cfg.GCN_SHARE_FEAT_PARAMS = checkpoint['gcn_share_feat_params']
if 'gcn_low_rank_params' in checkpoint.keys():
cfg.GCN_LOW_RANK_PARAMS = checkpoint['gcn_low_rank_params']
if 'gcn_low_rank_dim' in checkpoint.keys():
cfg.GCN_LOW_RANK_DIM = checkpoint['gcn_low_rank_dim']
if 'pooling_mode' in checkpoint.keys():
cfg.POOLING_MODE = checkpoint['pooling_mode']
if 'crop_with_max_pool' in checkpoint.keys():
cfg.ROI_CROP_WITH_MAX_POOL = checkpoint['crop_with_max_pool']
# pprint.pprint(cfg) ###
print('loaded checkpoint successfully!')
# initilize the network here.
if args.net == 'vgg16':
graphRCNN = vgg16(imdb._classes, imdb._attributes, imdb._relations,
4096)
elif args.net == 'res101':
graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations,
101)
elif args.net == 'res50':
graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations,
50)
elif args.net == 'res152':
graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations,
152)
graphRCNN.create_architecture()
# graphRCNN = nn.DataParallel(graphRCNN) ### for multi gpu
graphRCNN.load_state_dict(checkpoint['model'])
# graphRCNN = graphRCNN.module ### for multi gpu
im_info = im_info.cuda()
num_boxes = num_boxes.cuda()
gt_boxes = gt_boxes.cuda()
# make variable
im_data = Variable(im_data)
im_info = Variable(im_info)
num_boxes = Variable(num_boxes)
gt_boxes = Variable(gt_boxes)
if args.cuda > 0:
cfg.CUDA = True
# initilize the network here.
if args.net == 'vgg16':
graphRCNN = vgg16(imdb.classes, imdb.predicates, imdb.predicates, 4096)
elif args.net == 'res101':
graphRCNN = resnet(imdb.classes, imdb.predicates, imdb.predicates, 101)
elif args.net == 'res50':
graphRCNN = resnet(imdb.classes, imdb.predicates, imdb.predicates, 50)
elif args.net == 'res152':
graphRCNN = resnet(imdb.classes, imdb.predicates, imdb.predicates, 152)
graphRCNN.create_architecture()
lr = cfg.TRAIN.LEARNING_RATE
lr = args.lr_base
if args.mode == 0:
params = []
for key, value in dict(graphRCNN.named_parameters()).items():
im_info = im_info.cuda()
num_boxes = num_boxes.cuda()
gt_boxes = gt_boxes.cuda()
# make variable
im_data = Variable(im_data)
im_info = Variable(im_info)
num_boxes = Variable(num_boxes)
gt_boxes = Variable(gt_boxes)
if args.ngpu > 0:
cfg.CUDA = True
# initilize the network here.
if args.net == 'vgg16':
graphRCNN = vgg16(imdb._classes, imdb._attributes, imdb._relations, 4096) ## todo: ?4096, 101, 50, 152
# elif args.net == 'res101':
# graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations, 101)
# elif args.net == 'res50':
# graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations, 50)
# elif args.net == 'res152':
# graphRCNN = resnet(imdb._classes, imdb._attributes, imdb._relations, 152)
graphRCNN.create_architecture()
lr = cfg.TRAIN.LEARNING_RATE
if not args.resume:
params = []
for key, value in dict(graphRCNN.named_parameters()).items():
if value.requires_grad: