def inference(cfg, model, device, **kwargs):
cur_epoch = kwargs.get('cur_epoch', None)
dataset_name = cfg.DATASETS.TEST.NAME
num_gpus = cfg.NUM_GPUS
data_loader = build_dataloader(cfg, is_train=False)
dataset = data_loader.dataset
evaluator = data_loader.dataset.evaluator
evaluator.clean()
logger.info("Evaluating {} dataset({} video clips):".format(dataset_name, len(dataset)))
if is_master_proc():
for images, targets in tqdm(data_loader):
compute_on_dataset(images, targets, device, model, num_gpus, evaluator)
else:
for images, targets in data_loader:
compute_on_dataset(images, targets, device, model, num_gpus, evaluator)
result_str, acc_dict = evaluator.get()
logger.info(result_str)
if is_master_proc():
output_dir = cfg.OUTPUT_DIR
result_path = os.path.join(output_dir,
'result_{}.txt'.format(datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))) \
if cur_epoch is None else os.path.join(output_dir, 'result_{:04d}.txt'.format(cur_epoch))
with open(result_path, "w") as f:
f.write(result_str)
return acc_dict
def inference(rgb_cfg, rgb_model, rgbdiff_cfg, rgbdiff_model, device):
logger_name = rgb_cfg.INFER.NAME
dataset_name = rgb_cfg.DATASETS.TEST.NAME
output_dir = rgb_cfg.OUTPUT.DIR
rgb_data_loader = build_dataloader(rgb_cfg, train=False)
rgbdiff_data_loader = build_dataloader(rgbdiff_cfg, train=False)
dataset = rgb_data_loader.dataset
logger = setup_logger(logger_name)
logger.info("Evaluating {} dataset({} video clips):".format(
dataset_name, len(dataset)))
results_dict, cate_acc_dict, acc_top1, acc_top5 = \
compute_on_dataset(rgb_model, rgb_data_loader, rgbdiff_model, rgbdiff_data_loader, device)
top1_acc = np.mean(acc_top1)
top5_acc = np.mean(acc_top5)
result_str = '\ntotal - top_1 acc: {:.3f}, top_5 acc: {:.3f}\n'.format(
top1_acc, top5_acc)
classes = dataset.classes
for key in sorted(results_dict.keys(), key=lambda x: int(x)):
total_num = results_dict[key]
acc_num = cate_acc_dict[key]
cate_name = classes[int(key)]
if total_num != 0:
result_str += '{:<3} - {:<20} - acc: {:.2f}\n'.format(
key, cate_name, acc_num / total_num * 100)
else:
result_str += '{:<3} - {:<20} - acc: 0.0\n'.format(
key, cate_name, acc_num / total_num)
logger.info(result_str)
result_path = os.path.join(
output_dir,
'result_{}.txt'.format(datetime.now().strftime('%Y-%m-%d_%H-%M-%S')))
with open(result_path, "w") as f:
f.write(result_str)
for handler in logger.handlers:
logger.removeHandler(handler)
return {'top1': top1_acc, 'top5': top5_acc}
def train(cfg):
# Set up environment.
init_distributed_training(cfg)
local_rank_id = get_local_rank()
# Set random seed from configs.
np.random.seed(cfg.RNG_SEED + 10 * local_rank_id)
torch.manual_seed(cfg.RNG_SEED + 10 * local_rank_id)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Setup logging format.
logging.setup_logging(cfg.OUTPUT_DIR)
logger.info('init start')
# 迭代轮数从１开始计数
arguments = {"cur_epoch": 1}
device = get_device(local_rank_id)
model = build_recognizer(cfg, device)
criterion = build_criterion(cfg, device)
optimizer = build_optimizer(cfg, model)
lr_scheduler = build_lr_scheduler(cfg, optimizer)
checkpointer = CheckPointer(model,
optimizer=optimizer,
scheduler=lr_scheduler,
save_dir=cfg.OUTPUT_DIR,
save_to_disk=True)
if cfg.TRAIN.RESUME:
logger.info('resume start')
extra_checkpoint_data = checkpointer.load(map_location=device)
if isinstance(extra_checkpoint_data, dict):
arguments['cur_epoch'] = extra_checkpoint_data['cur_epoch']
if cfg.LR_SCHEDULER.IS_WARMUP:
logger.info('warmup start')
if lr_scheduler.finished:
optimizer.load_state_dict(
lr_scheduler.after_scheduler.optimizer.state_dict())
else:
optimizer.load_state_dict(
lr_scheduler.optimizer.state_dict())
lr_scheduler.optimizer = optimizer
lr_scheduler.after_scheduler.optimizer = optimizer
logger.info('warmup end')
logger.info('resume end')
data_loader = build_dataloader(cfg, is_train=True)
logger.info('init end')
synchronize()
do_train(cfg, arguments, data_loader, model, criterion, optimizer,
lr_scheduler, checkpointer, device)
def train(gpu, args, cfg):
rank = args.nr * args.gpus + gpu
setup(rank, args.world_size)
logger = setup_logger(cfg.TRAIN.NAME)
arguments = {"iteration": 0}
torch.cuda.set_device(gpu)
device = torch.device(f'cuda:{gpu}' if torch.cuda.is_available() else 'cpu')
map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
model = build_model(cfg, gpu, map_location=map_location)
criterion = build_criterion(cfg)
optimizer = build_optimizer(cfg, model)
lr_scheduler = build_lr_scheduler(cfg, optimizer)
checkpointer = CheckPointer(model, optimizer=optimizer, scheduler=lr_scheduler, save_dir=cfg.OUTPUT.DIR,
save_to_disk=True, logger=logger)
if args.resume:
if is_master_proc():
logger.info('resume ...')
extra_checkpoint_data = checkpointer.load(map_location=map_location, rank=rank)
if extra_checkpoint_data != dict():
arguments['iteration'] = extra_checkpoint_data['iteration']
if cfg.LR_SCHEDULER.IS_WARMUP:
if is_master_proc():
logger.info('warmup ...')
if lr_scheduler.finished:
optimizer.load_state_dict(lr_scheduler.after_scheduler.optimizer.state_dict())
else:
optimizer.load_state_dict(lr_scheduler.optimizer.state_dict())
lr_scheduler.optimizer = optimizer
lr_scheduler.after_scheduler.optimizer = optimizer
data_loader = build_dataloader(cfg, is_train=True, start_iter=arguments['iteration'])
synchronize()
do_train(args, cfg, arguments,
data_loader, model, criterion, optimizer, lr_scheduler,
checkpointer, device, logger)
cleanup()
def train(gpu, args, cfg):
rank = args.nr * args.gpus + gpu
setup(rank, args.world_size, args.gpus)
logger = setup_logger(cfg.TRAIN.NAME)
arguments = {"iteration": 0}
arguments['rank'] = rank
device = torch.device(
f'cuda:{gpu}' if torch.cuda.is_available() else 'cpu')
map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
model = build_model(cfg, map_location=map_location).to(device)
if cfg.MODEL.PRETRAINED != "":
if rank == 0 and logger:
logger.info(f'load pretrained: {cfg.MODEL.PRETRAINED}')
checkpointer = CheckPointer(model, logger=logger)
checkpointer.load(cfg.MODEL.PRETRAINED,
map_location=map_location,
rank=rank)
if args.gpus > 1:
model = DDP(model, device_ids=[gpu], find_unused_parameters=True)
criterion = build_criterion(cfg)
optimizer = build_optimizer(cfg, model)
lr_scheduler = build_lr_scheduler(cfg, optimizer)
checkpointer = CheckPointer(model,
optimizer=optimizer,
scheduler=lr_scheduler,
save_dir=cfg.OUTPUT.DIR,
save_to_disk=True,
logger=logger)
if args.resume:
if rank == 0:
logger.info('resume ...')
extra_checkpoint_data = checkpointer.load(map_location=map_location,
rank=rank)
if extra_checkpoint_data != dict():
arguments['iteration'] = extra_checkpoint_data['iteration']
if cfg.LR_SCHEDULER.WARMUP:
if rank == 0:
logger.info('warmup ...')
if lr_scheduler.finished:
optimizer.load_state_dict(
lr_scheduler.after_scheduler.optimizer.state_dict())
else:
optimizer.load_state_dict(
lr_scheduler.optimizer.state_dict())
lr_scheduler.optimizer = optimizer
lr_scheduler.after_scheduler.optimizer = optimizer
data_loader = build_dataloader(cfg,
train=True,
start_iter=arguments['iteration'],
world_size=args.world_size,
rank=rank)
model = do_train(args, cfg, arguments, data_loader, model, criterion,
optimizer, lr_scheduler, checkpointer, device, logger)
if rank == 0 and not args.stop_eval:
logger.info('Start final evaluating...')
torch.cuda.empty_cache() # speed up evaluating after training finished
do_evaluation(cfg, model, device)
cleanup()
