b_size = 10000 if args.num >= 10000 else args.num # This command is useless because the two dataset HICO-DEt and HOI-COCO have less 10000 verbs.
b_size = len(verb_feats) if b_size >= len(verb_feats) else b_size
tfconfig = tf.ConfigProto(device_count={"CPU": 12},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
allow_soft_placement=True)
# tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.2
tfconfig.gpu_options.allow_growth = True
sess = tf.Session(config=tfconfig)
# Generate_HICO_detection(output_file, HICO_dir)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
elif args.model.__contains__('VCOCO'):
os.environ['DATASET'] = 'VCOCO1'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
real_dataset_name = args.dataset if not args.dataset.startswith(
'gtobj365_coco') else 'gtobj365_coco'
output_file = cfg.LOCAL_DATA + '/feats/' + str(
args.iteration) + '_' + args.model + '_{}_{}.pkl'.format(
args.dataset, args.num)
print(output_file)
# if os.path.exists(output_file):
tfconfig = tf.ConfigProto(device_count={"CPU": 12},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
allow_soft_placement=True)
# init session
tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.2
sess = tf.Session(config=tfconfig)
# Generate_HICO_detection(output_file, HICO_dir)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
stride = 200
image, blobs, image_id = obtain_test_dataset1(args.object_thres,
args.human_thres,
stride=stride,
test_type=args.test_type,
model_name=args.model)
tmp_labels = tf.one_hot(tf.reshape(tf.cast(blobs['O_cls'], tf.int32),
shape=[
-1,
]),
tb_dir = cfg.ROOT_DIR + '/logs/' + args.model + '/'
# output directory where the models are saved
output_dir = cfg.ROOT_DIR + '/Weights/' + args.model + '/'
if os.path.exists(output_dir + 'checkpoint'):
args.Restore_flag = -1
if args.model.__contains__('unique_weights'):
args.Restore_flag = 6
augment_type = get_augment_type(args.model)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
with_pose = False
# if args.model.__contains__('pose'):
# with_pose = True
zero_shot_type = get_zero_shot_type(args.model)
assert args.model.__contains__('multi')
image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp = obtain_data1(
Pos_augment=args.Pos_augment,
Neg_select=args.Neg_select,
augment_type=augment_type,
', path = ' + weight)
output_file = cfg.ROOT_DIR + '/Results/' + str(
args.iteration) + '_' + args.model + '.pkl'
# init session
# gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.2)
tfconfig = tf.ConfigProto(gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=0.2))
# tfconfig.gpu_options.allow_growth=True
sess = tf.Session(config=tfconfig)
iCAN_Early_flag = 0
import os
os.environ['DATASET'] = 'VCOCO'
from networks.HOI import HOI
net = HOI(args.model)
net.create_architecture(False)
saver = tf.train.Saver()
saver.restore(sess, weight)
print('Pre-trained weights loaded.')
test_net(sess, net, Test_RCNN, prior_mask, Action_dic_inv, output_file,
args.object_thres, args.human_thres, args.prior_flag)
sess.close()
vcocoeval._do_eval(output_file, ovr_thresh=0.5)
sess = tf.Session(config=tfconfig)
# if args.model == 'iCAN_ResNet50_VCOCO':
# from networks.iCAN_ResNet50_VCOCO import ResNet50
# net = ResNet50(args.model)
# elif args.model == 'iCAN_ResNet50_VCOCO_Early':
# from networks.iCAN_ResNet50_VCOCO_Early import ResNet50
# net = ResNet50(args.model)
# else:
iCAN_Early_flag = 0
import os
os.environ['DATASET'] = 'VCOCO1'
from networks.HOI import HOI
net = HOI(args.model)
print('Pre-trained weights loaded.')
# net = DisentanglingNet(model_name=args.model)
image, blobs, image_id = obtain_data(Test_RCNN, prior_mask, Action_dic_inv,
output_file, args.object_thres,
args.human_thres, args.prior_flag)
print('blobs:', blobs)
print(image)
net.set_ph(image,
image_id,
H_num=blobs['H_num'],
spatial=blobs['sp'],
H_boxes=blobs['H_boxes'],
tb_dir = cfg.ROOT_DIR + '/logs/' + args.model + '/'
# output directory where the models are saved
output_dir = cfg.LOCAL_DATA + '/Weights/' + args.model + '/'
if args.Restore_flag == 5:
if os.path.exists(output_dir + 'checkpoint'):
args.Restore_flag = -1
elif args.model.__contains__('unique_weights'):
args.Restore_flag = 6
augment_type = get_augment_type(args.model)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
with_pose = False
# if args.model.__contains__('pose'):
# with_pose = True
coco = False
zero_shot_type = get_zero_shot_type(args.model)
large_neg_for_ho = False
if args.model.endswith('_aug5_new') or args.model.endswith('_aug6_new'):
large_neg_for_ho = True
image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp = obtain_data(
Pos_augment=args.Pos_augment,
'Noticeably, VCOCO only contains 26 verbs. We do not evaluate t2 and t3 carefully'
# output directory where the logs are saved
tb_dir = cfg.ROOT_DIR + '/logs/' + args.model + '/'
# output directory where the models are saved
output_dir = cfg.ROOT_DIR + '/Weights/' + args.model + '/'
import os
os.environ['DATASET'] = 'VCOCO'
from networks.HOI import HOI
augment_type = 0
if args.model.__contains__('_aug5'):
augment_type = 4
elif args.model.__contains__('_aug6'):
augment_type = 5
network = HOI(args.model)
image, image_id, num_pos, blobs = obtain_coco_data1(
args.Pos_augment, args.Neg_select)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
if not os.path.exists(tb_dir):
os.makedirs(tb_dir)
tfconfig = tf.ConfigProto(device_count={"CPU": 16},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8)
# tfconfig = tf.ConfigProto()
tfconfig.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
tfconfig.gpu_options.allow_growth = True
HICO_dir = cfg.LOCAL_DATA + '/Results/HICO/' + str(args.iteration) + '_' + args.model + '/'
tfconfig = tf.ConfigProto(device_count={"CPU": 16},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
allow_soft_placement=True)
# init session
# tfconfig = tf.ConfigProto(allow_soft_placement=True)
tfconfig.gpu_options.allow_growth=True
sess = tf.Session(config=tfconfig)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
if args.type == 'train':
image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp = obtain_data(
Pos_augment=0,
Neg_select=0,
augment_type=-1,
pattern_type=False)
net.set_ph(image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp)
else:
image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp = obtain_test_data(
Pos_augment=0,
iCAN_Early_flag = 0
import os
os.environ['DATASET'] = 'VCOCO'
from networks.HOI import HOI
augment_type = 0
if args.model.__contains__('_aug2'):
augment_type = 1
elif args.model.__contains__('_aug3'):
augment_type = 2
elif args.model.__contains__('_aug4'):
augment_type = 3
elif args.model.__contains__('_aug5'):
augment_type = 4
elif args.model.__contains__('_aug6'):
augment_type = 5
net = HOI(args.model)
image, image_id, num_pos, blobs = obtain_coco_data(args.Pos_augment,
args.Neg_select)
net.set_ph(image, image_id, num_pos, blobs['sp'], blobs['H_boxes'],
blobs['O_boxes'], blobs['gt_class_H'], blobs['gt_class_HO'],
blobs['gt_class_sp'], blobs['Mask_HO'], blobs['Mask_H'],
blobs['Mask_sp'], blobs['gt_class_C'])
train_net(net,
Trainval_GT,
Trainval_N,
output_dir,
tb_dir,
args.Pos_augment,
args.Neg_select,
iCAN_Early_flag,
HICO_dir = cfg.ROOT_DIR + '/Results/HICO/' + str(
args.iteration) + '_' + args.model + '/'
tfconfig = tf.ConfigProto(device_count={"CPU": 12},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
allow_soft_placement=True)
tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.8
sess = tf.Session(config=tfconfig)
# Generate_HICO_detection(output_file, HICO_dir)
if args.model.__contains__('VCOCO'):
os.environ['DATASET'] = 'VCOCO1'
from networks.HOI import HOI
net = HOI(model_name=args.model)
elif args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
stride = 200
test_type = 'default'
test_type = args.test_type
image, blobs, image_id = obtain_test_dataset_with_obj(
args.object_thres,
HICO_dir = cfg.LOCAL_DATA + '/Results/HICO/' + str(
args.iteration) + '_' + args.model + '/'
tfconfig = tf.ConfigProto(device_count={"CPU": 16},
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
allow_soft_placement=True)
# init session
# tfconfig = tf.ConfigProto(allow_soft_placement=True)
tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.1
sess = tf.Session(config=tfconfig)
if args.model.__contains__('res101'):
os.environ['DATASET'] = 'HICO_res101'
from networks.HOI import HOI
net = HOI(model_name=args.model)
elif args.model.__contains__('VCOCO'):
os.environ['DATASET'] = 'VCOCO1'
from networks.HOI import HOI
net = HOI(model_name=args.model)
else:
from networks.HOI import HOI
net = HOI(model_name=args.model)
if args.type == 'train':
if not args.model.__contains__('VCOCO'):
large_neg_for_ho = False
image, image_id, num_pos, Human_augmented, Object_augmented, action_HO, sp = obtain_data(
Pos_augment=0,
Neg_select=0,
augment_type=-1,