def main(argv=None):
_init_output_directories()
# step 1: create the negatives and positives directory
if FLAGS.do_full_prepare:
print('Loading labels from %s' % FLAGS.label_file)
lr = LabelRecord()
label_records = lr.load(FLAGS.label_file)
all_bounding_boxes = Box.get_all_bounding_boxes(label_records)
counter = 0
# fill examples, originals, negatives, and positives directories
print('Processing images...')
for (_, v) in label_records.items():
Progress.show_progress(counter)
image = CXRImage.get_image_data(v.filename, FLAGS.image_path)
basefilename = os.path.splitext(v.filename)[0]
if v.hasBoundingBox:
for i in range(0, v.boundingBoxes.shape[0]):
box = v.boundingBoxes[i, :]
#CXRImage.extract_center_and_write(image,box,1024,1024,FLAGS.positives_dir)
CXRImage.extract_anisotropic_scale_and_write(
image, box, FLAGS.image_size, FLAGS.image_size,
FLAGS.positives_dir)
CXRImage.write_image(image, FLAGS.examples_dir,
"%s.jpg" % basefilename)
else:
i = np.int32(
np.random.randint(0, all_bounding_boxes.shape[0] - 1))
box = all_bounding_boxes[i, :]
CXRImage.extract_anisotropic_scale_and_write(
image, box, FLAGS.image_size, FLAGS.image_size,
FLAGS.negatives_dir)
#CXRImage.extract_center_and_write(image,box,1024,1024,FLAGS.negatives_dir)
if (v.hasBoundingBox):
img = (CXRImage.xlate_image(image) * 255).astype(np.uint8)
CXRImage.write_image_with_bounding_boxes(
img, FLAGS.originals_dir, "%s.jpg" % basefilename,
v.boundingBoxes)
counter = counter + 1
# step 2: create the pre-training features by combining negatives and positives into pre_train.tfrecord
print('\nCreating pre-train file...')
rec = Record(1024, 1024, 1 if FLAGS.grayscale else 3)
total = rec.create_pre_train_file(FLAGS.positives_dir, FLAGS.negatives_dir,
FLAGS.pre_train_file)
print('\n%d files combined in %s' % (total, FLAGS.pre_train_file))
