"""
Script to extract CNN features using a caffe model.
Usage: python extract_features.py /path/to/input/image.jpg /path/to/output/features.txt
"""
from caffetools.extract.DeepFeatures import DeepFeatures
from caffetools.lmdb.numpyserializer import NumPySerializer
import argparse
# parse arguments
parser = argparse.ArgumentParser(description='Script to extract CNN features into LMDB')
parser.add_argument('input_image_path', help='path to input image')
parser.add_argument('output_features_path', help='path to output features text file')
args = parser.parse_args()
# extract features
d = DeepFeatures()
features = d.extract_features(args.input_image_path)
# save to text file
with open(args.output_features_path, 'w') as f:
features_row_vector = features.reshape(1,-1)
features_string = NumPySerializer.numpy_to_string(features_row_vector)
f.write(features_string)
print "Features saved to '%s'" % args.output_features_path
parser = argparse.ArgumentParser(description='Script to extract CNN features into numpy file')
parser.add_argument('images_list_path', help='path to text file of list of images')
parser.add_argument('images_path', help='path to base directory of images')
parser.add_argument('output_path', help='path to output numpy file')
args = parser.parse_args()
# check
if not os.path.exists(args.images_path):
raise OSError('images_path does not exist: {0}'.format(args.images_path))
# open images list
with open(args.images_list_path, 'r') as f:
images_list = [line.strip() for line in f]
# loop over images
d = DeepFeatures()
num_images = len(images_list)
data = None
for index, image in enumerate(tqdm(images_list)):
path_to_image = os.path.join(args.images_path, image)
# extract features
features = d.extract_features(path_to_image)
if not data:
feature_dim = features.shape[0]
data = np.zeros((num_images, feature_dim))
# put in numpy matrix
data[index, :] = features
# save to numpy file