# net = st_resnet.resnet_st_seg01.resnet50(pretrained=False, num_classes=args.num_classes)
# net.load_state_dict(torch.load(model_path), strict = True)
args.batch_size = 1
st_crf_layer = sec.sec_org_net.STCRFLayer(True)
print(args)
# if flag_use_cuda:
# net.cuda()
dataloader = VOCData(args)
max_iou = 0
iou_obj = common_function.iou_calculator()
num_train_batch = len(dataloader.dataloaders["train"])
# net.train(False)
with torch.no_grad():
train_iou = 0
eval_iou = 0
counter = 0
for data in dataloader.dataloaders["train"]:
inputs, labels, mask_gt, img, cues = data
# if flag_use_cuda:
# inputs = inputs.cuda(); labels = labels.cuda(); cues = cues.cuda()
#
# sm_mask = net(inputs)
print(args)
print(model_path)
if flag_use_cuda:
net.cuda()
dataloader = multi_scale.voc_data_mul_scale_w_cues.VOCData(args)
optimizer = optim.Adam(net.parameters(),
lr=args.lr) # L2 penalty: norm weight_decay=0.0001
# main_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.step_size)
main_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.5)
max_iou = 0
iou_obj_train = common_function.iou_calculator()
iou_obj_eval = common_function.iou_calculator()
num_train_batch = len(dataloader.dataloaders["train"])
weight_STBCE = 0.1
weight_dec = 0.9
iter_counter = 0
for epoch in range(args.epochs):
train_seed_loss = 0.0
train_expand_loss = 0.0
train_constraint_loss = 0.0
train_st_BEC_loss = 0.0
train_st_half_BCE_loss = 0.0
