fig, [0.1, 0.1, 0.8, 0.8], wcs=wcs)
fig.add_axes(ax)
ax.set_title('{}'.format(example[0], example[1]))
# Show the infrared.
ax.imshow(ir_stretch(ir_fits[0].data), cmap='cubehelix_r',
origin='lower')
# Show the radio.
ax.contour(radio_fits[0].data, colors='black',
levels=[nsig * sigma * sigmult ** i for i in range(15)],
linewidths=1, origin='lower', zorder=1)
# Plot predictions.
predictions = get_predictions(swire_tree, swire_coords, swire_test_sets, atlas_to_coords[example[0]], predictor_name)
if not predictions:
print('No predictions for {}'.format(example[0]))
continue
coords = [p[0] for p in predictions]
probabilities = [p[1] for p in predictions]
coords = wcs.all_world2pix(coords, 1)
ax.scatter(coords[:, 0], coords[:, 1], s=numpy.sqrt(numpy.array(probabilities)) * 200, color='white', edgecolor='black', linewidth=1, alpha=0.9, marker='o', zorder=2)
choice = numpy.argmax(probabilities)
ax.scatter(coords[choice, 0], coords[choice, 1], s=200 / numpy.sqrt(2), color='blue', marker='x', zorder=2.5)
try:
norris_coords, = wcs.all_world2pix([atlas_to_swire_coords[example[0]]], 1)
except KeyError:
print('No Norris cross-identification for {}'.format(example[0]))
continue
ax.scatter(norris_coords[0], norris_coords[1], marker='+', s=200, zorder=3, color='green')
lon, lat = ax.coords
lon.set_major_formatter('hh:mm:ss')
lon.set_axislabel('Right Ascension')
lat.set_axislabel('Declination')
fn = '{}_{}_{}'.format(classifier, labeller, example[0])
plt.savefig('/Users/alger/repos/crowdastro-projects/ATLAS-CDFS/images/examples/' + fn + '.png',
bbox_inches='tight', pad_inches=0)
plt.savefig('/Users/alger/repos/crowdastro-projects/ATLAS-CDFS/images/examples/' + fn + '.pdf',
bbox_inches='tight', pad_inches=0)
plt.clf()
def plot_predictions(cut=0.95,
labeller='norris',
dataset_name=None,
classifier=None):
"""Plot colour-colour diagram for predicted host galaxies.
labeller in {'norris', 'rgz'}
dataset_name in {'RGZ & Norris', ...}
"""
with h5py.File(CROWDASTRO_PATH, 'r') as f:
swire_numeric_cdfs = f['/swire/cdfs/numeric'][:, 2:2 + 4]
f_36 = swire_numeric_cdfs[:, 0]
f_45 = swire_numeric_cdfs[:, 1]
f_58 = swire_numeric_cdfs[:, 2]
f_80 = swire_numeric_cdfs[:, 3]
detection_58 = (f_58 != -99)
detection_80 = (f_80 != -99)
p = pipeline.unserialise_predictions(
pipeline.WORKING_DIR +
'{}_{}_cdfs_predictions'.format(classifier, labeller))
predictions = {}
for i in p:
predictions[i.dataset_name, i.quadrant] = i
swire_names, swire_coords, _ = pipeline.generate_swire_features(
overwrite=False, field='cdfs')
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False,
field='cdfs')
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords,
swire_labels,
overwrite=False,
field='cdfs')
xs = []
ys = []
colours = []
for q in range(4):
swire_set = swire_test_sets[:, pipeline.SET_NAMES['RGZ'], q]
if labeller == 'norris' and not dataset_name:
# predictions_set = predictions['RGZ & Norris', q].probabilities > cut
f_36_ = f_36[swire_set & swire_labels[:, 0]] #[predictions_set]
f_45_ = f_45[swire_set & swire_labels[:, 0]] #[predictions_set]
f_58_ = f_58[swire_set & swire_labels[:, 0]] #[predictions_set]
f_80_ = f_80[swire_set & swire_labels[:, 0]] #[predictions_set]
elif labeller == 'rgz' and not dataset_name:
f_36_ = f_36[swire_set & swire_labels[:, 1]]
f_45_ = f_45[swire_set & swire_labels[:, 1]]
f_58_ = f_58[swire_set & swire_labels[:, 1]]
f_80_ = f_80[swire_set & swire_labels[:, 1]]
if labeller == 'norris' and dataset_name:
predictions_set = predictions[dataset_name, q].probabilities > cut
f_36_ = f_36[swire_set][predictions_set]
f_45_ = f_45[swire_set][predictions_set]
f_58_ = f_58[swire_set][predictions_set]
f_80_ = f_80[swire_set][predictions_set]
probabilities = predictions[dataset_name,
q].probabilities[predictions_set]
detection_58_ = (f_58_ != -99)
detection_80_ = (f_80_ != -99)
detection_all_ = detection_58_ & detection_80_
ratio_58_36 = numpy.log10(f_58_[detection_all_] /
f_36_[detection_all_])
ratio_80_45 = numpy.log10(f_80_[detection_all_] /
f_45_[detection_all_])
probabilities = probabilities[detection_all_]
xs.extend(ratio_58_36)
ys.extend(ratio_80_45)
colours.extend(probabilities)
assert len(xs) == len(ys)
assert len(xs) == len(colours)
plot_basic()
if dataset_name:
plt.scatter(xs,
ys,
s=20,
marker='^',
linewidth=0,
alpha=0.5,
c=numpy.array(colours),
cmap='winter')
else:
plt.scatter(xs, ys, s=25, c='r', marker='^', linewidth=0)
plt.xlim((-0.75, 1.0))
plt.ylim((-0.75, 1.0))
plt.xlabel('$\\log_{10}(S_{5.8}/S_{3.6})$')
plt.ylabel('$\\log_{10}(S_{8.0}/S_{4.5})$')
plt.subplots_adjust(left=0.2, bottom=0.15, right=0.95, top=0.95)
plt.colorbar()
plt.show()
def plot_grid(field='cdfs'):
# Load predictions.
lr_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'LogisticRegression_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'LogisticRegression_rgz_{}_predictions'.format(field)))
rf_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'RandomForestClassifier_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'RandomForestClassifier_rgz_{}_predictions'.format(field)))
cnn_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_rgz_{}_predictions'.format(field)))
# Convert to the format we need. e.g. {'RGZ' -> [acc, acc, acc, acc]}
lr_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
lr_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
rf_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
rf_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
cnn_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
cnn_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
for predictions in lr_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
lr_norris_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
else:
lr_rgz_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
for predictions in rf_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
rf_norris_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
else:
rf_rgz_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
for predictions in cnn_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
cnn_norris_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
else:
cnn_rgz_accuracies[dataset_name][predictions.quadrant] = predictions.balanced_accuracy
if field == 'cdfs':
# Load RGZ cross-identifications and compute a balanced accuracy with them.
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False, field=field)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False, field=field)
label_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
label_norris_accuracies = {sstr: [1] * 4 for sstr in pipeline.SET_NAMES} # By definition.
for dataset_name in pipeline.SET_NAMES:
for quadrant in range(4):
test_set = swire_test_sets[:, pipeline.SET_NAMES[dataset_name], quadrant]
predictions = swire_labels[test_set, 1]
trues = swire_labels[test_set, 0]
ba = balanced_accuracy(trues, predictions)
label_rgz_accuracies[dataset_name][quadrant] = ba
colours = ['grey', 'magenta', 'blue', 'orange']
markers = ['o', '^', 'x', 's']
handles = {}
plt.figure(figsize=(5, 5))
accuracy_map = defaultdict(lambda: defaultdict(dict)) # For table output.
output_sets = [
('LR', [lr_norris_accuracies, lr_rgz_accuracies]),
('CNN', [cnn_norris_accuracies, cnn_rgz_accuracies]),
('RF', [rf_norris_accuracies, rf_rgz_accuracies]),
]
if field == 'cdfs':
output_sets.append(('Labels', [label_norris_accuracies, label_rgz_accuracies]))
for j, (classifier_name, classifier_set) in enumerate(output_sets):
for i, set_name in enumerate(norris_labelled_sets):
if 'compact' not in set_name: # Skip compact.
ax = plt.subplot(2, 1, {'RGZ & Norris & resolved': 1, 'RGZ & Norris': 2}[set_name])
ax.set_ylim((80, 100))
ax.set_xlim((-0.5, 1.5))
ax.set_xticks([0, 1])#, 2])
ax.set_xticklabels(['Norris',
# 'RGZ N',
'RGZ',
], rotation='horizontal')
if i == 2:
plt.xlabel('Labels')
plt.ylabel('{}\nBalanced accuracy\n(per cent)'.format(titlemap[set_name]))
ax.title.set_fontsize(16)
ax.xaxis.label.set_fontsize(12)
ax.yaxis.label.set_fontsize(9)
for tick in ax.get_xticklabels() + ax.get_yticklabels():
tick.set_fontsize(10)
ax.grid(which='major', axis='y', color='#EEEEEE')
for k in range(4):
if 'compact' in set_name:
continue
if j != 3: # !Labels
ax.scatter([0 + (j - 1) / 5], classifier_set[0][set_name][k] * 100,
color=colours[j], marker=markers[j], linewidth=1, edgecolor='k')
rgz_offset = ((j - 1.5) / 6) if field == 'cdfs' else (j - 1) / 5
handles[j] = ax.scatter([1 + rgz_offset],
classifier_set[1][fullmap[set_name]][k] * 100,
color=colours[j], marker=markers[j], linewidth=1, edgecolor='k')
# ax.scatter([1 + (j - 1) / 5], classifier_set[1][set_name][k] * 100,
# color=colours[j], marker=markers[j], linewidth=1, edgecolor='k')
# Compute for table.
for labeller in ['Norris', 'RGZ N', 'RGZ']:
if labeller == 'Norris':
mean = numpy.mean(classifier_set[0][set_name]) * 100
stdev = numpy.std(classifier_set[0][set_name]) * 100
elif labeller == 'RGZ N':
continue
# mean = numpy.mean(classifier_set[1][set_name]) * 100
# stdev = numpy.std(classifier_set[1][set_name]) * 100
elif labeller == 'RGZ':
mean = numpy.mean(classifier_set[1][fullmap[set_name]]) * 100
stdev = numpy.std(classifier_set[1][fullmap[set_name]]) * 100
accuracy_map[labeller][classifier_name][titlemap[set_name]] = '${:.02f} \\pm {:.02f}$'.format(mean, stdev)
# Assemble table.
col_labeller = []
col_classifier = []
col_compact = []
col_resolved = []
col_all = []
for labeller in ['Norris', 'RGZ N', 'RGZ']:
if labeller == 'RGZ N':
continue
for classifier in ['CNN', 'LR', 'RF'] + ['Labels'] if field == 'cdfs' else []:
col_labeller.append(labeller)
col_classifier.append(classifier)
col_compact.append(accuracy_map[labeller][classifier]['Compact'])
col_resolved.append(accuracy_map[labeller][classifier]['Resolved'])
col_all.append(accuracy_map[labeller][classifier]['All'])
out_table = astropy.table.Table([col_labeller, col_classifier, col_compact, col_resolved, col_all],
names=['Labeller', 'Classifier', "Mean `Compact' accuracy\\\\(per cent)",
"Mean `Resolved' accuracy\\\\(per cent)",
"Mean `All' accuracy\\\\(per cent)"])
out_table.write('../{}_accuracy_table.tex'.format(field), format='latex')
plt.figlegend([handles[j] for j in sorted(handles)], ['LR', 'CNN', 'RF'] + (['Labels'] if field == 'cdfs' else []), 'lower center', ncol=4, fontsize=10)
plt.subplots_adjust(bottom=0.2, hspace=0.25)
plt.savefig('../images/{}_ba_grid.pdf'.format(field),
bbox_inches='tight', pad_inches=0)
plt.savefig('../images/{}_ba_grid.png'.format(field),
bbox_inches='tight', pad_inches=0)
def get_examples():
titlemap = {
'RGZ & Norris & compact': 'Compact',
'RGZ & Norris & resolved': 'Resolved',
'RGZ & Norris': 'All',
'RGZ & compact': 'Compact',
'RGZ & resolved': 'Resolved',
'RGZ': 'All',
}
swire_names, swire_coords, _ = pipeline.generate_swire_features(
overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names, overwrite=False)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(
swire_coords, overwrite=False)
cids = list(
pipeline.cross_identify_all(swire_names, swire_coords, swire_test_sets,
swire_labels[:, 0]))
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_swire_norris = {}
key_to_atlas = {}
atlas_to_ras = {}
atlas_to_decs = {}
id_to_atlas = {}
atlas_to_zid = {}
atlas_to_id = {}
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
id_to_atlas[row['Component ID (Franzen)']] = name
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
atlas_to_swire_norris[name] = swire
atlas_to_id[name] = row['Component ID (Franzen)']
atlas_to_ras[name] = row['Component RA (Franzen)']
atlas_to_decs[name] = row['Component DEC (Franzen)']
atlas_to_zid[name] = row['Component Zooniverse ID (RGZ)']
atlas_to_rgz = {}
atlas_to_radio_consensus = {}
atlas_to_ir_consensus = {}
for row in astropy.io.ascii.read(pipeline.RGZ_PATH):
name = id_to_atlas[row['atlas_id']]
atlas_to_radio_consensus[name] = row['consensus.radio_level']
atlas_to_ir_consensus[name] = row['consensus.ir_level']
atlas_to_rgz[name] = row['SWIRE.designation']
cross_identifications = collections.defaultdict(
dict) # ATLAS -> labeller -> SWIRE
for cid in cids:
if cid.labeller == 'norris' and 'Norris' not in cid.dataset_name:
continue
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES['RGZ & Norris'],
cid.quadrant].nonzero()[0]
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_norris:
continue
if name not in atlas_to_swire_predictor:
continue
cross_identifications[name]['Norris'] = atlas_to_swire_norris[name]
cross_identifications[name]['RGZ'] = atlas_to_rgz.get(name, None)
cross_identifications[name][
labeller, cid.classifier,
titlemap[cid.dataset_name]] = atlas_to_swire_predictor[name]
"""
For each classifier, pull out examples where:
- RGZ and Norris agree, but the classifier disagrees.
Only include RGZ & Norris dataset ("All").
"""
classifier_to_example = collections.defaultdict(set)
for atlas, cids_ in cross_identifications.items():
if cids_['Norris'] != cids_['RGZ']:
continue
for classifier, swire in cids_.items():
if classifier[2] != 'All' or classifier[1] in {
'Random', 'Groundtruth'
}:
continue
if swire != cids_['Norris']:
classifier_to_example[classifier].add(
(atlas, atlas_to_zid[atlas], atlas_to_id[atlas]))
return classifier_to_example
def plot(field='cdfs'):
log.debug('Getting SWIRE, ATLAS features.')
swire_names, swire_coords, _ = pipeline.generate_swire_features(
overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False,
field=field)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(
swire_coords, swire_labels, overwrite=False, field=field)
log.debug('Calling cross-identify.')
cids = list(
pipeline.cross_identify_all(swire_names,
swire_coords,
swire_labels,
swire_test_sets,
swire_labels[:, 0],
field=field))
# Also load the nearest-neighbour cross-identifications.
cids += [
pipeline.CrossIdentifications.from_hdf5(
pipeline.WORKING_DIR +
'NearestNeighbour_{}_cross_ids_{}_RGZ & Norris.h5'.format(
field, q)) for q in range(4 if field == 'cdfs' else 1)
]
swire_tree = scipy.spatial.KDTree(swire_coords[swire_test_sets[:, 0, 0]])
failed_coords = []
if field == 'cdfs':
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
rgzcat = astropy.io.ascii.read(pipeline.RGZ_PATH)
atlas_to_swire_expert = {}
atlas_to_swire_rgz = {}
key_to_atlas = {}
atlas_id_to_name = {}
is_compact = {}
for row in table:
name = row['Component Name (Franzen)']
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
if not swire or not swire.startswith('SWIRE') or not name:
continue
atlas_id_to_name[row['Component ID (Franzen)']] = name
atlas_to_swire_expert[name] = swire
is_compact[name] = pipeline.compact_test(row)
for row in rgzcat:
swire_name = row['SWIRE.designation']
if not swire_name or swire_name == '-99':
continue
name = atlas_id_to_name.get(row['atlas_id'], None)
atlas_to_swire_rgz[name] = swire_name
else:
atlas_to_swire_expert = {}
with astropy.io.fits.open(
pipeline.MIDDELBERG_TABLE4_PATH) as elais_components_fits:
elais_components = elais_components_fits[1].data
atlas_cid_to_name = {}
atlas_names = [] # Indices correspond to table 4 rows.
atlas_name_to_compact = {}
for component in elais_components:
cid = component['CID']
name = component['ATELAIS']
atlas_names.append(name)
atlas_cid_to_name[cid] = name
row = {
'Component S (Franzen)':
component['Sint'], # Fitting in with the CDFS API...
'Component S_ERR (Franzen)': component['e_Sint'],
'Component Sp (Franzen)': component['Sp'],
'Component Sp_ERR (Franzen)': component['e_Sp']
}
atlas_name_to_compact[name] = pipeline.compact_test(row)
with open(pipeline.MIDDELBERG_TABLE5_PATH) as elais_file:
# Took this code from pipeline.py, probably should make it a function
lines = [line.split('|') for line in elais_file]
for line in lines:
if 'ATELAISJ' not in line[0]:
continue
line_cids = line[1]
if 'C0' not in line_cids and 'C1' not in line_cids:
continue
line_cids = [cid.strip() for cid in line_cids.split(',')]
swire_coord_re = re.search(
r'SWIRE4J(\d\d)(\d\d)(\d\d\.\d\d)(-\d\d)(\d\d)(\d\d\.\d)',
line[2])
if not swire_coord_re:
continue
swire_coord_list = swire_coord_re.groups()
coord = astropy.coordinates.SkyCoord(
ra='{} {} {}'.format(*swire_coord_list[:3]),
dec='{} {} {}'.format(*swire_coord_list[3:]),
unit=('hourangle', 'deg'))
coord = (coord.ra.deg, coord.dec.deg)
# Nearest SWIRE...
dist, nearest = swire_tree.query(coord)
if dist > 5 / 60 / 60:
logging.debug(
'No SWIRE match found for Middelberg cross-identification {}'
.format(line[0]))
logging.debug('Nearest is {} ({:.01f} arcsec)'.format(
numpy.array(swire_names)[swire_test_sets[:, 0,
0]][nearest],
dist * 60 * 60))
logging.debug('Middelberg: {}'.format(
swire_coord_re.group()))
failed_coords.append(coord)
continue
name = numpy.array(swire_names)[swire_test_sets[:, 0,
0]][nearest]
for cid in line_cids:
atlas_to_swire_expert[atlas_cid_to_name[cid]] = name
labeller_classifier_to_accuracies = collections.defaultdict(list)
# Augment the CIDs by duplicating the "resolved" cross-ids to make the "all" set.
resolved_cids_copy = [
copy.copy(cid) for cid in cids if 'resolved' in cid.dataset_name
]
for cid in resolved_cids_copy:
cid.dataset_name = cid.dataset_name.replace(' & resolved', '')
cids.extend(resolved_cids_copy)
for cid in cids:
if cid.labeller == 'norris' and 'Norris' not in cid.dataset_name:
continue
if cid.classifier in {'Groundtruth', 'Random', 'NearestNeighbour'}:
# Deal with these later as they are special.
continue
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
n_compact = 0
if field == 'cdfs':
atlas_keys = atlas_test_sets[:, pipeline.
SET_NAMES[whatset[cid.dataset_name]],
cid.quadrant].nonzero()[0]
# For each ATLAS object in RGZ & Norris...
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_expert[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
n_total += 1
else:
# Only one test set for ELAIS.
atlas_indices = atlas_test_sets[:, 0, 0].nonzero()[0]
assert atlas_test_sets.shape[0] == len(atlas_names)
for index in atlas_indices:
# Screen resolved here.
atlas_name = atlas_names[index]
if atlas_name not in atlas_to_swire_expert:
n_skipped += 1
continue
if atlas_name not in atlas_to_swire_predictor:
n_skipped += 1
continue
if 'resolved' in cid.dataset_name and atlas_name_to_compact[
atlas_name]:
n_compact += 1
continue
swire_middelberg = atlas_to_swire_expert[atlas_name]
swire_predictor = atlas_to_swire_predictor[atlas_name]
n_correct += swire_middelberg == swire_predictor
n_total += 1
# print('Compact: {:.02%}'.format(n_compact / (n_total + n_compact)))
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
labeller_classifier_to_accuracies[labeller, cid.classifier,
titlemap[cid.dataset_name]].append(
n_correct / n_total)
# Groundtruth, random, and NN classifiers exist only for the RGZ & Norris set, but we want to test on all subsets.
# This section duplicates the classifiers and evaluates them on all subsets.
for cid in cids:
if cid.classifier not in {'Groundtruth', 'Random', 'NearestNeighbour'}:
continue
for dataset_name in [
'RGZ & Norris', 'RGZ & Norris & resolved',
'RGZ & Norris & compact'
]:
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
if field == 'cdfs':
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:,
pipeline.SET_NAMES[dataset_name],
cid.quadrant].nonzero()[0]
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_expert[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
if cid.classifier == 'NearestNeighbour' and swire_norris != swire_predictor:
pass
n_total += 1
else:
atlas_indices = atlas_test_sets[:, 0, 0].nonzero()[0]
assert atlas_test_sets.shape[0] == len(atlas_names)
for index in atlas_indices:
# Screen resolved here (because the test sets aren't useful for that for ELAIS)
atlas_name = atlas_names[index]
if 'resolved' in dataset_name and atlas_name_to_compact[
atlas_name]:
continue
if atlas_name not in atlas_to_swire_expert:
n_skipped += 1
continue
if atlas_name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_middelberg = atlas_to_swire_expert[atlas_name]
swire_predictor = atlas_to_swire_predictor[atlas_name]
n_correct += swire_middelberg == swire_predictor
n_total += 1
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
print(labeller, cid.classifier, titlemap[dataset_name], n_correct,
n_total, n_correct / n_total)
labeller_classifier_to_accuracies[labeller, cid.classifier,
titlemap[dataset_name]].append(
n_correct / n_total)
if field == 'cdfs':
# Compute accuracy for RGZ.
for dataset_name in pipeline.SET_NAMES:
for quadrant in range(4):
# N.B. Disabled using the pipeline for RGZ.
# # Compact objects are cross-identified in a separate pipeline, which is slow so I don't want to reproduce it here.
# # So I'll read the compact object cross-identifications from the LR(RGZ) cross-identification set, since it ought
# # to be the same.
# corresponding_set, = [cid for cid in cids if cid.quadrant == quadrant
# and cid.dataset_name == dataset_name
# and cid.labeller == 'rgz'
# and cid.classifier == 'LogisticRegression']
# atlas_to_swire_lr = dict(zip(corresponding_set.radio_names, corresponding_set.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
n_compact = 0
atlas_keys = atlas_test_sets[:, pipeline.
SET_NAMES[whatset[dataset_name]],
quadrant].nonzero()[0]
# For each ATLAS object in RGZ & Norris...
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_rgz: # or name not in atlas_to_swire_lr:
n_skipped += 1
continue
if False and is_compact[name]:
swire_predictor = atlas_to_swire_lr[name]
else:
swire_predictor = atlas_to_swire_rgz[name]
swire_norris = atlas_to_swire_expert[name]
n_correct += swire_norris == swire_predictor
n_total += 1
labeller_classifier_to_accuracies[
'RGZ', 'Label',
titlemap[dataset_name]].append(n_correct / n_total)
labeller_classifier_to_accuracy = {}
labeller_classifier_to_stdev = {}
for key, accuracies in labeller_classifier_to_accuracies.items():
print('Best {}:'.format(key), max(accuracies))
labeller_classifier_to_accuracy[key] = numpy.mean(accuracies)
labeller_classifier_to_stdev[key] = numpy.std(accuracies)
random_acc = {
k[2]: v * 100
for k, v in labeller_classifier_to_accuracy.items() if k[1] == 'Random'
}
random_stdev = {
k[2]: v * 100
for k, v in labeller_classifier_to_stdev.items() if k[1] == 'Random'
}
best_acc = {
k[2]: v * 100
for k, v in labeller_classifier_to_accuracy.items()
if k[1] == 'Groundtruth'
}
best_stdev = {
k[2]: v * 100
for k, v in labeller_classifier_to_stdev.items()
if k[1] == 'Groundtruth'
}
print('Best: {} +- {}'.format(best_acc, best_stdev))
print('Random: {} +- {}'.format(random_acc, random_stdev))
plt.figure()
colours = ['grey', 'magenta', 'blue', 'orange', 'grey']
markers = ['o', '^', 'x', 's', '*']
handles = {}
print('Data set & Labeller & Classifier & Mean accuracy (\\%)\\\\')
for k, set_name in enumerate(norris_labelled_sets[1:]):
if 'resolved' in set_name:
# https://github.com/MatthewJA/radio/issues/22
continue
k -= 1
print_set_name = titlemap[set_name]
ax = plt.subplot(1, 1, 1 + k) # 22
print('{} & Norris & Perfect & ${:.02f} \\pm {:.02f}$\\\\'.format(
print_set_name, best_acc[titlemap[set_name]],
best_stdev[titlemap[set_name]]))
print('{} & Norris & Random & ${:.02f} \\pm {:.02f}$\\\\'.format(
print_set_name, random_acc[titlemap[set_name]],
random_stdev[titlemap[set_name]]))
plt.hlines(best_acc[titlemap[set_name]],
-0.5,
2.5,
linestyles='solid',
colors='green',
linewidth=1,
zorder=1)
plt.fill_between(
[-1, 2],
[best_acc[titlemap[set_name]] - best_stdev[titlemap[set_name]]] *
2,
[best_acc[titlemap[set_name]] + best_stdev[titlemap[set_name]]] *
2,
linestyle='dashed',
color='green',
alpha=0.2,
linewidth=1,
zorder=1)
plt.hlines(random_acc[titlemap[set_name]],
-0.5,
2.5,
linestyles='solid',
colors='blue',
linewidth=1,
zorder=1,
alpha=0.7)
plt.fill_between([-1, 2],
[
random_acc[titlemap[set_name]] -
random_stdev[titlemap[set_name]]
] * 2,
[
random_acc[titlemap[set_name]] +
random_stdev[titlemap[set_name]]
] * 2,
linestyle='dashed',
color='blue',
alpha=0.2,
linewidth=1,
zorder=1)
for i, labeller in enumerate(['Norris', 'RGZ']):
for j, classifier in enumerate(
['LogisticRegression', 'CNN', 'RandomForestClassifier'] +
(['Label', 'NearestNeighbour'] if field ==
'cdfs' else ['NearestNeighbour'])):
ys = numpy.array(labeller_classifier_to_accuracies[
labeller, classifier, titlemap[set_name]]) * 100
if classifier != 'NearestNeighbour':
x_offset = i + (
j - 1
) / 5 if labeller == 'Norris' or field == 'elais' else i + (
j - 1.5) / 6
else:
# NN
plt.axhline(numpy.mean(ys),
color='grey',
linestyle='-.',
linewidth=1)
if field == 'cdfs':
plt.fill_between([-1, 2],
[numpy.mean(ys) - numpy.std(ys)] * 2,
[numpy.mean(ys) + numpy.std(ys)] * 2,
color='grey',
linestyle='-.',
alpha=0.2,
linewidth=1)
x_offset = 2
if classifier == 'Label' and labeller == 'RGZ':
plt.annotate('{:.1%}'.format(numpy.mean(ys) / 100),
(x_offset, 72.5),
ha='center',
va='bottom')
plt.arrow(x_offset,
72.5,
0,
-1.5,
head_width=0.05,
head_length=1,
ec='k',
fc='k')
xs = [x_offset] * len(ys)
print('{} & {} & {} & ${:.02f} \\pm {:.02f}$\\\\'.format(
print_set_name, labeller, classifier, numpy.mean(ys),
numpy.std(ys)))
ax.set_xlim((-0.5, 1.5))
ax.set_ylim((70, 100))
ax.set_xticks([0, 1])
ax.set_xticklabels(['Norris', 'RGZ'])
ax.set_yticklabels(
['{}\%'.format(x) for x in range(70, 101, 5)])
handles[j] = plt.scatter(xs,
ys,
color=colours[j],
marker=markers[j],
zorder=2,
edgecolor='k',
linewidth=1)
# if k == 0: # 22
# plt.xlabel('Labels')
plt.ylabel('Cross-identification\naccuracy (per cent)'.format(
titlemap[set_name]))
# ax.title.set_fontsize(16)
# ax.xaxis.label.set_fontsize(12)
# ax.yaxis.label.set_fontsize(9)
# for tick in ax.get_xticklabels() + ax.get_yticklabels():
# tick.set_fontsize(10)
ax.grid(which='major', axis='y', color='#DDDDDD')
# Print the table.
print('\\hline')
print('Labeller & Classifier & Mean `Resolved\' & Mean `All\'\\')
print('&& accuracy (per cent) & accuracy (per cent)\\')
print('\\hline')
for labeller in ['Norris', 'RGZ']:
for classifier in [
'CNN', 'LogisticRegression', 'RandomForestClassifier',
'Groundtruth', 'Random', 'Label', 'NearestNeighbour'
]:
if labeller == 'RGZ' and classifier in {
'Groundtruth', 'Random', 'NearestNeighbour'
}:
continue
if labeller == 'Norris' and classifier == 'Label':
continue
print(
'{} & {} & ${:.1f} \\pm {:.1f}$ & ${:.1f} \\pm {:.1f}$'.format(
labeller, classifier,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier,
'Resolved']).mean() *
100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier,
'Resolved']).std() *
100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier,
'All']).mean() * 100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier,
'All']).std() * 100))
plt.gca().tick_params(axis='both',
which='major',
direction='out',
length=5)
plt.gca().tick_params(axis='y', which='minor', direction='out', length=3)
plt.gca().minorticks_on()
plt.figlegend([handles[j] for j in sorted(handles)], ['LR', 'CNN', 'RF'] +
(['Labels'] if field == 'cdfs' else []),
'lower center',
ncol=4)
plt.subplots_adjust(bottom=0.25, hspace=0.25, left=0.3)
plt.savefig('../images/{}_cross_identification_grid.pdf'.format(field))
plt.savefig('../images/{}_cross_identification_grid.png'.format(field))
def main(classifier='CNN', labeller='Norris'):
# Load SWIRE stuff.
swire_names, swire_coords, swire_features = pipeline.generate_swire_features(
overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False)
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords,
overwrite=False)
swire_tree = KDTree(swire_coords)
swire_name_to_index = {n: i for i, n in enumerate(swire_names)}
atlas_names = []
atlas_compactnesses = []
atlas_coords = []
atlas_norris_swire = []
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
if not row['Component Zooniverse ID (RGZ)']:
continue
compactness = pipeline.compactness(row)
atlas_names.append(name)
atlas_compactnesses.append(compactness)
atlas_coords.append(
(row['Component RA (Franzen)'], row['Component DEC (Franzen)']))
atlas_norris_swire.append(row['Source SWIRE (Norris)'])
ys = []
xs_entropy = []
xs_margin = []
no_groundtruth = []
correct = []
for name, compactness, coords, swire in zip(atlas_names,
atlas_compactnesses,
atlas_coords,
atlas_norris_swire):
predictor_name = '{}_{}'.format(classifier, labeller)
predictions = get_predictions(swire_tree, swire_coords,
numpy.array(swire_names),
swire_test_sets, coords, predictor_name)
if not predictions:
print('No predictions for {}'.format(name))
continue
chosen_swire = predictions[numpy.argmax([p
for _, p in predictions])][0]
predictions = [p for _, p in predictions]
predictions_softmax = [
numpy.exp(p) / sum(numpy.exp(p) for p in predictions)
for p in predictions
]
if len(predictions_softmax) == 1:
entropy_ambiguity = 0
margin_ambiguity = 0
else:
entropy_ambiguity = -sum(p * numpy.log(p)
for p in predictions_softmax if p)
predictions.sort()
margin_ambiguity = 1 - (predictions[-1] - predictions[-2])
ys.append(compactness)
xs_entropy.append(entropy_ambiguity)
xs_margin.append(margin_ambiguity)
no_groundtruth.append(not swire or not swire.startswith('SWIRE'))
correct.append(swire == chosen_swire)
ys = numpy.array(ys)
xs_margin = numpy.array(xs_margin)
xs_entropy = numpy.array(xs_entropy)
no_groundtruth = numpy.array(no_groundtruth, dtype=bool)
correct = numpy.array(correct, dtype=bool)
print(sum(1 for y in ys if y <= 1))
plt.subplot(1, 2, 1)
plt.scatter(xs_margin[no_groundtruth],
ys[no_groundtruth],
marker='x',
color='black',
alpha=0.05)
plt.scatter(xs_margin[~no_groundtruth & correct],
ys[~no_groundtruth & correct],
marker='x',
color='blue',
alpha=0.7)
plt.scatter(xs_margin[~no_groundtruth & ~correct],
ys[~no_groundtruth & ~correct],
marker='x',
color='magenta',
alpha=0.7)
plt.title('Margin')
plt.xlabel('1 - margin')
plt.ylabel('$1.3 SNR S / 10 S_p$')
plt.yscale('log')
plt.axhline(1, min(xs_margin), max(xs_margin))
plt.subplot(1, 2, 2)
plt.scatter(xs_entropy[no_groundtruth],
ys[no_groundtruth],
marker='x',
color='black',
alpha=0.05)
plt.scatter(xs_entropy[~no_groundtruth & correct],
ys[~no_groundtruth & correct],
marker='x',
color='blue',
alpha=0.7)
plt.scatter(xs_entropy[~no_groundtruth & ~correct],
ys[~no_groundtruth & ~correct],
marker='x',
color='magenta',
alpha=0.7)
plt.title('Entropy')
plt.xlabel('Entropy')
plt.ylabel('$1.3 SNR S / 10 S_p$')
plt.yscale('log')
plt.axhline(1,
min(xs_entropy),
max(xs_entropy),
zorder=-100,
linestyle='--',
color='black')
plt.show()
def plot(field='cdfs'):
log.debug('Getting SWIRE, ATLAS features.')
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False, field=field)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False, field=field)
log.debug('Calling cross-identify.')
cids = list(pipeline.cross_identify_all(swire_names, swire_coords, swire_labels, swire_test_sets, swire_labels[:, 0], field=field))
# Also load the nearest-neighbour cross-identifications.
cids += [pipeline.CrossIdentifications.from_hdf5(
pipeline.WORKING_DIR + 'NearestNeighbour_{}_cross_ids_{}_RGZ & Norris.h5'.format(field, q)) for q in range(4 if field == 'cdfs' else 1)]
swire_tree = scipy.spatial.KDTree(swire_coords[swire_test_sets[:, 0, 0]])
failed_coords = []
if field == 'cdfs':
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
rgzcat = astropy.io.ascii.read(pipeline.RGZ_PATH)
atlas_to_swire_expert = {}
atlas_to_swire_rgz = {}
key_to_atlas = {}
atlas_id_to_name = {}
is_compact = {}
for row in table:
name = row['Component Name (Franzen)']
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
if not swire or not swire.startswith('SWIRE') or not name:
continue
atlas_id_to_name[row['Component ID (Franzen)']] = name
atlas_to_swire_expert[name] = swire
is_compact[name] = pipeline.compact_test(row)
for row in rgzcat:
swire_name = row['SWIRE.designation']
if not swire_name or swire_name == '-99':
continue
name = atlas_id_to_name.get(row['atlas_id'], None)
atlas_to_swire_rgz[name] = swire_name
else:
atlas_to_swire_expert = {}
with astropy.io.fits.open(pipeline.MIDDELBERG_TABLE4_PATH) as elais_components_fits:
elais_components = elais_components_fits[1].data
atlas_cid_to_name = {}
atlas_names = [] # Indices correspond to table 4 rows.
atlas_name_to_compact = {}
for component in elais_components:
cid = component['CID']
name = component['ATELAIS']
atlas_names.append(name)
atlas_cid_to_name[cid] = name
row = {'Component S (Franzen)': component['Sint'], # Fitting in with the CDFS API...
'Component S_ERR (Franzen)': component['e_Sint'],
'Component Sp (Franzen)': component['Sp'],
'Component Sp_ERR (Franzen)': component['e_Sp']}
atlas_name_to_compact[name] = pipeline.compact_test(row)
with open(pipeline.MIDDELBERG_TABLE5_PATH) as elais_file:
# Took this code from pipeline.py, probably should make it a function
lines = [line.split('|') for line in elais_file]
for line in lines:
if 'ATELAISJ' not in line[0]:
continue
line_cids = line[1]
if 'C0' not in line_cids and 'C1' not in line_cids:
continue
line_cids = [cid.strip() for cid in line_cids.split(',')]
swire_coord_re = re.search(r'SWIRE4J(\d\d)(\d\d)(\d\d\.\d\d)(-\d\d)(\d\d)(\d\d\.\d)', line[2])
if not swire_coord_re:
continue
swire_coord_list = swire_coord_re.groups()
coord = astropy.coordinates.SkyCoord(
ra='{} {} {}'.format(*swire_coord_list[:3]),
dec='{} {} {}'.format(*swire_coord_list[3:]),
unit=('hourangle', 'deg'))
coord = (coord.ra.deg, coord.dec.deg)
# Nearest SWIRE...
dist, nearest = swire_tree.query(coord)
if dist > 5 / 60 / 60:
logging.debug('No SWIRE match found for Middelberg cross-identification {}'.format(line[0]))
logging.debug('Nearest is {} ({:.01f} arcsec)'.format(numpy.array(swire_names)[swire_test_sets[:, 0, 0]][nearest], dist * 60 * 60))
logging.debug('Middelberg: {}'.format(swire_coord_re.group()))
failed_coords.append(coord)
continue
name = numpy.array(swire_names)[swire_test_sets[:, 0, 0]][nearest]
for cid in line_cids:
atlas_to_swire_expert[atlas_cid_to_name[cid]] = name
labeller_classifier_to_accuracies = collections.defaultdict(list)
# Augment the CIDs by duplicating the "resolved" cross-ids to make the "all" set.
resolved_cids_copy = [copy.copy(cid) for cid in cids if 'resolved' in cid.dataset_name]
for cid in resolved_cids_copy:
cid.dataset_name = cid.dataset_name.replace(' & resolved', '')
cids.extend(resolved_cids_copy)
for cid in cids:
if cid.labeller == 'norris' and 'Norris' not in cid.dataset_name:
continue
if cid.classifier in {'Groundtruth', 'Random', 'NearestNeighbour'}:
# Deal with these later as they are special.
continue
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
n_compact = 0
if field == 'cdfs':
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES[whatset[cid.dataset_name]], cid.quadrant].nonzero()[0]
# For each ATLAS object in RGZ & Norris...
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_expert[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
n_total += 1
else:
# Only one test set for ELAIS.
atlas_indices = atlas_test_sets[:, 0, 0].nonzero()[0]
assert atlas_test_sets.shape[0] == len(atlas_names)
for index in atlas_indices:
# Screen resolved here.
atlas_name = atlas_names[index]
if atlas_name not in atlas_to_swire_expert:
n_skipped += 1
continue
if atlas_name not in atlas_to_swire_predictor:
n_skipped += 1
continue
if 'resolved' in cid.dataset_name and atlas_name_to_compact[atlas_name]:
n_compact += 1
continue
swire_middelberg = atlas_to_swire_expert[atlas_name]
swire_predictor = atlas_to_swire_predictor[atlas_name]
n_correct += swire_middelberg == swire_predictor
n_total += 1
# print('Compact: {:.02%}'.format(n_compact / (n_total + n_compact)))
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
labeller_classifier_to_accuracies[labeller, cid.classifier, titlemap[cid.dataset_name]].append(n_correct / n_total)
# Groundtruth, random, and NN classifiers exist only for the RGZ & Norris set, but we want to test on all subsets.
# This section duplicates the classifiers and evaluates them on all subsets.
for cid in cids:
if cid.classifier not in {'Groundtruth', 'Random', 'NearestNeighbour'}:
continue
for dataset_name in ['RGZ & Norris', 'RGZ & Norris & resolved', 'RGZ & Norris & compact']:
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
if field == 'cdfs':
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES[dataset_name], cid.quadrant].nonzero()[0]
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_norris = atlas_to_swire_expert[name]
swire_predictor = atlas_to_swire_predictor[name]
n_correct += swire_norris == swire_predictor
if cid.classifier == 'NearestNeighbour' and swire_norris != swire_predictor:
pass
n_total += 1
else:
atlas_indices = atlas_test_sets[:, 0, 0].nonzero()[0]
assert atlas_test_sets.shape[0] == len(atlas_names)
for index in atlas_indices:
# Screen resolved here (because the test sets aren't useful for that for ELAIS)
atlas_name = atlas_names[index]
if 'resolved' in dataset_name and atlas_name_to_compact[atlas_name]:
continue
if atlas_name not in atlas_to_swire_expert:
n_skipped += 1
continue
if atlas_name not in atlas_to_swire_predictor:
n_skipped += 1
continue
swire_middelberg = atlas_to_swire_expert[atlas_name]
swire_predictor = atlas_to_swire_predictor[atlas_name]
n_correct += swire_middelberg == swire_predictor
n_total += 1
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
print(labeller, cid.classifier, titlemap[dataset_name], n_correct, n_total, n_correct / n_total)
labeller_classifier_to_accuracies[labeller, cid.classifier, titlemap[dataset_name]].append(n_correct / n_total)
if field == 'cdfs':
# Compute accuracy for RGZ.
for dataset_name in pipeline.SET_NAMES:
for quadrant in range(4):
# N.B. Disabled using the pipeline for RGZ.
# # Compact objects are cross-identified in a separate pipeline, which is slow so I don't want to reproduce it here.
# # So I'll read the compact object cross-identifications from the LR(RGZ) cross-identification set, since it ought
# # to be the same.
# corresponding_set, = [cid for cid in cids if cid.quadrant == quadrant
# and cid.dataset_name == dataset_name
# and cid.labeller == 'rgz'
# and cid.classifier == 'LogisticRegression']
# atlas_to_swire_lr = dict(zip(corresponding_set.radio_names, corresponding_set.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
n_compact = 0
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES[whatset[dataset_name]], quadrant].nonzero()[0]
# For each ATLAS object in RGZ & Norris...
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_expert:
n_skipped += 1
continue
if name not in atlas_to_swire_rgz:# or name not in atlas_to_swire_lr:
n_skipped += 1
continue
if False and is_compact[name]:
swire_predictor = atlas_to_swire_lr[name]
else:
swire_predictor = atlas_to_swire_rgz[name]
swire_norris = atlas_to_swire_expert[name]
n_correct += swire_norris == swire_predictor
n_total += 1
labeller_classifier_to_accuracies['RGZ', 'Label', titlemap[dataset_name]].append(n_correct / n_total)
labeller_classifier_to_accuracy = {}
labeller_classifier_to_stdev = {}
for key, accuracies in labeller_classifier_to_accuracies.items():
print('Best {}:'.format(key), max(accuracies))
labeller_classifier_to_accuracy[key] = numpy.mean(accuracies)
labeller_classifier_to_stdev[key] = numpy.std(accuracies)
random_acc = {k[2]: v * 100
for k, v in labeller_classifier_to_accuracy.items()
if k[1] == 'Random'}
random_stdev = {k[2]: v * 100
for k, v in labeller_classifier_to_stdev.items()
if k[1] == 'Random'}
best_acc = {k[2]: v * 100
for k, v in labeller_classifier_to_accuracy.items()
if k[1] == 'Groundtruth'}
best_stdev = {k[2]: v * 100
for k, v in labeller_classifier_to_stdev.items()
if k[1] == 'Groundtruth'}
print('Best: {} +- {}'.format(best_acc, best_stdev))
print('Random: {} +- {}'.format(random_acc, random_stdev))
plt.figure()
colours = ['grey', 'magenta', 'blue', 'orange', 'grey']
markers = ['o', '^', 'x', 's', '*']
handles = {}
print('Data set & Labeller & Classifier & Mean accuracy (\\%)\\\\')
for k, set_name in enumerate(norris_labelled_sets[1:]):
if 'resolved' in set_name:
# https://github.com/MatthewJA/radio/issues/22
continue
k -= 1
print_set_name = titlemap[set_name]
ax = plt.subplot(1, 1, 1 + k) # 22
print('{} & Norris & Perfect & ${:.02f} \\pm {:.02f}$\\\\'.format(print_set_name, best_acc[titlemap[set_name]], best_stdev[titlemap[set_name]]))
print('{} & Norris & Random & ${:.02f} \\pm {:.02f}$\\\\'.format(print_set_name, random_acc[titlemap[set_name]], random_stdev[titlemap[set_name]]))
plt.hlines(best_acc[titlemap[set_name]], -0.5, 2.5, linestyles='solid', colors='green', linewidth=1, zorder=1)
plt.fill_between([-1, 2],
[best_acc[titlemap[set_name]] - best_stdev[titlemap[set_name]]] * 2,
[best_acc[titlemap[set_name]] + best_stdev[titlemap[set_name]]] * 2,
linestyle='dashed', color='green', alpha=0.2, linewidth=1, zorder=1)
plt.hlines(random_acc[titlemap[set_name]], -0.5, 2.5, linestyles='solid', colors='blue', linewidth=1, zorder=1, alpha=0.7)
plt.fill_between([-1, 2],
[random_acc[titlemap[set_name]] - random_stdev[titlemap[set_name]]] * 2,
[random_acc[titlemap[set_name]] + random_stdev[titlemap[set_name]]] * 2,
linestyle='dashed', color='blue', alpha=0.2, linewidth=1, zorder=1)
for i, labeller in enumerate(['Norris', 'RGZ']):
for j, classifier in enumerate(['LogisticRegression', 'CNN', 'RandomForestClassifier'] + (['Label', 'NearestNeighbour'] if field == 'cdfs' else ['NearestNeighbour'])):
ys = numpy.array(labeller_classifier_to_accuracies[labeller, classifier, titlemap[set_name]]) * 100
if classifier != 'NearestNeighbour':
x_offset = i + (j - 1) / 5 if labeller == 'Norris' or field == 'elais' else i + (j - 1.5) / 6
else:
# NN
plt.axhline(numpy.mean(ys), color='grey', linestyle='-.', linewidth=1)
if field == 'cdfs':
plt.fill_between([-1, 2], [numpy.mean(ys) - numpy.std(ys)] * 2, [numpy.mean(ys) + numpy.std(ys)] * 2, color='grey', linestyle='-.', alpha=0.2, linewidth=1)
x_offset = 2
if classifier == 'Label' and labeller == 'RGZ':
plt.annotate('{:.1%}'.format(numpy.mean(ys) / 100), (x_offset, 72.5), ha='center', va='bottom')
plt.arrow(x_offset, 72.5, 0, -1.5, head_width=0.05, head_length=1, ec='k', fc='k')
xs = [x_offset] * len(ys)
print('{} & {} & {} & ${:.02f} \\pm {:.02f}$\\\\'.format(print_set_name, labeller, classifier, numpy.mean(ys), numpy.std(ys)))
ax.set_xlim((-0.5, 1.5))
ax.set_ylim((70, 100))
ax.set_xticks([0, 1])
ax.set_xticklabels(['Norris', 'RGZ'])
ax.set_yticklabels(['{}\%'.format(x) for x in range(70, 101, 5)])
handles[j] = plt.scatter(xs, ys, color=colours[j], marker=markers[j], zorder=2, edgecolor='k', linewidth=1)
# if k == 0: # 22
# plt.xlabel('Labels')
plt.ylabel('Cross-identification\naccuracy (per cent)'.format(titlemap[set_name]))
# ax.title.set_fontsize(16)
# ax.xaxis.label.set_fontsize(12)
# ax.yaxis.label.set_fontsize(9)
# for tick in ax.get_xticklabels() + ax.get_yticklabels():
# tick.set_fontsize(10)
ax.grid(which='major', axis='y', color='#DDDDDD')
# Print the table.
print('\\hline')
print('Labeller & Classifier & Mean `Resolved\' & Mean `All\'\\')
print('&& accuracy (per cent) & accuracy (per cent)\\')
print('\\hline')
for labeller in ['Norris', 'RGZ']:
for classifier in ['CNN', 'LogisticRegression', 'RandomForestClassifier', 'Groundtruth', 'Random', 'Label', 'NearestNeighbour']:
if labeller == 'RGZ' and classifier in {'Groundtruth', 'Random', 'NearestNeighbour'}:
continue
if labeller == 'Norris' and classifier == 'Label':
continue
print('{} & {} & ${:.1f} \\pm {:.1f}$ & ${:.1f} \\pm {:.1f}$'.format(
labeller, classifier,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier, 'Resolved']).mean() * 100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier, 'Resolved']).std() * 100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier, 'All']).mean() * 100,
numpy.array(
labeller_classifier_to_accuracies[labeller, classifier, 'All']).std() * 100))
plt.gca().tick_params(axis='both', which='major', direction='out', length=5)
plt.gca().tick_params(axis='y', which='minor', direction='out', length=3)
plt.gca().minorticks_on()
plt.figlegend([handles[j] for j in sorted(handles)], ['LR', 'CNN', 'RF'] + (['Labels'] if field == 'cdfs' else []), 'lower center', ncol=4)
plt.subplots_adjust(bottom=0.25, hspace=0.25, left=0.3)
plt.savefig('../images/{}_cross_identification_grid.pdf'.format(field))
plt.savefig('../images/{}_cross_identification_grid.png'.format(field))
def print_table(field='cdfs'):
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False, field=field)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False, field=field)
cids = list(pipeline.cross_identify_all(swire_names, swire_coords, swire_labels, swire_test_sets, swire_labels[:, 0], field=field))
atlas_to_swire = collections.defaultdict(dict) # ATLAS -> predictor -> SWIRE
swire_name_to_coord = {}
for name, coord in zip(swire_names, swire_coords):
swire_name_to_coord[name] = coord
atlas_to_swire_expert = {}
key_to_atlas = {}
atlas_to_ras = {}
atlas_to_decs = {}
id_to_atlas = {}
atlas_to_id = {}
atlas_to_zooniverse_id = {}
if field == 'cdfs':
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
id_to_atlas[row['Component ID (Franzen)']] = name
atlas_to_id[name] = row['Component ID (Franzen)']
atlas_to_zooniverse_id[name] = row['Component Zooniverse ID (RGZ)']
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
atlas_to_swire_expert[name] = swire
atlas_to_ras[name] = row['Component RA (Franzen)']
atlas_to_decs[name] = row['Component DEC (Franzen)']
else:
swire_scoords = astropy.coordinates.SkyCoord(ra=swire_coords[:, 0],
dec=swire_coords[:, 1],
unit='deg')
with astropy.io.fits.open(pipeline.MIDDELBERG_TABLE4_PATH) as elais_components_fits:
elais_components = elais_components_fits[1].data
component_to_name = {}
for i, component in enumerate(elais_components):
name = component['ATELAIS']
id_to_atlas[component['CID']] = name
atlas_to_id[name] = component['CID']
atlas_to_zooniverse_id[name] = ''
key_to_atlas[i] = name
coord = astropy.coordinates.SkyCoord(
ra='{} {} {}'.format(component['RAh'], component['RAm'], component['RAs']),
dec='-{} {} {}'.format(component['DEd'], component['DEm'], component['DEs']),
unit=('hourangle', 'deg'))
coord = (coord.ra.deg, coord.dec.deg)
atlas_to_ras[name] = coord[0]
atlas_to_decs[name] = coord[1]
# Load SWIRE cross-identification from Table 5.
with open(pipeline.MIDDELBERG_TABLE5_PATH) as elais_file:
lines = [line.split('|') for line in elais_file]
for line in lines:
if 'ATELAISJ' not in line[0]:
continue
line_cids = line[1]
if 'C0' not in line_cids and 'C1' not in line_cids:
continue
line_cids = [cid.strip() for cid in line_cids.split(',')]
swire_coord_re = re.search(r'SWIRE4J(\d\d)(\d\d)(\d\d\.\d\d)(-\d\d)(\d\d)(\d\d\.\d)', line[2])
if not swire_coord_re:
continue
swire_coord_list = swire_coord_re.groups()
coord = astropy.coordinates.SkyCoord(
ra='{} {} {}'.format(*swire_coord_list[:3]),
dec='{} {} {}'.format(*swire_coord_list[3:]),
unit=('hourangle', 'deg'))
# Nearest SWIRE...
seps = coord.separation(swire_scoords)
nearest = numpy.argmin(seps)
dist = seps[nearest]
if dist.deg > 5 / 60 / 60:
continue
name = swire_names[nearest]
for cid in line_cids:
atlas_to_swire_expert[id_to_atlas[cid]] = name
atlas_to_rgz = {}
atlas_to_radio_consensus = {}
atlas_to_ir_consensus = {}
if field == 'cdfs':
for row in astropy.io.ascii.read(pipeline.RGZ_PATH):
name = id_to_atlas[row['atlas_id']]
atlas_to_radio_consensus[name] = row['consensus.radio_level']
atlas_to_ir_consensus[name] = row['consensus.ir_level']
atlas_to_rgz[name] = row['SWIRE.designation']
titlemap = {
'RGZ & Norris & compact': 'Compact',
'RGZ & Norris & resolved': 'Resolved',
'RGZ & Norris': 'All',
'RGZ & compact': 'Compact',
'RGZ & resolved': 'Resolved',
'RGZ': 'All',
}
known_predictors = set()
for cid in cids:
if cid.labeller == 'norris' and 'Norris' not in cid.dataset_name:
continue
if cid.classifier in {'Groundtruth', 'Random'}:
continue
if field == 'cdfs':
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES['RGZ & Norris'], cid.quadrant].nonzero()[0]
else:
atlas_keys = atlas_test_sets[:, 0, 0].nonzero()[0]
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
n_total = 0
n_correct = 0
n_skipped = 0
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
continue
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
predictor_name = '{}({} / {})'.format(
{'LogisticRegression': 'LR', 'CNN': 'CNN', 'RandomForestClassifier': 'RF'}[cid.classifier],
labeller, titlemap[cid.dataset_name])
known_predictors.add(predictor_name)
for i in atlas_keys:
name = key_to_atlas[i]
swire_predictor = atlas_to_swire_predictor.get(name, '')
atlas_to_swire[name][predictor_name] = swire_predictor
known_predictors = sorted(known_predictors)
atlases = sorted(atlas_to_swire)
ras = []
decs = []
expert_xids = []
expert_xid_ras = []
expert_xid_decs = []
rgzs = []
rgz_ras = []
rgz_decs = []
rcs = []
ircs = []
cids = []
zids = []
predictor_columns = collections.defaultdict(list)
predictor_ras = collections.defaultdict(list)
predictor_decs = collections.defaultdict(list)
for atlas in atlases:
for predictor in known_predictors:
predictor_columns[predictor].append(atlas_to_swire[atlas].get(predictor, ''))
predictor_ras[predictor].append(swire_name_to_coord.get(atlas_to_swire[atlas].get(predictor, ''), (None, None))[0])
predictor_decs[predictor].append(swire_name_to_coord.get(atlas_to_swire[atlas].get(predictor, ''), (None, None))[1])
ras.append(atlas_to_ras[atlas])
decs.append(atlas_to_decs[atlas])
rgzs.append(atlas_to_rgz.get(atlas, ''))
cids.append(atlas_to_id[atlas])
zids.append(atlas_to_zooniverse_id[atlas])
rgz_ras.append(swire_name_to_coord.get(atlas_to_rgz.get(atlas, ''), (None, None))[0])
rgz_decs.append(swire_name_to_coord.get(atlas_to_rgz.get(atlas, ''), (None, None))[1])
rcs.append(atlas_to_radio_consensus.get(atlas, 0.0))
ircs.append(atlas_to_ir_consensus.get(atlas, 0.0))
expert_xids.append(atlas_to_swire_expert.get(atlas, ''))
expert_xid_ras.append(swire_name_to_coord.get(atlas_to_swire_expert.get(atlas, ''), (None, None))[0])
expert_xid_decs.append(swire_name_to_coord.get(atlas_to_swire_expert.get(atlas, ''), (None, None))[1])
expert = 'Norris' if field == 'cdfs' else 'Middelberg'
table = astropy.table.Table(
data=[atlases, ras, decs,
cids, zids,
expert_xids, expert_xid_ras, expert_xid_decs,
rgzs, rgz_ras, rgz_decs,
rcs, ircs] + [k for p in known_predictors for k in (predictor_columns[p], predictor_ras[p], predictor_decs[p])],
names=['ATLAS', 'RA', 'Dec',
'CID', 'Zooniverse ID',
expert, expert + ' RA', expert + ' Dec',
'RGZ', 'RGZ RA', 'RGZ Dec',
'RGZ radio consensus', 'RGZ IR consensus'] + [k for p in known_predictors for k in (p, p + ' RA', p + ' Dec')])
table['RGZ radio consensus'].format = '{:.4f}'
table['RGZ IR consensus'].format = '{:.4f}'
table.write('/Users/alger/data/Crowdastro/predicted_cross_ids_table_21_03_18_{}.csv'.format(field), format='csv')
table.write('/Users/alger/data/Crowdastro/predicted_cross_ids_table_21_03_18_{}.tex'.format(field), format='latex')
def main(classifier='CNN', labeller='Norris'):
# Load SWIRE stuff.
swire_names, swire_coords, swire_features = pipeline.generate_swire_features(overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False)
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, overwrite=False)
swire_tree = KDTree(swire_coords)
swire_name_to_index = {n: i for i, n in enumerate(swire_names)}
atlas_names = []
atlas_compactnesses = []
atlas_coords = []
atlas_norris_swire = []
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
if not row['Component Zooniverse ID (RGZ)']:
continue
compactness = pipeline.compactness(row)
atlas_names.append(name)
atlas_compactnesses.append(compactness)
atlas_coords.append((row['Component RA (Franzen)'], row['Component DEC (Franzen)']))
atlas_norris_swire.append(row['Source SWIRE (Norris)'])
ys = []
xs_entropy = []
xs_margin = []
no_groundtruth = []
correct = []
for name, compactness, coords, swire in zip(atlas_names, atlas_compactnesses, atlas_coords, atlas_norris_swire):
predictor_name = '{}_{}'.format(classifier, labeller)
predictions = get_predictions(swire_tree, swire_coords, numpy.array(swire_names), swire_test_sets, coords, predictor_name)
if not predictions:
print('No predictions for {}'.format(name))
continue
chosen_swire = predictions[numpy.argmax([p for _, p in predictions])][0]
predictions = [p for _, p in predictions]
predictions_softmax = [numpy.exp(p) / sum(numpy.exp(p) for p in predictions) for p in predictions]
if len(predictions_softmax) == 1:
entropy_ambiguity = 0
margin_ambiguity = 0
else:
entropy_ambiguity = -sum(p * numpy.log(p) for p in predictions_softmax if p)
predictions.sort()
margin_ambiguity = 1 - (predictions[-1] - predictions[-2])
ys.append(compactness)
xs_entropy.append(entropy_ambiguity)
xs_margin.append(margin_ambiguity)
no_groundtruth.append(not swire or not swire.startswith('SWIRE'))
correct.append(swire == chosen_swire)
ys = numpy.array(ys)
xs_margin = numpy.array(xs_margin)
xs_entropy = numpy.array(xs_entropy)
no_groundtruth = numpy.array(no_groundtruth, dtype=bool)
correct = numpy.array(correct, dtype=bool)
print(sum(1 for y in ys if y <= 1))
plt.subplot(1, 2, 1)
plt.scatter(xs_margin[no_groundtruth], ys[no_groundtruth], marker='x', color='black', alpha=0.05)
plt.scatter(xs_margin[~no_groundtruth & correct], ys[~no_groundtruth & correct], marker='x', color='blue', alpha=0.7)
plt.scatter(xs_margin[~no_groundtruth & ~correct], ys[~no_groundtruth & ~correct], marker='x', color='magenta', alpha=0.7)
plt.title('Margin')
plt.xlabel('1 - margin')
plt.ylabel('$1.3 SNR S / 10 S_p$')
plt.yscale('log')
plt.axhline(1, min(xs_margin), max(xs_margin))
plt.subplot(1, 2, 2)
plt.scatter(xs_entropy[no_groundtruth], ys[no_groundtruth], marker='x', color='black', alpha=0.05)
plt.scatter(xs_entropy[~no_groundtruth & correct], ys[~no_groundtruth & correct], marker='x', color='blue', alpha=0.7)
plt.scatter(xs_entropy[~no_groundtruth & ~correct], ys[~no_groundtruth & ~correct], marker='x', color='magenta', alpha=0.7)
plt.title('Entropy')
plt.xlabel('Entropy')
plt.ylabel('$1.3 SNR S / 10 S_p$')
plt.yscale('log')
plt.axhline(1, min(xs_entropy), max(xs_entropy), zorder=-100, linestyle='--', color='black')
plt.show()
def get_examples():
titlemap = {
'RGZ & Norris & compact': 'Compact',
'RGZ & Norris & resolved': 'Resolved',
'RGZ & Norris': 'All',
'RGZ & compact': 'Compact',
'RGZ & resolved': 'Resolved',
'RGZ': 'All',
}
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names, overwrite=False)
(_, atlas_test_sets), (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, overwrite=False)
cids = list(pipeline.cross_identify_all(swire_names, swire_coords, swire_test_sets, swire_labels[:, 0]))
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_swire_norris = {}
key_to_atlas = {}
atlas_to_ras = {}
atlas_to_decs = {}
id_to_atlas = {}
atlas_to_zid = {}
atlas_to_id = {}
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
id_to_atlas[row['Component ID (Franzen)']] = name
key_to_atlas[row['Key']] = name
swire = row['Source SWIRE (Norris)']
atlas_to_swire_norris[name] = swire
atlas_to_id[name] = row['Component ID (Franzen)']
atlas_to_ras[name] = row['Component RA (Franzen)']
atlas_to_decs[name] = row['Component DEC (Franzen)']
atlas_to_zid[name] = row['Component Zooniverse ID (RGZ)']
atlas_to_rgz = {}
atlas_to_radio_consensus = {}
atlas_to_ir_consensus = {}
for row in astropy.io.ascii.read(pipeline.RGZ_PATH):
name = id_to_atlas[row['atlas_id']]
atlas_to_radio_consensus[name] = row['consensus.radio_level']
atlas_to_ir_consensus[name] = row['consensus.ir_level']
atlas_to_rgz[name] = row['SWIRE.designation']
cross_identifications = collections.defaultdict(dict) # ATLAS -> labeller -> SWIRE
for cid in cids:
if cid.labeller == 'norris' and 'Norris' not in cid.dataset_name:
continue
atlas_to_swire_predictor = dict(zip(cid.radio_names, cid.ir_names))
# For each ATLAS object in RGZ & Norris...
atlas_keys = atlas_test_sets[:, pipeline.SET_NAMES['RGZ & Norris'], cid.quadrant].nonzero()[0]
if 'Norris' in cid.dataset_name and cid.labeller == 'rgz':
labeller = 'RGZ N'
elif cid.labeller == 'rgz':
labeller = 'RGZ'
else:
labeller = 'Norris'
for i in atlas_keys:
name = key_to_atlas[i]
if name not in atlas_to_swire_norris:
continue
if name not in atlas_to_swire_predictor:
continue
cross_identifications[name]['Norris'] = atlas_to_swire_norris[name]
cross_identifications[name]['RGZ'] = atlas_to_rgz.get(name, None)
cross_identifications[name][labeller, cid.classifier, titlemap[cid.dataset_name]] = atlas_to_swire_predictor[name]
"""
For each classifier, pull out examples where:
- RGZ and Norris agree, but the classifier disagrees.
Only include RGZ & Norris dataset ("All").
"""
classifier_to_example = collections.defaultdict(set)
for atlas, cids_ in cross_identifications.items():
if cids_['Norris'] != cids_['RGZ']:
continue
for classifier, swire in cids_.items():
if classifier[2] != 'All' or classifier[1] in {'Random', 'Groundtruth'}:
continue
if swire != cids_['Norris']:
classifier_to_example[classifier].add((atlas, atlas_to_zid[atlas], atlas_to_id[atlas]))
return classifier_to_example
def plot_predictions(cut=0.95, labeller='norris', dataset_name=None, classifier=None):
"""Plot colour-colour diagram for predicted host galaxies.
labeller in {'norris', 'rgz'}
dataset_name in {'RGZ & Norris', ...}
"""
with h5py.File(CROWDASTRO_PATH, 'r') as f:
swire_numeric_cdfs = f['/swire/cdfs/numeric'][:, 2:2 + 4]
f_36 = swire_numeric_cdfs[:, 0]
f_45 = swire_numeric_cdfs[:, 1]
f_58 = swire_numeric_cdfs[:, 2]
f_80 = swire_numeric_cdfs[:, 3]
detection_58 = (f_58 != -99)
detection_80 = (f_80 != -99)
p = pipeline.unserialise_predictions(
pipeline.WORKING_DIR + '{}_{}_cdfs_predictions'.format(classifier, labeller))
predictions = {}
for i in p:
predictions[i.dataset_name, i.quadrant] = i
swire_names, swire_coords, _ = pipeline.generate_swire_features(overwrite=False, field='cdfs')
swire_labels = pipeline.generate_swire_labels(swire_names, swire_coords, overwrite=False, field='cdfs')
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords, swire_labels, overwrite=False, field='cdfs')
xs = []
ys = []
colours = []
for q in range(4):
swire_set = swire_test_sets[:, pipeline.SET_NAMES['RGZ'], q]
if labeller == 'norris' and not dataset_name:
# predictions_set = predictions['RGZ & Norris', q].probabilities > cut
f_36_ = f_36[swire_set & swire_labels[:, 0]]#[predictions_set]
f_45_ = f_45[swire_set & swire_labels[:, 0]]#[predictions_set]
f_58_ = f_58[swire_set & swire_labels[:, 0]]#[predictions_set]
f_80_ = f_80[swire_set & swire_labels[:, 0]]#[predictions_set]
elif labeller == 'rgz' and not dataset_name:
f_36_ = f_36[swire_set & swire_labels[:, 1]]
f_45_ = f_45[swire_set & swire_labels[:, 1]]
f_58_ = f_58[swire_set & swire_labels[:, 1]]
f_80_ = f_80[swire_set & swire_labels[:, 1]]
if labeller == 'norris' and dataset_name:
predictions_set = predictions[dataset_name, q].probabilities > cut
f_36_ = f_36[swire_set][predictions_set]
f_45_ = f_45[swire_set][predictions_set]
f_58_ = f_58[swire_set][predictions_set]
f_80_ = f_80[swire_set][predictions_set]
probabilities = predictions[dataset_name, q].probabilities[predictions_set]
detection_58_ = (f_58_ != -99)
detection_80_ = (f_80_ != -99)
detection_all_ = detection_58_ & detection_80_
ratio_58_36 = numpy.log10(f_58_[detection_all_] / f_36_[detection_all_])
ratio_80_45 = numpy.log10(f_80_[detection_all_] / f_45_[detection_all_])
probabilities = probabilities[detection_all_]
xs.extend(ratio_58_36)
ys.extend(ratio_80_45)
colours.extend(probabilities)
assert len(xs) == len(ys)
assert len(xs) == len(colours)
plot_basic()
if dataset_name:
plt.scatter(xs, ys, s=20, marker='^', linewidth=0, alpha=0.5, c=numpy.array(colours), cmap='winter')
else:
plt.scatter(xs, ys, s=25, c='r', marker='^', linewidth=0)
plt.xlim((-0.75, 1.0))
plt.ylim((-0.75, 1.0))
plt.xlabel('$\\log_{10}(S_{5.8}/S_{3.6})$')
plt.ylabel('$\\log_{10}(S_{8.0}/S_{4.5})$')
plt.subplots_adjust(left=0.2, bottom=0.15, right=0.95, top=0.95)
plt.colorbar()
plt.show()
def plot_grid(field='cdfs'):
# Load predictions.
lr_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR +
'LogisticRegression_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR +
'LogisticRegression_rgz_{}_predictions'.format(field)))
rf_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR +
'RandomForestClassifier_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR +
'RandomForestClassifier_rgz_{}_predictions'.format(field)))
cnn_predictions = itertools.chain(
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_norris_{}_predictions'.format(field)),
pipeline.unserialise_predictions(
pipeline.WORKING_DIR + 'CNN_rgz_{}_predictions'.format(field)))
# Convert to the format we need. e.g. {'RGZ' -> [acc, acc, acc, acc]}
lr_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
lr_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
rf_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
rf_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
cnn_norris_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
cnn_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
for predictions in lr_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
lr_norris_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
else:
lr_rgz_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
for predictions in rf_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
rf_norris_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
else:
rf_rgz_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
for predictions in cnn_predictions:
dataset_name = predictions.dataset_name
if predictions.labeller == 'norris':
cnn_norris_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
else:
cnn_rgz_accuracies[dataset_name][
predictions.quadrant] = predictions.balanced_accuracy
if field == 'cdfs':
# Load RGZ cross-identifications and compute a balanced accuracy with them.
swire_names, swire_coords, _ = pipeline.generate_swire_features(
overwrite=False, field=field)
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False,
field=field)
(_, atlas_test_sets), (_,
swire_test_sets) = pipeline.generate_data_sets(
swire_coords,
swire_labels,
overwrite=False,
field=field)
label_rgz_accuracies = {sstr: [0] * 4 for sstr in pipeline.SET_NAMES}
label_norris_accuracies = {
sstr: [1] * 4
for sstr in pipeline.SET_NAMES
} # By definition.
for dataset_name in pipeline.SET_NAMES:
for quadrant in range(4):
test_set = swire_test_sets[:, pipeline.SET_NAMES[dataset_name],
quadrant]
predictions = swire_labels[test_set, 1]
trues = swire_labels[test_set, 0]
ba = balanced_accuracy(trues, predictions)
label_rgz_accuracies[dataset_name][quadrant] = ba
colours = ['grey', 'magenta', 'blue', 'orange']
markers = ['o', '^', 'x', 's']
handles = {}
plt.figure(figsize=(5, 5))
accuracy_map = defaultdict(lambda: defaultdict(dict)) # For table output.
output_sets = [
('LR', [lr_norris_accuracies, lr_rgz_accuracies]),
('CNN', [cnn_norris_accuracies, cnn_rgz_accuracies]),
('RF', [rf_norris_accuracies, rf_rgz_accuracies]),
]
if field == 'cdfs':
output_sets.append(
('Labels', [label_norris_accuracies, label_rgz_accuracies]))
for j, (classifier_name, classifier_set) in enumerate(output_sets):
for i, set_name in enumerate(norris_labelled_sets):
if 'compact' not in set_name: # Skip compact.
ax = plt.subplot(2, 1, {
'RGZ & Norris & resolved': 1,
'RGZ & Norris': 2
}[set_name])
ax.set_ylim((80, 100))
ax.set_xlim((-0.5, 1.5))
ax.set_xticks([0, 1]) #, 2])
ax.set_xticklabels(
[
'Norris',
# 'RGZ N',
'RGZ',
],
rotation='horizontal')
if i == 2:
plt.xlabel('Labels')
plt.ylabel('{}\nBalanced accuracy\n(per cent)'.format(
titlemap[set_name]))
ax.title.set_fontsize(16)
ax.xaxis.label.set_fontsize(12)
ax.yaxis.label.set_fontsize(9)
for tick in ax.get_xticklabels() + ax.get_yticklabels():
tick.set_fontsize(10)
ax.grid(which='major', axis='y', color='#EEEEEE')
for k in range(4):
if 'compact' in set_name:
continue
if j != 3: # !Labels
ax.scatter([0 + (j - 1) / 5],
classifier_set[0][set_name][k] * 100,
color=colours[j],
marker=markers[j],
linewidth=1,
edgecolor='k')
rgz_offset = ((j - 1.5) /
6) if field == 'cdfs' else (j - 1) / 5
handles[j] = ax.scatter(
[1 + rgz_offset],
classifier_set[1][fullmap[set_name]][k] * 100,
color=colours[j],
marker=markers[j],
linewidth=1,
edgecolor='k')
# ax.scatter([1 + (j - 1) / 5], classifier_set[1][set_name][k] * 100,
# color=colours[j], marker=markers[j], linewidth=1, edgecolor='k')
# Compute for table.
for labeller in ['Norris', 'RGZ N', 'RGZ']:
if labeller == 'Norris':
mean = numpy.mean(classifier_set[0][set_name]) * 100
stdev = numpy.std(classifier_set[0][set_name]) * 100
elif labeller == 'RGZ N':
continue
# mean = numpy.mean(classifier_set[1][set_name]) * 100
# stdev = numpy.std(classifier_set[1][set_name]) * 100
elif labeller == 'RGZ':
mean = numpy.mean(
classifier_set[1][fullmap[set_name]]) * 100
stdev = numpy.std(
classifier_set[1][fullmap[set_name]]) * 100
accuracy_map[labeller][classifier_name][
titlemap[set_name]] = '${:.02f} \\pm {:.02f}$'.format(
mean, stdev)
# Assemble table.
col_labeller = []
col_classifier = []
col_compact = []
col_resolved = []
col_all = []
for labeller in ['Norris', 'RGZ N', 'RGZ']:
if labeller == 'RGZ N':
continue
for classifier in ['CNN', 'LR', 'RF'] + ['Labels'
] if field == 'cdfs' else []:
col_labeller.append(labeller)
col_classifier.append(classifier)
col_compact.append(accuracy_map[labeller][classifier]['Compact'])
col_resolved.append(accuracy_map[labeller][classifier]['Resolved'])
col_all.append(accuracy_map[labeller][classifier]['All'])
out_table = astropy.table.Table(
[col_labeller, col_classifier, col_compact, col_resolved, col_all],
names=[
'Labeller', 'Classifier', "Mean `Compact' accuracy\\\\(per cent)",
"Mean `Resolved' accuracy\\\\(per cent)",
"Mean `All' accuracy\\\\(per cent)"
])
out_table.write('../{}_accuracy_table.tex'.format(field), format='latex')
plt.figlegend([handles[j] for j in sorted(handles)], ['LR', 'CNN', 'RF'] +
(['Labels'] if field == 'cdfs' else []),
'lower center',
ncol=4,
fontsize=10)
plt.subplots_adjust(bottom=0.2, hspace=0.25)
plt.savefig('../images/{}_ba_grid.pdf'.format(field),
bbox_inches='tight',
pad_inches=0)
plt.savefig('../images/{}_ba_grid.png'.format(field),
bbox_inches='tight',
pad_inches=0)
def main(examples=None, classifier='CNN', labeller='Norris'):
# Load SWIRE stuff.
swire_names, swire_coords, swire_features = pipeline.generate_swire_features(
overwrite=False)
swire_labels = pipeline.generate_swire_labels(swire_names,
swire_coords,
overwrite=False)
_, (_, swire_test_sets) = pipeline.generate_data_sets(swire_coords,
overwrite=False)
swire_tree = KDTree(swire_coords)
swire_name_to_index = {n: i for i, n in enumerate(swire_names)}
# Load ATLAS coords.
table = astropy.io.ascii.read(pipeline.TABLE_PATH)
atlas_to_coords = {}
atlas_to_swire_coords = {}
for row in table:
name = row['Component Name (Franzen)']
if not name:
continue
atlas_to_coords[name] = row['Component RA (Franzen)'], row[
'Component DEC (Franzen)']
index = swire_name_to_index.get(row['Source SWIRE (Norris)'] or '')
if index:
atlas_to_swire_coords[name] = swire_coords[index]
ir_stretch = astropy.visualization.LogStretch(0.001)
if examples is None:
examples = examples_incorrect.get_examples()
examples = examples[labeller, classifier, 'All']
for example in examples:
print('Plotting {}'.format(example))
predictor_name = '{}_{}'.format(classifier, labeller)
cid = example[2]
# Load FITS stuff.
try:
radio_fits = astropy.io.fits.open(CDFS_PATH + cid + '_radio.fits')
except FileNotFoundError:
if example[1]: # Has Zooniverse ID
print('{} not in RGZ'.format(cid))
continue
ir_fits = astropy.io.fits.open(CDFS_PATH + cid + '_ir.fits')
wcs = astropy.wcs.WCS(radio_fits[0].header)
# Compute info for contour levels. (also from Enno Middelberg)
median = numpy.median(radio_fits[0].data)
mad = numpy.median(numpy.abs(radio_fits[0].data - median))
sigma = mad / mad2sigma
# Set up the plot.
fig = plt.figure()
ax = astropy.visualization.wcsaxes.WCSAxes(fig, [0.1, 0.1, 0.8, 0.8],
wcs=wcs)
fig.add_axes(ax)
ax.set_title('{}'.format(example[0], example[1]))
# Show the infrared.
ax.imshow(ir_stretch(ir_fits[0].data),
cmap='cubehelix_r',
origin='lower')
# Show the radio.
ax.contour(radio_fits[0].data,
colors='black',
levels=[nsig * sigma * sigmult**i for i in range(15)],
linewidths=1,
origin='lower',
zorder=1)
# Plot predictions.
predictions = get_predictions(swire_tree, swire_coords,
swire_test_sets,
atlas_to_coords[example[0]],
predictor_name)
if not predictions:
print('No predictions for {}'.format(example[0]))
continue
coords = [p[0] for p in predictions]
probabilities = [p[1] for p in predictions]
coords = wcs.all_world2pix(coords, 1)
ax.scatter(coords[:, 0],
coords[:, 1],
s=numpy.sqrt(numpy.array(probabilities)) * 200,
color='white',
edgecolor='black',
linewidth=1,
alpha=0.9,
marker='o',
zorder=2)
choice = numpy.argmax(probabilities)
ax.scatter(coords[choice, 0],
coords[choice, 1],
s=200 / numpy.sqrt(2),
color='blue',
marker='x',
zorder=2.5)
try:
norris_coords, = wcs.all_world2pix(
[atlas_to_swire_coords[example[0]]], 1)
except KeyError:
print('No Norris cross-identification for {}'.format(example[0]))
continue
ax.scatter(norris_coords[0],
norris_coords[1],
marker='+',
s=200,
zorder=3,
color='green')
lon, lat = ax.coords
lon.set_major_formatter('hh:mm:ss')
lon.set_axislabel('Right Ascension')
lat.set_axislabel('Declination')
fn = '{}_{}_{}'.format(classifier, labeller, example[0])
plt.savefig(
'/Users/alger/repos/crowdastro-projects/ATLAS-CDFS/images/examples/'
+ fn + '.png',
bbox_inches='tight',
pad_inches=0)
plt.savefig(
'/Users/alger/repos/crowdastro-projects/ATLAS-CDFS/images/examples/'
+ fn + '.pdf',
bbox_inches='tight',
pad_inches=0)
plt.clf()
