def prepare(cfg, checkpoint):
engine = build_engine(cfg.val_engine)
load_weights(engine.model, checkpoint, map_location='cpu')
device = torch.cuda.current_device()
engine = MMDataParallel(engine.to(device),
device_ids=[torch.cuda.current_device()])
dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader = build_dataloader(dataset, 1, 1, dist=False, shuffle=False)
return engine, dataloader
def trainval(cfg, logger):
for mode in cfg.modes:
assert mode in ('train', 'val')
dataloaders = dict()
engines = dict()
if 'train' in cfg.modes:
dataset = build_dataset(cfg.data.train)
dataloaders['train'] = build_dataloader(dataset,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
pin_memory=True)
engine = build_engine(cfg.train_engine)
engine = MMDataParallel(engine.cuda(),
device_ids=[torch.cuda.current_device()])
engines['train'] = engine
if 'val' in cfg.modes:
# TODO implement validation
dataset = build_dataset(cfg.data.val)
dataloaders['val'] = build_dataloader(dataset,
1,
cfg.data.workers_per_gpu,
pin_memory=True,
shuffle=False)
engine = build_engine(cfg.val_engine)
engine = MMDataParallel(engine.cuda(),
device_ids=[torch.cuda.current_device()])
engines['val'] = engine
hook_pool = HookPool(cfg.hooks, cfg.modes, logger)
looper = EpochBasedLooper(cfg.modes, dataloaders, engines, hook_pool,
logger, cfg.work_dir)
if 'weights' in cfg:
looper.load_weights(**cfg.weights)
if 'train' in cfg.modes:
if 'optimizer' in cfg:
looper.load_optimizer(**cfg.optimizer)
if 'meta' in cfg:
looper.load_meta(**cfg.meta)
else:
if 'optimizer' in cfg:
logger.warning('optimizer is not needed in train mode')
if 'meta' in cfg:
logger.warning('meta is not needed in train mode')
looper.start(cfg.max_epochs)
def trainval(cfg, distributed, logger):
for mode in cfg.modes:
assert mode in ('train', 'val')
dataloaders = dict()
engines = dict()
find_unused_parameters = cfg.get('find_unused_parameters', False)
if 'train' in cfg.modes:
dataset = build_dataset(cfg.data.train)
dataloaders['train'] = build_dataloader(dataset,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
dist=distributed,
seed=cfg.get('seed', None))
engine = build_engine(cfg.train_engine)
if distributed:
engine = MMDistributedDataParallel(
engine.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
engine = MMDataParallel(engine.cuda(),
device_ids=[torch.cuda.current_device()])
engines['train'] = engine
if 'val' in cfg.modes:
dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloaders['val'] = build_dataloader(dataset,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
engine = build_engine(cfg.val_engine)
if distributed:
engine = MMDistributedDataParallel(
engine.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
engine = MMDataParallel(engine.cuda(),
device_ids=[torch.cuda.current_device()])
engines['val'] = engine
hook_pool = HookPool(cfg.hooks, cfg.modes, logger)
looper = EpochBasedLooper(cfg.modes, dataloaders, engines, hook_pool,
logger, cfg.workdir)
if isinstance(looper, EpochBasedLooper):
looper.hook_pool.register_hook(dict(typename='WorkerInitHook'))
if distributed:
looper.hook_pool.register_hook(
dict(typename='DistSamplerSeedHook'))
if 'weights' in cfg:
looper.load_weights(**cfg.weights)
if 'train' in cfg.modes:
if 'optimizer' in cfg:
looper.load_optimizer(**cfg.optimizer)
if 'meta' in cfg:
looper.load_meta(**cfg.meta)
else:
if 'optimizer' in cfg:
logger.warning('optimizer is not needed in train mode')
if 'meta' in cfg:
logger.warning('meta is not needed in train mode')
looper.start(cfg.max_epochs)