def train(cfg, local_rank, distributed, use_tensorboard=False):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model, optimizer, opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False, find_unused_parameters=True,
)
arguments = {"iteration": 0, "iter_size": cfg.SOLVER.ITER_SIZE}
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
checkpointer = DetectronCheckpointer(
cfg, model, optimizer, scheduler, output_dir, save_to_disk
)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
if use_tensorboard:
meters = TensorboardLogger(
log_dir=os.path.join(cfg['OUTPUT_DIR'], 'log/'),
start_iter=arguments['iteration'],
delimiter=" ")
else:
meters = MetricLogger(delimiter=" ")
do_train(
model,
data_loader,
optimizer,
scheduler,
checkpointer,
device,
checkpoint_period,
arguments,
meters
)
return model
def train_cdb(cfg, local_rank, distributed, use_tensorboard=False):
model = build_detection_model(cfg)
device = torch.device(cfg.MODEL.DEVICE)
model.to(device)
model_cdb = ConvConcreteDB(cfg, model.backbone.out_channels)
model_cdb.to(device)
optimizer = make_optimizer(cfg, model)
scheduler = make_lr_scheduler(cfg, optimizer)
optimizer_cdb = make_cdb_optimizer(cfg, model_cdb)
scheduler_cdb = make_lr_cdb_scheduler(cfg, optimizer_cdb)
# Initialize mixed-precision training
use_mixed_precision = cfg.DTYPE == "float16"
amp_opt_level = 'O1' if use_mixed_precision else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level)
model_cdb, optimizer_cdb, = amp.initialize(model_cdb,
optimizer_cdb,
opt_level=amp_opt_level)
if distributed:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
find_unused_parameters=True,
)
model_cdb = torch.nn.parallel.DistributedDataParallel(
model_cdb,
device_ids=[local_rank],
output_device=local_rank,
# this should be removed if we update BatchNorm stats
broadcast_buffers=False,
find_unused_parameters=True,
)
arguments = {"iteration": 0}
output_dir = cfg.OUTPUT_DIR
save_to_disk = get_rank() == 0
checkpoint_period = cfg.SOLVER.CHECKPOINT_PERIOD
# TODO: check whether the *_cdb is properly loaded for inference when using 1 GPU
checkpointer = DetectronCheckpointer(cfg,
model,
optimizer,
scheduler,
output_dir,
save_to_disk,
model_cdb=model_cdb)
extra_checkpoint_data = checkpointer.load(cfg.MODEL.WEIGHT)
arguments.update(extra_checkpoint_data)
print('Do CFG_DIRE', cfg.PATH_DATA_TRAIN)
data_loader = make_data_loader(
cfg,
is_train=True,
is_distributed=distributed,
start_iter=arguments["iteration"],
)
if use_tensorboard:
meters = TensorboardLogger(log_dir=os.path.join(
cfg['OUTPUT_DIR'], 'log/'),
start_iter=arguments['iteration'],
delimiter=" ")
else:
meters = MetricLogger(delimiter=" ")
do_train_cdb(model, model_cdb, data_loader, optimizer, optimizer_cdb,
scheduler, scheduler_cdb, checkpointer, device,
checkpoint_period, arguments, meters, cfg)
return model