def main():
import argparse
from attalos.dataset.mscoco_prep import MSCOCODatasetPrep
parser = argparse.ArgumentParser(description='Extract text features.')
parser.add_argument('--dataset_dir',
dest='dataset_dir',
type=str,
help='Directory with input data')
parser.add_argument('--dataset_type',
dest='dataset_type',
default='mscoco',
choices=[
'mscoco', 'visualgenome', 'iaprtc', 'generic',
'espgame', 'nuswide'
])
parser.add_argument('--split',
dest='split',
default='train',
choices=['train', 'test', 'val'])
parser.add_argument('--output_fname',
dest='output_fname',
default='captions_text.json.gz',
type=str,
help='Output json filename')
args = parser.parse_args()
if args.dataset_type == 'mscoco':
from attalos.dataset.mscoco_prep import MSCOCODatasetPrep
print('Processing MSCOCO Data')
dataset_prep = MSCOCODatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'visualgenome':
print('Processing Visual Genome Data')
from attalos.dataset.vg_prep import VGDatasetPrep
dataset_prep = VGDatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'iaprtc':
print('Processing IAPRTC-12 data')
from attalos.dataset.iaprtc12_prep import IAPRTC12DatasetPrep
dataset_prep = IAPRTC12DatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'generic':
print('Processing generic dataset')
from attalos.dataset.generic_prep import GenericDatasetPrep
dataset_prep = GenericDatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'espgame':
print('Processing espgame')
from attalos.dataset.espgame_prep import ESPGameDatasetPrep
dataset_prep = ESPGameDatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'nuswide':
print('Processing nuswide data')
from attalos.dataset.nuswide_prep import NUSWideDatasetPrep
dataset_prep = NUSWideDatasetPrep(args.dataset_dir, split=args.split)
else:
raise NotImplementedError('Dataset type {} not supported'.format(
args.dataset_type))
process_dataset(dataset_prep, args.output_fname)
def main(_):
import argparse
parser = argparse.ArgumentParser(description='Extract image features using Inception model.')
parser.add_argument('--dataset_dir',
dest='dataset_dir',
type=str,
help='Directory with input images')
parser.add_argument('--dataset_type',
dest='dataset_type',
default='mscoco',
choices=['mscoco', 'visualgenome', 'iaprtc', 'generic', 'espgame'])
parser.add_argument('--split',
dest='split',
default='train',
choices=['train', 'test', 'val'])
parser.add_argument('--output_fname',
dest='output_fname',
default='image_features.hdf5',
type=str,
help='Output hd5f filename')
parser.add_argument('--working_dir',
dest='working_dir',
default=tempfile.gettempdir(),
type=str,
help='Working directory for hdf5 file creation')
args = parser.parse_args()
if args.dataset_type == 'mscoco':
print('Processing MSCOCO Data')
from attalos.dataset.mscoco_prep import MSCOCODatasetPrep
dataset_prep = MSCOCODatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'visualgenome':
print('Processing Visual Genome Data')
from attalos.dataset.vg_prep import VGDatasetPrep
dataset_prep = VGDatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'iaprtc':
print('Processing IAPRTC-12 data')
from attalos.dataset.iaprtc12_prep import IAPRTC12DatasetPrep
dataset_prep = IAPRTC12DatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'generic':
print('Processing espgame data')
from attalos.dataset.generic_prep import GenericDatasetPrep
dataset_prep = GenericDatasetPrep(args.dataset_dir, split=args.split)
elif args.dataset_type == 'espgame':
print('Processing espgame data')
from attalos.dataset.espgame_prep import ESPGameDatasetPrep
dataset_prep = ESPGameDatasetPrep(args.dataset_dir, split=args.split)
else:
raise NotImplementedError('Dataset type {} not supported'.format(args.dataset_type))
process_dataset(dataset_prep, args.output_fname, working_dir=args.working_dir)
def load_metadata(self):
"""
Load the ESP Game Metadata to allow for efficient iteration
Returns:
"""
if self.split == SplitType.TRAIN:
split_name = 'train'
elif self.split == SplitType.TEST:
split_name = 'test'
else:
raise NotImplementedError('Split type not yet implemented')
if self.image_file_handle is None:
self.image_file_handle = tarfile.open(self.image_filename)
self.metadata_file_handle = tarfile.open(self.metadata_filename)
dictionary_lookup = {}
for i, word in enumerate(
self.metadata_file_handle.extractfile(
'espgame_dictionary.txt')):
dictionary_lookup[i] = word.strip()
filenames = []
filelist_filename = 'espgame_{}_list.txt'.format(split_name)
for filename in self.metadata_file_handle.extractfile(
filelist_filename):
img_id = os.path.basename(filename.strip()) + '.jpg'
filenames.append(img_id)
# Get valid tags for files
item_info = {}
annotationlist_filename = 'espgame_{}_annot.hvecs'.format(split_name)
annotationlist_file = self.metadata_file_handle.extractfile(
annotationlist_filename)
tag_onehot = IAPRTC12DatasetPrep.parse_LEAR_annotation_file(
annotationlist_file)
for img_index in range(tag_onehot.shape[0]):
image_id = filenames[img_index]
tags = []
for tag_id in np.nonzero(tag_onehot[img_index])[0]:
tags.append(dictionary_lookup[tag_id])
image_filename = 'ESP-ImageSet/images/{}'.format(image_id)
item_info[image_id] = {
'fname': image_filename,
'id': image_id,
'tags': tags,
'captions': []
}
return item_info
def load_metadata(self):
"""
Load the ESP Game Metadata to allow for efficient iteration
Returns:
"""
if self.split == SplitType.TRAIN:
split_name = 'train'
elif self.split == SplitType.TEST:
split_name = 'test'
else:
raise NotImplementedError('Split type not yet implemented')
if self.image_file_handle is None:
self.image_file_handle = tarfile.open(self.image_filename)
self.metadata_file_handle = tarfile.open(self.metadata_filename)
dictionary_lookup = {}
for i, word in enumerate(self.metadata_file_handle.extractfile('espgame_dictionary.txt')):
dictionary_lookup[i] = word.strip()
filenames = []
filelist_filename = 'espgame_{}_list.txt'.format(split_name)
for filename in self.metadata_file_handle.extractfile(filelist_filename):
img_id = os.path.basename(filename.strip()) + '.jpg'
filenames.append(img_id)
# Get valid tags for files
item_info = {}
annotationlist_filename = 'espgame_{}_annot.hvecs'.format(split_name)
annotationlist_file = self.metadata_file_handle.extractfile(annotationlist_filename)
tag_onehot = IAPRTC12DatasetPrep.parse_LEAR_annotation_file(annotationlist_file)
for img_index in range(tag_onehot.shape[0]):
image_id = filenames[img_index]
tags = []
for tag_id in np.nonzero(tag_onehot[img_index])[0]:
tags.append(dictionary_lookup[tag_id])
image_filename = 'ESP-ImageSet/images/{}'.format(image_id)
item_info[image_id] = {'fname': image_filename,
'id': image_id,
'tags': tags,
'captions': []}
return item_info