def vis_graph(lang_name, imdb_name, vis):
if 0:
vis_bar_top()
if 1:
lang_db = get_language_model(lang_name)
imdb = get_imdb(imdb_name)
unseen_labels = imdb.get_labels(1)
unseen_seen_labels = imdb.get_labels(2)
seen_labels = []
for s in unseen_seen_labels:
if s in unseen_labels:
continue
seen_labels.append(s)
lang_db.build_tree(2, unseen_seen_labels, imdb_name)
#vis_distance(lang_db, imdb, seen_labels)
vis_top_classes(lang_db, imdb, seen_labels)
parser.add_argument(
'--lm',
dest='lang_mod',
help='Initiate which language model to use with euc_distance',
default='w2v_wiki_300D',
type=str)
if len(sys.argv) == 1:
parser.print_help()
sys.exit(1)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print('Called with args: ')
print args
if args.method == 'pclass':
evaluate_pr_class(args.fn)
elif args.method == 'apredict':
show_actual_predicted(args.fn)
elif args.method == 'avg_class':
average_gt_classes(args.fn)
elif args.method == 'avgeuc':
average_cosine(args.fn, get_language_model(args.lang_mod))
else:
evaluate_flat_map(args.fn)
cls = words[1].lower()
if cls not in test_cls:
test_f.write('%s\n' % cls)
test_cls.append(cls)
classes.remove(cls)
print "Total test classes: {}".format(len(test_cls))
train_cls = []
for c in classes:
if lang_db.word_vector(c) is not None:
train_cls.append(c)
train_f.write('%s\n' % c)
print "Total train classes: {}".format(len(train_cls))
if __name__ == '__main__':
this_dir = osp.dirname(__file__)
lib_path = osp.join(this_dir, '../..', 'lib')
sys.path.insert(0, lib_path)
from bt_datasets.factory import get_imdb
from language_models.language_factory import get_language_model
lang_db = get_language_model('glove_wiki_300D')
imdb = get_imdb('imagenet1k_val')
generate_imagenet_single(lang_db, imdb)
#generate_splitted_classes(lang_db)
help='randomize (do not use a fixed seed)',
action='store_true')
parser.add_argument('--loss',
dest='loss',
help='loss function to run',
default='hinge',
type=str)
if len(sys.argv) == 1:
parser.print_help()
sys.exit(1)
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print('Called with args:')
print(args)
if not args.randomize:
# fix the random seeds (numpy and caffe) for reproducibility
np.random.seed(42)
lang_db = get_language_model(args.lang_name)
imdb = get_imdb(args.imdb_name)
train_bts(lang_db, imdb, max_iters=args.max_iters, loss=args.loss)
def clean_train_test(lang_name, train_u_s, train_u, test):
from language_models.language_factory import get_language_model
lang_db = get_language_model(lang_name)
print lang_db
train_u_s_read = open(train_u_s).readlines()
print "len(train_us)", len(train_u_s_read)
unique_lbl_t = {}
unique_lbl_u = {}
unique_lbl_us = {}
train_u_read = open(train_u).readlines()
print "len(train_u)", len(train_u_s_read)
test_read = open(test).readlines()
print "len(train)", len(train_u_s_read)
with open(train_u_s, "w") as train_u_s_file, \
open(train_u, "w") as train_u_file, \
open(test, "w") as test_file:
for img_lbls in train_u_s_read:
line = img_lbls.split(",")
img = line[0].strip()
lbls = [x.strip() for x in line[1:]]
new_lbls = []
for lbl in lbls:
lbl = lbl.strip()
vec = lang_db.word_vector(lbl)
if vec is not None:
new_lbls.append(lbl)
if lbl is not None:
unique_lbl_us[lbl] = 1
img_path = img.split("\\")
path = img_path[0] + "/" + img_path[1]
train_u_s_file.write(path + "," + ",".join(set(new_lbls)) + "\n")
print "s/u", path, new_lbls
for img_lbls in train_u_read:
line = img_lbls.split(",")
img = line[0].strip()
lbls = [x.strip() for x in line[1:]]
new_lbls = []
for lbl in lbls:
lbl = lbl.strip()
vec = lang_db.word_vector(lbl.strip())
if vec is not None:
new_lbls.append(lbl.strip())
if lbl is not unique_lbl_us:
unique_lbl_u[lbl] = 1
img_path = img.split("\\")
path = img_path[0] + "/" + img_path[1]
print "unseen", path, new_lbls
train_u_file.write(path + "," + ",".join(set(new_lbls)) + "\n")
for img_lbls in test_read:
line = img_lbls.split(",")
img = line[0].strip()
lbls = [x.strip() for x in line[1:]]
new_lbls = []
for lbl in lbls:
lbl = lbl.strip()
vec = lang_db.word_vector(lbl)
if vec is not None:
new_lbls.append(lbl)
if lbl is not unique_lbl_us:
unique_lbl_t[lbl] = 1
img_path = img.split("\\")
path = img_path[0] + "/" + img_path[1]
print "test", path, new_lbls
test_file.write(path + "," + ",".join(set(new_lbls)) + "\n")