dim=0).to(device)
scores.append(f(x).item())
summary[v] = scores
machine_summary = tools.clip2frame(summary)
return evaluate(machine_summary, category_dict, args.topk_mAP)
if __name__ == '__main__':
device = torch.device(
"cuda") if torch.cuda.is_available() else torch.device("cpu")
loader_dict = dict(shuffle=True,
batch_size=args.batch_size,
pin_memory=True,
num_workers=16)
# train_path = os.path.join("/home/xuanteng/Highlight/proDataset/TrainingSet")
dataset = Pairs(args.train_path, args.domain, args.num_per_group)
train_loader = DataLoader(dataset, **loader_dict)
# recoder = Recorder("{}_{}".format(args.dataset, args.domain))
f = model.FNet().to(device)
h = model.HNet().to(device)
limloss = model.LIMloss().to(device)
opt = optim.SGD(list(f.parameters()) + list(h.parameters()),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
mAPs = []
for epoch_idx in range(args.epoch):
# logging.info("in epoch {}:\n".format(epoch_idx))
train(epoch_idx)
if epoch_idx % args.interval == 0:
# limloss = nn.DataParallel(limloss, device_ids=device_ids)
opt = optim.SGD(list(rnn.parameters()) + list(f.parameters()) +
list(h.parameters()),
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
scheduder = lr_scheduler.StepLR(opt, step_size=step_size, gamma=gamma)
# dataloader
dataloader_args = dict(shuffle=False,
batch_size=batch_size,
pin_memory=True,
num_workers=12)
train_loader = DataLoader(Pairs(domain, seq_len), **dataloader_args)
def train(domain):
rnn.train()
f.train()
h.train()
limloss.train()
if not os.path.exists(os.path.join('./', 'train_results')):
os.mkdir(os.path.join('./', 'train_results'))
if not os.path.exists(os.path.join('./', 'model_params')):
os.mkdir(os.path.join('./', 'model_params'))
with open(
os.path.join('./', 'train_results', '{}_train.txt'.format(domain)),
'w') as file: