mata = dict()
# make dirs
mkdir_or_exist(osp.abspath(cfg.savepath))
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
cfg.log_file = osp.join(cfg.savepath, f'{timestamp}.log')
# create text log
# build model
model = build_network(cfg.model,
train_cfg=cfg.train_cfg,
test_cfg=cfg.test_cfg)
load(cfg.load_from, model, None)
# build dataset
datasets = build_dataset(cfg.data.test)
# put model on gpu
if torch.cuda.is_available():
# model = DataParallel(model.cuda(), device_ids=cfg.gpu_ids)
model = model.cuda()
# create data_loader
data_loader = build_dataloader(datasets, cfg.data.val_samples_per_gpu,
cfg.data.val_workers_per_gpu,
len(cfg.gpu_ids))
save_cfg = False
for i in range(len(cfg.test_pipeling)):
if 'Normalize' == cfg.test_pipeling[i].type:
save_cfg = True
save_path = osp.join(cfg.savepath,
# build model
model = build_network(cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
logger.info('-' * 20 + 'finish build model' + '-' * 20)
logger.info('Total Parameters: %d, Trainable Parameters: %s',
model.net_parameters['Total'],
str(model.net_parameters['Trainable']))
optimizer = build_optimizer(model, cfg.optimizer)
if cfg.resume_from:
start_epoch, ite_num = resume(cfg.resume_from, model, optimizer, logger, )
elif cfg.load_from:
load(cfg.load_from, model, logger)
# model = convert_syncbn_model(model)
# build dataset
datasets = build_dataset(cfg.data.train)
logger.info('-' * 20 + 'finish build dataset' + '-' * 20)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# put model on gpu
if torch.cuda.is_available():
# device = torch.device(cfg.gpu_ids)
model = model.to(device)
model.train()
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
# model = DDP(model, delay_allreduce=True)
# model = DDP(model, delay_allreduce=True)
# create data_loader
data_loader = build_dataloader(
datasets,