def main():
parser = argparse.ArgumentParser()
parser.add_argument('--caption_file',
'-cf',
type=str,
help='caption file',
default='sample_testing_text.txt')
parser.add_argument('--method_dir',
'-md',
type=str,
default='',
help='data directory')
parser.add_argument('--data_set',
'-ds',
type=str,
default='faces',
help='data set name')
parser.add_argument('--vector_type',
'-vt',
type=int,
default=4,
choices=(0, 8),
help='method to encode caption, options: '
'0. uni-skip, 1. bi-skip, 2. combine-skip, '
'3. one-hot, 4. glove')
parser.add_argument('--out_file',
'-of',
default='test_caption_vectors.hdf5',
type=str,
help='output file name')
parser.add_argument('--dict_file',
'-df',
default='onehot_hair_eyes.hdf5',
type=str,
help='input dictionary name')
parser.add_argument('--glove_file',
'-gf',
type=str,
help='glove file',
default='glove/glove.6B.300d.txt')
args = parser.parse_args()
if args.method_dir == '':
print('need to specify method_dir!')
exit(1)
with open(join(args.data_set, args.caption_file)) as f:
captions = f.read().splitlines()
captions = [cap for cap in captions]
captions = [cap.split(',') for cap in captions]
captions = dict([[cap[0], cap[1].split()]
for i, cap in enumerate(captions)])
if args.vector_type < 3:
for key, val in captions.items():
captions[key] = [
'the girl has ' + val[0] + ' ' + val[1] + ' and ' + val[2] +
' ' + val[3]
]
model = skipthoughts.load_model()
parallel_keys = [[] for i in range(cpus)]
quantity = (len(captions) + cpus - 1) // cpus
for i, key_val in enumerate(captions.items()):
parallel_keys[i // quantity].append(key_val[0])
out_q = Queue()
thrds =\
[Process(target=get_vector, args=(parallel_keys[i], model,
captions, out_q))
for i in range(cpus)]
for i, thrd in enumerate(thrds):
thrd.start()
caption_vectors = {}
for i in range(len(thrds)):
caption_vectors.update(out_q.get())
for i, thrd in enumerate(thrds):
thrd.join()
elif args.vector_type == 3:
h = h5py.File(join(args.data_dir, args.dict_file), 'r')
for key, val in captions.items():
new_val = [h['hair'][val[0]].value, h['eyes'][val[2]].value]
new_val = [
np.eye(h['hair'].attrs['size'])[new_val[0]],
np.eye(h['eyes'].attrs['size'])[new_val[1]]
]
caption_vectors[key] = new_val
elif args.vector_type == 4:
wordvecs = open(args.glove_file, 'r').read().splitlines()
wordvecs = [wordvec.split() for wordvec in wordvecs]
wordvecs = dict(
[[wordvec[0],
np.array([wordvec[1:]], dtype=np.float32)]
for wordvec in wordvecs])
caption_vectors = {}
for key, val in captions.items():
#new_val = [val[0], val[2]] if val[1] == 'hair' else [val[2], val[0]]
caption_vectors[key] = wordvecs[val[0]]
filename = join(args.data_set, args.method_dir, args.out_file)
if os.path.isfile(filename):
os.remove(filename)
h = h5py.File(filename, 'w')
for key in caption_vectors.keys():
h.create_dataset(key, data=caption_vectors[key])
h.close()
def save_caption(args):
img_dir = join(args.data_set, args.imgs_dir)
image_files = [f for f in os.listdir(img_dir) if 'jpg' in f]
print('number of train images: %d' % len(image_files))
tag_file = join(args.data_set, args.tag_name)
tags = open(tag_file, 'r').read().splitlines()
tags = [tag.split(',') for tag in tags]
tags = [[tag[0], tag[1].split('\t')] for tag in tags]
tags = [[tag[0], [words.split(':')[0].split() for words in tag[1]]]
for tag in tags]
tags = [[tag[0], [words for words in tag[1] if wanted_words(words)]]
for tag in tags]
tags = dict([[tag[0], to_same(tag[1])] for tag in tags
if wanted_tag(tag[1])])
if args.vector_type <= 2:
image_captions = {}
for key, val in tags.items():
image_captions[str(key) + '.jpg'] = [
'the girl has ' + val[0][0] + ' ' + val[0][1] + ' and ' +
val[1][0] + ' ' + val[1][1]
]
model = skipthoughts.load_model()
encoded_captions = {}
parallel_keys = [[] for i in range(cpus)]
quantity = (len(image_captions) + cpus - 1) // cpus
for i, key_val in enumerate(image_captions.items()):
parallel_keys[i // quantity].append(key_val[0])
out_q = Queue()
thrds =\
[Process(target=get_vector, args=(model, parallel_keys[i],
image_captions, out_q))
for i in range(cpus)]
for thrd in thrds:
thrd.start()
encoded_captions = {}
for i in range(cpus):
encoded_captions.update(out_q.get())
for thrd in thrds:
thrd.join()
elif args.vector_type == 3:
hair_list, eye_list = [], []
for key, val in tags.items():
if val[0][0] not in hair_list: hair_list.append(val[0][0])
if val[1][0] not in eyes_list: eyes_list.append(val[1][0])
h = h5py.File(join(data_dir, dict_file), 'w')
h.create_group('hair')
h.create_group('eyes')
for key in hair_list:
h['hair'].create_dataset(key, data=hair_list.index(key), dtype='i')
for key in eyes_list:
h['eyes'].create_dataset(key, data=eyes_list.index(key), dtype='i')
h['hair'].attrs['size'] = len(hair_list)
h['eyes'].attrs['size'] = len(eyes_list)
h.close()
encoded_captions = {}
for key, val in tags.items():
encoded_captions[str(key) + '.jpg'] = [
hair_list.index(val[0][0]),
eyes_list.index(val[1][0])
]
num_classes_hair = len(hair_list)
num_classes_eyes = len(eyes_list)
one_hot_hair =\
np.eye(num_classes_hair)[encoded_captions[str(key)+'.jpg'][0]]
one_hot_eyes =\
np.eye(num_classes_eyes)[encoded_captions[str(key)+'.jpg'][1]]
encoded_captions[str(key)+'.jpg'] = \
np.array([np.concatenate((one_hot_hair,one_hot_eyes),axis=0)])
elif args.vector_type == 4:
wordvecs = open(args.glove_file, 'r').read().splitlines()
wordvecs = [wordvec.split() for wordvec in wordvecs]
wordvecs = dict(
[[wordvec[0],
np.array([wordvec[1:]], dtype=np.float32)]
for wordvec in wordvecs])
encoded_captions = {}
for key, val in tags.items():
print(val)
if val[0][0] == 'twintails':
encoded_captions[key+'.jpg'] =\
(wordvecs['twin'] + wordvecs['tails'] + wordvecs[val[0][1]]*2)/4
else:
encoded_captions[key+'.jpg'] =\
(wordvecs[val[0][0]] + wordvecs[val[0][1]])/2
h = h5py.File(join(args.data_set, args.method_dir, args.out_file), 'w')
for key in encoded_captions:
h.create_dataset(key, data=encoded_captions[key])
h.close()