def __init__(self, img_paths, outfile_name=("data/align/fundus_photo_aligner_%s.csv" % str(uuid.uuid4())[:8])):
path = '/'.join(img_paths[0].split('/')[:-1])
img_names = [img_path.split('/')[-1] for img_path in img_paths]
ext = img_names[0].split('.')[-1]
ids = set([img_path.split('_')[0] for img_path in img_names])
assert((len(ids) * 2) == len(img_names))
grouped_img_names = [["%s_left" % id, "%s_right" % id] for id in ids]
self.labels = {}
print ''.join([" %i%%|" % (j*10) for j in xrange(1,10)] + [' 100%|'])
pb = NyanBar(tasks=len(grouped_img_names))
with open(outfile_name, 'w') as csvfile:
writer = csv.writer(csvfile)
writer.writerow(['image','horizontal_flip','vertical_flip'])
for img_names in grouped_img_names:
img_paths = ["%s/%s.%s" % (path, name, ext) for name in img_names]
fundus_photos = [FundusPhoto(img_path) for img_path in img_paths]
for i, fundus_photo in enumerate(fundus_photos):
fundus_photo.refine_center()
direct = fundus_photo.has_notch()
lens_flip_vector = np.array([0,0] if direct else [1,1])
left_right_flip_vector = np.array([0,1] if i == 0 else [0,0])
final_flip_vector = (lens_flip_vector + left_right_flip_vector) % 2
writerow = [img_names[i]] + final_flip_vector.tolist()
writer.writerow(writerow)
self.labels[img_names[i]] = direct
pb.task_done()
pb.finish()
print("Done! created: %s" % outfile_name)
def benchmark_put_with_many_filters():
hosts = ["bloom1", "bloom2", "bloom3", "bloom4"]
client = BloomRouter(hosts, "g{}".format(random.randint(1, 100000), filter_count=16))
progress = NyanBar(tasks=testsize)
for i in range(testsize):
progress.task_done()
keys = [str(uuid.uuid4()) for _ in range(num_keys)]
client.add(keys)
progress.finish()
def put_then_get_with_one_filter():
hosts = ["bloom1", "bloom2", "bloom3", "bloom4"]
client = BloomRouter(hosts, "f{}".format(random.randint(1, 100000)), filter_count=1)
progress = NyanBar(tasks=testsize)
for i in range(testsize):
progress.task_done()
keys = [str(uuid.uuid4()) for _ in range(num_keys)]
client.add(keys)
assert client.get(keys)
progress.finish()
def package_data(directory, datasets, image_shape, outfile_path):
"""
Outputs a *.npz file of image data in 3 sets
:type directory: string
:param directory: directory of image data
:type datasets: tuple of dictionaries, string -> Any
:param datasets: dictionaries should be in the order
training, validation, test. The key is the image
filename and the value is the true label
:type image_shape: tuple of ints
:type outfile_path: string
"""
assert(len(datasets) == 3)
xsize = numpy.prod(image_shape)
x_datasets = [numpy.zeros((len(dataset), xsize), dtype=numpy.uint8) for dataset in datasets]
y_datasets = [numpy.array(dataset.values(), dtype=numpy.uint8) for dataset in datasets]
print "| " + ("⚐ ⚑ " * 19) + "-|"
pb = NyanBar(tasks=sum([len(dataset) for dataset in datasets]))
for j, dataset in enumerate(datasets):
for i, image_name in enumerate(dataset.keys()):
pb.task_done()
im = Image.open(directory + image_name)
x_datasets[j][i, :] = numpy.array(im.getdata(), dtype=numpy.uint8).reshape(xsize)
pb.finish()
print '... saving data'
# cPickle too slow (takes many minutes for tens of thousands of images over 100x100x3)
saveme = [x_datasets[0], y_datasets[0], x_datasets[1], y_datasets[1], x_datasets[2], y_datasets[2]]
numpy.savez(open(outfile_path, 'wb'), *saveme)
print 'done'
with open(label_csv, 'rU') as csvfile:
reader = csv.reader(csvfile, dialect=csv.excel_tab, delimiter=',')
next(reader, None) # skip header
print ''.join([" %i%%|" % (j*10) for j in xrange(1,10)] + [' 100%|'])
pb = NyanBar(tasks=(sum(1 for row in open(label_csv, 'rU'))-1))
for row in reader:
imgname = row[0]
true_label = 1 if (row[1] == 'tr' or row[1] == 'br' or row[1] == '1') else 0
title_prefix = "True Label: %s" % true_label
predicted_label = predict_single_image(inpath + imgname, title_prefix=title_prefix, save_path=(outpath + imgname), true_label=true_label)
true_labels.append(true_label)
predicted_labels.append(predicted_label)
pb.task_done()
pb.finish()
accuracy, precision = binary_accuracy_precision(true_labels, predicted_labels)
print("Accuracy: %.3f. Precision: %.3f." % (accuracy, precision))
elif len(sys.argv) == 3: # batch of doubled images
true_label_csv = sys.argv[1]
inpath = sys.argv[2]
true_labels = []
img_names = [] # for reference of label order
img_paths = []
# read csv
# we're going to assume true_label_csv has the double images next to each other
# whether left/right comes first doesn't matter since their label is identical
with open(true_label_csv, 'rU') as csvfile:
reader = csv.reader(csvfile, dialect=csv.excel_tab, delimiter=',')
