def extract_unique_craters(CP, craters_unique):
"""Top level function that extracts craters from model predictions,
converts craters from pixel to real (degree, km) coordinates, and filters
out duplicate detections across images.
Parameters
----------
CP : dict
Crater Parameters needed to run the code.
craters_unique : array
Empty master array of unique crater tuples in the form
(long, lat, radius).
Returns
-------
craters_unique : array
Filled master array of unique crater tuples.
"""
# Load/generate model preds
try:
preds = h5py.File(CP['dir_preds'], 'r')[CP['datatype']]
print("Loaded model predictions successfully")
except:
print("Couldnt load model predictions, generating")
preds = get_model_preds(CP)
# need for long/lat bounds
P = h5py.File(CP['dir_data'], 'r')
llbd, pbd, distcoeff = ('longlat_bounds', 'pix_bounds',
'pix_distortion_coefficient')
#r_moon = 1737.4
dim = (float(CP['dim']), float(CP['dim']))
N_matches_tot = 0
for i in range(CP['start_of_images'],
CP['start_of_images'] + CP['n_imgs']):
id = proc.get_id(i)
coords = tmt.template_match_t(preds[i])
# convert, add to master dist
if len(coords) > 0:
new_craters_unique = estimate_longlatdiamkm(
dim, P[llbd][id], P[distcoeff][id][0], coords)
N_matches_tot += len(coords)
# Only add unique (non-duplicate) craters
if len(craters_unique) > 0:
craters_unique = add_unique_craters(new_craters_unique,
craters_unique, CP['llt2'],
CP['rt2'])
else:
craters_unique = np.concatenate(
(craters_unique, new_craters_unique))
np.save(CP['dir_result'], craters_unique)
return craters_unique
def get_data(CP):
"""Reads in or generates model predictions.
Parameters
----------
CP : dict
Containins directory locations for loading data and storing
predictions.
Returns
-------
craters : h5py
Model predictions.
"""
n_imgs, dtype = CP['n_imgs'], CP['datatype']
data = h5py.File(CP['dir_data'], 'r')
Data = {
dtype: [
data['input_images'][:n_imgs].astype('float32'),
data['target_masks'][:n_imgs].astype('float32')
]
}
data.close()
craters = pd.HDFStore(CP['crater_data'], 'r')
csvs = []
minrad, maxrad, cutrad, n_csvs, dim = 3, 50, 0.8, len(craters), 256
diam = 'Diameter (pix)'
for i in range(n_csvs):
csv = craters[proc.get_id(i, 2)]
# remove small/large/half craters
csv = csv[(csv[diam] < 2 * maxrad) & (csv[diam] > 2 * minrad)]
csv = csv[(csv['x'] + cutrad * csv[diam] / 2 <= dim)]
csv = csv[(csv['y'] + cutrad * csv[diam] / 2 <= dim)]
csv = csv[(csv['x'] - cutrad * csv[diam] / 2 > 0)]
csv = csv[(csv['y'] - cutrad * csv[diam] / 2 > 0)]
if len(csv) < 1: # Exclude csvs with few craters
csvs.append([-1])
else:
csv_coords = np.asarray((csv['x'], csv['y'], csv[diam] / 2)).T
csvs.append(csv_coords)
return Data, csvs
def extract_unique_craters(CP, craters_unique):
"""Top level function that extracts craters from model predictions,
converts craters from pixel to real (degree, km) coordinates, and filters
out duplicate detections across images.
Parameters
----------
CP : dict
Crater Parameters needed to run the code.
craters_unique : array
Empty master array of unique crater tuples in the form
(long, lat, radius).
Returns
-------
craters_unique : array
Filled master array of unique crater tuples.
"""
# Load/generate model preds
try:
preds = h5py.File(CP['dir_preds'], 'r')[CP['datatype']]
print("Loaded model predictions successfully")
except:
print("Couldnt load model predictions, generating")
preds = get_model_preds(CP)
# need for long/lat bounds
P = h5py.File(CP['dir_data'], 'r')
llbd, pbd, distcoeff = ('longlat_bounds', 'pix_bounds',
'pix_distortion_coefficient')
#r_moon = 1737.4
dim = (float(CP['dim']), float(CP['dim']))
N_matches_tot = 0
for i in range(CP['start_of_images'],CP['start_of_images']+CP['n_imgs']):
id = proc.get_id(i)
coords = tmt.template_match_t(preds[i])
def get_metrics(data, craters, dim, model, beta=1):
"""Function that prints pertinent metrics at the end of each epoch.
Parameters
----------
data : hdf5
Input images.
craters : hdf5
Pandas arrays of human-counted crater data.
dim : int
Dimension of input images (assumes square).
model : keras model object
Keras model
beta : int, optional
Beta value when calculating F-beta score. Defaults to 1.
"""
X, Y = data[0], data[1]
dim =256
# Get csvs of human-counted craters
csvs = []
X = X[1:]
Y=Y[1:]
minrad, maxrad, cutrad, n_csvs = 1, 50, 0.8, len(X)
diam = 'Diameter (pix)'
for i in range(1,n_csvs):
try:
csv = craters[proc.get_id(i,4)]
except:
csvs.append([-1])
print ('Skipping iteration number =' ,i ,' as no crater is available in that area')
continue
# remove small/large/half craters
csv = csv[(csv[diam] < 2 * maxrad) & (csv[diam] > 2 * minrad)]
csv = csv[(csv['x'] + cutrad * csv[diam] / 2 <= dim)]
csv = csv[(csv['y'] + cutrad * csv[diam] / 2 <= dim)]
csv = csv[(csv['x'] - cutrad * csv[diam] / 2 > 0)]
csv = csv[(csv['y'] - cutrad * csv[diam] / 2 > 0)]
if len(csv) < 1: # Exclude csvs with few craters
csvs.append([-1])
else:
try:
csv_coords = np.asarray((csv['x'], csv['y'], csv[diam] / 2)).T
csvs.append(csv_coords)
except:
print('A fix for mars')
# Calculate custom metrics
print("")
print("*********Custom Loss*********")
recall, precision, fscore = [], [], []
frac_new, frac_new2, maxrad = [], [], []
err_lo, err_la, err_r = [], [], []
frac_duplicates = []
# print(len(csvs[1]))
# print(len(csvs[2]))
preds = model.predict(X)
for i in range(len(csvs)):
if len(csvs[i]) < 1:
continue
try:
if (csvs[i].count(-1) ==1):
print ('Skipping bcoz csvs ==-1')
continue
except:
print ('out of try block')
(N_match, N_csv, N_detect, maxr,
elo, ela, er, frac_dupes) = tmt.template_match_t2c(preds[i], csvs[i],
rmv_oor_csvs=0)
if N_match > 0:
p = float(N_match) / float(N_match + (N_detect - N_match))
r = float(N_match) / float(N_csv)
f = (1 + beta**2) * (r * p) / (p * beta**2 + r)
diff = float(N_detect - N_match)
fn = diff / (float(N_detect) + diff)
fn2 = diff / (float(N_csv) + diff)
recall.append(r)
precision.append(p)
fscore.append(f)
frac_new.append(fn)
frac_new2.append(fn2)
maxrad.append(maxr)
err_lo.append(elo)
err_la.append(ela)
err_r.append(er)
frac_duplicates.append(frac_dupes)
else:
print("skipping iteration %d,N_csv=%d,N_detect=%d,N_match=%d" %
(i, N_csv, N_detect, N_match))
print("binary XE score = %f" % model.evaluate(X, Y))
if len(recall) > 0:
print("mean and std of N_match/N_csv (recall) = %f, %f" %
(np.mean(recall), np.std(recall)))
print("""mean and std of N_match/(N_match + (N_detect-N_match))
(precision) = %f, %f""" % (np.mean(precision), np.std(precision)))
print("mean and std of F_%d score = %f, %f" %
(beta, np.mean(fscore), np.std(fscore)))
print("""mean and std of (N_detect - N_match)/N_detect (fraction
of craters that are new) = %f, %f""" %
(np.mean(frac_new), np.std(frac_new)))
print("""mean and std of (N_detect - N_match)/N_csv (fraction of
"craters that are new, 2) = %f, %f""" %
(np.mean(frac_new2), np.std(frac_new2)))
print("median and IQR fractional longitude diff = %f, 25:%f, 75:%f" %
(np.median(err_lo), np.percentile(err_lo, 25),
np.percentile(err_lo, 75)))
print("median and IQR fractional latitude diff = %f, 25:%f, 75:%f" %
(np.median(err_la), np.percentile(err_la, 25),
np.percentile(err_la, 75)))
print("median and IQR fractional radius diff = %f, 25:%f, 75:%f" %
(np.median(err_r), np.percentile(err_r, 25),
np.percentile(err_r, 75)))
print("mean and std of frac_duplicates: %f, %f" %
(np.mean(frac_duplicates), np.std(frac_duplicates)))
print("""mean and std of maximum detected pixel radius in an image =
%f, %f""" % (np.mean(maxrad), np.std(maxrad)))
print("""absolute maximum detected pixel radius over all images =
%f""" % np.max(maxrad))
print("")
def extract_unique_craters(CP, craters_unique):
"""Top level function that extracts craters from model predictions,
converts craters from pixel to real (degree, km) coordinates, and filters
out duplicate detections across images.
Parameters
----------
CP : dict
Crater Parameters needed to run the code.
craters_unique : array
Empty master array of unique crater tuples in the form
(long, lat, radius).
Returns
-------
craters_unique : array
Filled master array of unique crater tuples.
"""
# Load/generate model preds
try:
preds = h5py.File(CP['dir_preds'], 'r')[CP['datatype']]
print("Loaded model predictions successfully")
except:
print("Couldnt load model predictions, generating")
preds = get_model_preds(CP)
Data, Carters = get_data(CP)
# need for long/lat bounds
P = h5py.File(CP['dir_data'], 'r')
llbd, pbd, distcoeff = ('longlat_bounds', 'pix_bounds',
'pix_distortion_coefficient')
#r_moon = 1737.4
dim = (float(CP['dim']), float(CP['dim']))
N_matches_tot = 0
if not os.path.exists(CP['result_img']):
os.mkdir(CP['result_img'])
lenstr = ""
lenstr1 = "true_carter"
lenstr2 = "detect_carter"
lenstr3 = "undetected_carter"
num = 0
num1 = 0
num2 = 0
num3 = 0
for i in range(CP['n_imgs']):
id = proc.get_id(i, 2)
print("Drawing picture:%d" % i)
input_images = Data[CP['datatype']][0][i]
imgs = Image.fromarray(input_images.astype('uint8')).convert('RGB')
img = cv2.cvtColor(np.asarray(imgs), cv2.COLOR_RGB2BGR)
coords = tmt.template_match_t(preds[i])
num = num + len(coords)
lenstr = lenstr + " " + str(len(coords))
matplotlib.image.imsave(CP['result_img'] + "/" + str(i) + '_mask.jpg',
preds[i])
true_carter, detect_carter, Undetected_carter = get_coords_classification(
coords, Carters[i])
lenstr1 = lenstr1 + " " + str(len(true_carter))
num1 = num1 + len(true_carter)
lenstr2 = lenstr2 + " " + str(len(detect_carter))
num2 = num2 + len(detect_carter)
lenstr3 = lenstr3 + " " + str(len(Undetected_carter))
num3 = num3 + len(Undetected_carter)
draw_pic(img, coords, Carters[i],
CP['result_img'] + "/" + str(i) + '.jpg')
if len(coords) > 0:
# for i in range(len(coords)):
new_craters_unique = estimate_longlatdiamkm(
dim, P[llbd][id], P[distcoeff][id][0], coords)
N_matches_tot += len(coords)
#print(id,new_craters_unique)
# Only add unique (non-duplicate) craters
if len(craters_unique) > 0:
craters_unique = add_unique_craters(new_craters_unique,
craters_unique, CP['llt2'],
CP['rt2'])
else:
craters_unique = np.concatenate(
(craters_unique, new_craters_unique))
print(lenstr)
print("total num:%d" % num)
print(lenstr1)
print(num1)
print(lenstr2)
print(num2)
print(lenstr3)
print(num3)
np.save(CP['dir_result'], craters_unique)
return craters_unique
