def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(
dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
eval_cfg = cfg.get('evaluation', {})
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(
CocoDistEvalRecallHook(val_dataset_cfg, **eval_cfg))
else:
dataset_type = getattr(datasets, val_dataset_cfg.type)
runner.register_hook(
MyDistEvalHook(val_dataset_cfg, **eval_cfg))
# if issubclass(dataset_type, datasets.CocoDataset):
# runner.register_hook(
# CocoDistEvalmAPHook(val_dataset_cfg, **eval_cfg))
# else:
# runner.register_hook(
# DistEvalmAPHook(val_dataset_cfg, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
logger = get_logger(cfg.log_level)
# init distributed environment if necessary
if args.launcher == 'none':
dist = False
logger.info('Disabled distributed training.')
else:
dist = True
init_dist(**cfg.dist_params)
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
if rank != 0:
logger.setLevel('ERROR')
logger.info('Enabled distributed training.')
try:
dataset = getattr(data, cfg.dataset_type)
train_loader, val_loader = dataset(cfg.data_root, batch_size = cfg.total_bs, imgsize=cfg.img_size)
except:
print("Dataset Type {} are Not Implemented. ")
exit()
# build model
try:
model = ptcv_get_model(cfg.model.type, pretrained=cfg.model.pretrained)
except:
print("Model Not Implemented. ")
exit()
if dist:
model = DistributedDataParallel(model.cuda(), device_ids=[torch.cuda.current_device()])
else:
model = DataParallel(model, device_ids=cfg.gpus).cuda()
# build runner and register hooks
runner = Runner(model,batch_processor,cfg.optimizer,cfg.work_dir,log_level=cfg.log_level)
runner.register_training_hooks(
lr_config=cfg.lr_config,
optimizer_config=cfg.optimizer_config,
checkpoint_config=cfg.checkpoint_config,
log_config=cfg.log_config)
if dist:
runner.register_hook(DistSamplerSeedHook())
# load param (if necessary) and run
if cfg.get('resume_from') is not None:
runner.resume(cfg.resume_from)
elif cfg.get('load_from') is not None:
runner.load_checkpoint(cfg.load_from)
runner.run([train_loader, val_loader], cfg.workflow, cfg.total_epochs)
def runner(self):
def parse_losses(losses):
log_vars = OrderedDict()
for loss_name, loss_value in losses.items():
if isinstance(loss_value, torch.Tensor):
log_vars[loss_name] = loss_value.mean()
elif isinstance(loss_value, list):
log_vars[loss_name] = sum(_loss.mean()
for _loss in loss_value)
else:
raise TypeError(
'{} is not a tensor or list of tensors'.format(
loss_name))
loss = sum(_value for _key, _value in log_vars.items()
if 'loss' in _key)
log_vars['loss'] = loss
for name in log_vars:
log_vars[name] = log_vars[name].item()
return loss, log_vars
def batch_processor(model, data, train_mode):
losses = model(**data)
# losses = model(data)
loss, log_vars = parse_losses(losses)
outputs = dict(loss=loss,
log_vars=log_vars,
num_samples=len(data['batchdata'].data))
return outputs
self.runner = Runner(self.model, batch_processor, self.optimizer,
self.arg.work_dir)
optimizer_config = DistOptimizerHook(
grad_clip=dict(max_norm=20, norm_type=2))
if not "policy" in self.arg.policy:
lr_config = dict(policy='step', step=self.arg.step)
else:
lr_config = dict(**self.arg.policy)
checkpoint_config = dict(interval=5)
log_config = dict(interval=20,
hooks=[
dict(type='TextLoggerHook'),
dict(type='TensorboardLoggerHook')
])
self.runner.register_training_hooks(lr_config, optimizer_config,
checkpoint_config, log_config)
self.runner.register_hook(DistSamplerSeedHook())
Feeder = import_class(self.arg.feeder)
self.runner.register_hook(
DistEvalTopKAccuracyHook(Feeder(**self.arg.test_feeder_args),
interval=self.arg.test_interval,
k=(1, 5)))
def _dist_train(model, dataset, cfg, logger=None, timestamp=None, meta=None):
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
data_loaders = [
build_dataloader(
ds,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True,
shuffle=True,
replace=getattr(cfg.data, 'sampling_replace', False),
seed=cfg.seed,
drop_last=getattr(cfg.data, 'drop_last', False)) for ds in dataset
]
# put model on gpus
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(
model,
batch_processor,
optimizer,
cfg.work_dir,
logger=logger,
meta=meta)
# an ugly walkaround to make the .log and .log.json filenames the same
runner.timestamp = timestamp
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register custom hooks
for hook in cfg.get('custom_hooks', ()):
if hook.type == 'DeepClusterHook':
common_params = dict(dist_mode=True, data_loaders=data_loaders)
else:
common_params = dict(dist_mode=True)
runner.register_hook(build_hook(hook, common_params))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model,
dataset,
cfg,
validate=False,
logger=None,
timestamp=None):
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
data_loaders = [
build_dataloader(ds,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True,
seed=cfg.seed) for ds in dataset
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model,
batch_processor,
optimizer,
cfg.work_dir,
logger=logger)
# an ugly walkaround to make the .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
eval_cfg = cfg.get('evaluation', {})
runner.register_hook(DistEvalHook(val_dataset_cfg, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
data_loaders = [
build_dataloader(ds,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True) for ds in dataset
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
eval_cfg = cfg.evaluation
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(
CocoDistEvalRecallHook(val_dataset_cfg, **eval_cfg))
else:
if isinstance(val_dataset_cfg, dict):
runner.register_hook(
KaggleEvalHook(val_dataset_cfg, **eval_cfg))
elif isinstance(val_dataset_cfg, list):
for vdc in val_dataset_cfg:
runner.register_hook(KaggleEvalHook(vdc, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False, **kwargs):
logger = get_root_logger(cfg.log_level)
# prepare data loaders
data_loaders = [
build_dataloader(
dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True,
repeat_samples=cfg.train_cfg.repeat_samples,
**kwargs)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda(), device_ids=[torch.cuda.current_device()])
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(
model,
#batch_processor=batch_processor,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger)
#runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
# logger)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(val_dataset_cfg))
else:
dataset_type = getattr(datasets, val_dataset_cfg.type)
if issubclass(dataset_type, datasets.CocoDataset):
runner.register_hook(CocoDistEvalmAPHook(val_dataset_cfg))
else:
runner.register_hook(DistEvalmAPHook(val_dataset_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.videos_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
# Hard-coding type of optimizer for now
print('Training #Params: ',
len(list(filter(lambda p: p.requires_grad, model.parameters()))))
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad,
model.parameters()),
lr=cfg.optimizer.lr,
momentum=cfg.optimizer.momentum,
weight_decay=cfg.optimizer.weight_decay)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if cfg.data.val.type in ['RawFramesDataset', 'VideoDataset']:
runner.register_hook(
DistEvalTopKAccuracyHook(cfg.data.val, k=(1, 5)))
if cfg.data.val.type == 'AVADataset':
runner.register_hook(AVADistEvalmAPHook(cfg.data.val))
# if validate:
# if isinstance(model.module, RPN):
# # TODO: implement recall hooks for other datasets
# runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
# else:
# if cfg.data.val.type == 'CocoDataset':
# runner.register_hook(CocoDistEvalmAPHook(cfg.data.val))
# else:
# runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
data_loaders = [
build_dataloader(ds,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True) for ds in dataset
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
eval_cfg = cfg.get('evaluation', {})
dataset_type = DATASETS.get(val_dataset_cfg.type)
if issubclass(dataset_type, datasets.CocoPoseDataset):
runner.register_hook(
CocoPoseDistEvalmAPHook(val_dataset_cfg, **eval_cfg))
elif issubclass(dataset_type, datasets.CocoDataset):
runner.register_hook(
CocoDistEvalmAPHook(val_dataset_cfg, **eval_cfg))
else:
runner.register_hook(DistEvalmAPHook(val_dataset_cfg, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(
dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True,
)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
runner = Runner(
model,
batch_processor,
cfg.optimizer,
cfg.work_dir,
cfg.log_level,
)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(
cfg.lr_config,
optimizer_config,
cfg.checkpoint_config,
cfg.log_config,
)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
else:
if cfg.data.val.type == "CocoDataset":
runner.register_hook(CocoDistEvalmAPHook(cfg.data.val))
else:
runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
else:
runner.load_checkpoint(tmp)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
shutil.copyfile(tmp, bw)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
# model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(val_dataset_cfg))
else:
dataset_type = getattr(datasets, val_dataset_cfg.type)
if issubclass(dataset_type, datasets.CocoDataset):
runner.register_hook(CocoDistEvalmAPHook(val_dataset_cfg))
else:
runner.register_hook(DistEvalmAPHook(val_dataset_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, data_loaders, batch_processor, cfg):
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(
dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level, mean_teacher=cfg.mean_teacher)
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
eval_cfg = cfg.get('evaluation', {})
eval_hook = eval_cfg.pop('eval_hook', 'CocoDistEvalmAPHook')
EvalHook = getattr(core, eval_hook)
runner.register_hook(EvalHook(val_dataset_cfg, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
if cfg.mean_teacher:
runner.load_mean_teacher_checkpoint(cfg)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, logger, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.videos_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Start: vj changes
find_unused_parameters = True
# End: vj chagnes
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(model, batch_processor, optimizer, cfg.work_dir,
logger)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if cfg.data.val.type in ['RawFramesDataset', 'VideoDataset']:
runner.register_hook(
DistEvalTopKAccuracyHook(cfg.data.val, k=(1, 5)))
if cfg.data.val.type == 'AVADataset':
runner.register_hook(AVADistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
for m in model.modules():
if isinstance(m, torch.nn.SyncBatchNorm):
m._specify_ddp_gpu_num(1)
model = MMDistributedDataParallel(model.cuda())
# model = DistributedDataParallel(model.cuda())
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset_cfg = cfg.data.val
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(val_dataset_cfg))
else:
dataset_type = getattr(datasets, val_dataset_cfg.type)
if issubclass(dataset_type, datasets.CocoDataset):
runner.register_hook(CocoDistEvalmAPHook(val_dataset_cfg))
else:
runner.register_hook(DistEvalmAPHook(val_dataset_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
try:
pad_size = cfg.data.pad_size
print("using padding size")
except:
pad_size = None
data_loaders = [
build_dataloader(dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
pad_size=pad_size,
dist=True)
]
rank = int(os.environ['RANK'])
num_gpus = torch.cuda.device_count()
# put model on gpus
model = MMDistributedDataParallel(model.cuda(rank % num_gpus))
torch.cuda.empty_cache()
# build runner
runner = Runner(model, batch_processor, cfg.optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
else:
if cfg.data.val.type == 'CocoDataset' or cfg.data.val.type == 'CocoZipDataset':
runner.register_hook(CocoDistEvalmAPHook(cfg.data.val))
else:
runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# with torch.no_grad():
# for j in range(2):
# print(j)
# for i, data_batch in enumerate(data_loaders[0]):
# _ = model(**data_batch)
# # break
# build runner
runner = Runner(model, batch_processor, cfg.optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
else:
if cfg.data.val.type == 'CocoDataset':
runner.register_hook(CocoDistEvalmAPHook(cfg.data.val))
else:
runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, dataset, cfg, validate=False):
# prepare data loaders
data_loaders = [
build_dataloader(dataset,
cfg.data.videos_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda(),
device_ids=[torch.cuda.current_device()])
# build runner
runner = Runner(model, batch_processor, cfg.optimizer, cfg.work_dir,
cfg.log_level)
# register hooks
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if cfg.data.val.type in ['RawFramesDataset', 'VideoDataset']:
runner.register_hook(
DistEvalTopKAccuracyHook(cfg.data.val, k=(1, 5)))
if cfg.data.val.type == 'AVADataset':
runner.register_hook(AVADistEvalmAPHook(cfg.data.val))
# if validate:
# if isinstance(model.module, RPN):
# # TODO: implement recall hooks for other datasets
# runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
# else:
# if cfg.data.val.type == 'CocoDataset':
# runner.register_hook(CocoDistEvalmAPHook(cfg.data.val))
# else:
# runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model, datasets, cfg, validate=False, logger=None):
# prepare data loaders
data_loaders = [
build_dataloader(
dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True) for dataset in datasets
]
# put model on gpus
model = MMDistributedDataParallel(model.cuda())
# build runner
runner = NASRunner(model, batch_processor, None, cfg.work_dir, cfg.log_level, cfg=cfg, logger=logger)
# register hooks
weight_optim_config = DistOptimizerHook(**cfg.optimizer.weight_optim.optimizer_config)
arch_optim_config = ArchDistOptimizerHook(**cfg.optimizer.arch_optim.optimizer_config)
runner.register_training_hooks(cfg.lr_config, weight_optim_config, arch_optim_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
if isinstance(model.module, RPN):
# TODO: implement recall hooks for other datasets
runner.register_hook(CocoDistEvalRecallHook(cfg.data.val))
else:
if cfg.dataset_type == 'CocoDataset':
# runner.register_hook(CocoDistEvalmAPHook_(datasets[1]))
runner.register_hook(CocoDistEvalmAPHook(cfg.data.val_))
else:
runner.register_hook(DistEvalmAPHook(cfg.data.val))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs, cfg.arch_update_epoch)
def train_detector(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
if 'imgs_per_gpu' in cfg.data:
logger.warning('"imgs_per_gpu" is deprecated in MMDet V2.0. '
'Please use "samples_per_gpu" instead')
if 'samples_per_gpu' in cfg.data:
logger.warning(
f'Got "imgs_per_gpu"={cfg.data.imgs_per_gpu} and '
f'"samples_per_gpu"={cfg.data.samples_per_gpu}, "imgs_per_gpu"'
f'={cfg.data.imgs_per_gpu} is used in this experiments')
else:
logger.warning(
'Automatically set "samples_per_gpu"="imgs_per_gpu"='
f'{cfg.data.imgs_per_gpu} in this experiments')
cfg.data.samples_per_gpu = cfg.data.imgs_per_gpu
data_loaders = [
build_dataloader(
ds,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
# cfg.gpus will be ignored if distributed
len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(model.cuda(cfg.gpu_ids[0]),
device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model,
batch_processor,
optimizer,
cfg.work_dir,
logger=logger,
meta=meta)
# defrost backbone hook
when_defrost = cfg.get('when_defrost')
if when_defrost is not None:
if when_defrost < 0:
raise RuntimeError('when_defrost < 0')
frozen_stages = cfg.get('frozen_stages', -1)
defrost_backbone = DefrostBackbone(when_defrost, frozen_stages)
runner.register_hook(defrost_backbone)
# log hook
custom_log = CustomLog(cfg.data.samples_per_gpu, when_defrost,
os.path.join(cfg.work_dir, 'log.txt'))
runner.register_hook(custom_log)
# an ugly walkaround to make the .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg,
distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
samples_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(
cfg.resume_from,
map_location=lambda storage, loc: storage.cuda(cfg.gpu_ids[0]))
elif cfg.load_from:
runner.load_checkpoint(
cfg.load_from,
map_location=lambda storage, loc: storage.cuda(cfg.gpu_ids[0]))
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model,
train_dataset,
cfg,
eval_dataset=None,
vis_dataset=None,
validate=False,
logger=None):
# prepare data loaders
data_loaders = [
build_data_loader(train_dataset,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True)
]
if cfg.apex.synced_bn:
# using apex synced BN
model = apex.parallel.convert_syncbn_model(model)
model = model.cuda()
# build optimizer
optimizer = build_optimizer(model, cfg.optimizer)
# Initialize mixed-precision training
if cfg.apex.use_mixed_precision:
amp_opt_level = 'O1' if cfg.apex.type == "float16" else 'O0'
model, optimizer = amp.initialize(model,
optimizer,
opt_level=amp_opt_level,
loss_scale=cfg.apex.loss_scale)
# put model on gpus
model = MMDistributedDataParallel(model)
# build runner
runner = Runner(model, batch_processor, optimizer, cfg.work_dir,
cfg.log_level, logger)
# register optimizer hooks
if cfg.apex.use_mixed_precision:
optimizer_config = DistApexOptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
logger.info("Register Optimizer Hook...")
runner.register_training_hooks(cfg.lr_config,
optimizer_config,
cfg.checkpoint_config,
log_config=None)
# register self-defined logging hooks
for info in cfg.log_config['hooks']:
assert isinstance(info, dict) and 'type' in info
if info['type'] in ['TensorboardLoggerHook']:
logger.info("Register Tensorboard Logger Hook...")
runner.register_hook(TensorboardLoggerHook(
interval=cfg.log_config.interval,
register_logWithIter_keyword=['loss']),
priority='VERY_LOW')
if info['type'] in ['TextLoggerHook']:
logger.info("Register Text Logger Hook...")
runner.register_hook(TextLoggerHook(
interval=cfg.log_config.interval, ),
priority='VERY_LOW')
logger.info("Register SamplerSeed Hook...")
runner.register_hook(DistSamplerSeedHook())
logger.info("Register EmptyCache Hook...")
runner.register_hook(EmptyCacheHook(before_epoch=True,
after_iter=False,
after_epoch=True),
priority='VERY_LOW')
# register eval hooks
if validate:
interval = cfg.get('validate_interval', 1)
if eval_dataset is not None:
logger.info("Register Evaluation Hook...")
runner.register_hook(
DistStereoEvalHook(cfg, eval_dataset, interval))
if vis_dataset is not None:
logger.info("Register Visualization hook...")
runner.register_hook(DistStereoVisHook(vis_dataset, cfg, interval))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (Dataset): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
# step 1: give default values and override (if exist) from cfg.data
loader_cfg = {
**dict(
seed=cfg.get('seed'),
drop_last=False,
dist=distributed,
num_gpus=len(cfg.gpu_ids)),
**({} if torch.__version__ != 'parrots' else dict(
prefetch_num=2,
pin_memory=False,
)),
**dict((k, cfg.data[k]) for k in [
'samples_per_gpu',
'workers_per_gpu',
'shuffle',
'seed',
'drop_last',
'prefetch_num',
'pin_memory',
'persistent_workers',
] if k in cfg.data)
}
# step 2: cfg.data.train_dataloader has highest priority
train_loader_cfg = dict(loader_cfg, **cfg.data.get('train_dataloader', {}))
data_loaders = [build_dataloader(ds, **train_loader_cfg) for ds in dataset]
# determine whether use adversarial training precess or not
use_adverserial_train = cfg.get('use_adversarial_train', False)
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', True)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
if use_adverserial_train:
# Use DistributedDataParallelWrapper for adversarial training
model = DistributedDataParallelWrapper(
model,
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizers(model, cfg.optimizer)
runner = EpochBasedRunner(
model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
if use_adverserial_train:
# The optimizer step process is included in the train_step function
# of the model, so the runner should NOT include optimizer hook.
optimizer_config = None
else:
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
samples_per_gpu=1,
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
drop_last=False,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def train_detector(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
if 'imgs_per_gpu' in cfg.data:
logger.warning('"imgs_per_gpu" is deprecated in MMDet V2.0. '
'Please use "samples_per_gpu" instead')
if 'samples_per_gpu' in cfg.data:
logger.warning(
f'Got "imgs_per_gpu"={cfg.data.imgs_per_gpu} and '
f'"samples_per_gpu"={cfg.data.samples_per_gpu}, "imgs_per_gpu"'
f'={cfg.data.imgs_per_gpu} is used in this experiments')
else:
logger.warning(
'Automatically set "samples_per_gpu"="imgs_per_gpu"='
f'{cfg.data.imgs_per_gpu} in this experiments')
cfg.data.samples_per_gpu = cfg.data.imgs_per_gpu
data_loaders = [
build_dataloader(
ds,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
# cfg.gpus will be ignored if distributed
len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
print(cfg)
print("-"*20)
print(cfg.optimizer)
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(
model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
# Support batch_size > 1 in validation
val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
if val_samples_per_gpu > 1:
# Replace 'ImageToTensor' to 'DefaultFormatBundle'
cfg.data.val.pipeline = replace_ImageToTensor(
cfg.data.val.pipeline)
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
samples_per_gpu=val_samples_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
# user-defined hooks
if cfg.get('custom_hooks', None):
custom_hooks = cfg.custom_hooks
assert isinstance(custom_hooks, list), \
f'custom_hooks expect list type, but got {type(custom_hooks)}'
for hook_cfg in cfg.custom_hooks:
assert isinstance(hook_cfg, dict), \
'Each item in custom_hooks expects dict type, but got ' \
f'{type(hook_cfg)}'
hook_cfg = hook_cfg.copy()
priority = hook_cfg.pop('priority', 'NORMAL')
hook = build_from_cfg(hook_cfg, HOOKS)
runner.register_hook(hook, priority=priority)
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def _dist_train(model,
dataset,
cfg,
validate=False,
logger=None,
timestamp=None,
meta=None):
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
data_loaders = [
build_dataloader(ds,
cfg.data.imgs_per_gpu,
cfg.data.workers_per_gpu,
dist=True,
seed=cfg.seed) for ds in dataset
]
# put model on gpus
find_unused_parameters = cfg.get('find_unused_parameters', True) # False
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model,
batch_processor,
optimizer,
cfg.work_dir,
logger=logger,
meta=meta)
# an ugly walkaround to make the .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg)
else:
optimizer_config = DistOptimizerHook(**cfg.optimizer_config)
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=True,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
runner.register_hook(DistEvalHook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def train_detector(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
if 'imgs_per_gpu' in cfg.data:
logger.warning('"imgs_per_gpu" is deprecated in MMDet V2.0. '
'Please use "samples_per_gpu" instead')
if 'samples_per_gpu' in cfg.data:
logger.warning(
f'Got "imgs_per_gpu"={cfg.data.imgs_per_gpu} and '
f'"samples_per_gpu"={cfg.data.samples_per_gpu}, "imgs_per_gpu"'
f'={cfg.data.imgs_per_gpu} is used in this experiments')
else:
logger.warning(
'Automatically set "samples_per_gpu"="imgs_per_gpu"='
f'{cfg.data.imgs_per_gpu} in this experiments')
cfg.data.samples_per_gpu = cfg.data.imgs_per_gpu
data_loaders = [
build_dataloader(
ds,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
# cfg.gpus will be ignored if distributed
len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(model.cuda(cfg.gpu_ids[0]),
device_ids=cfg.gpu_ids)
# if just swa training is performed,
# skip building the runner for the traditional training
if not cfg.get('only_swa_training', False):
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg,
distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
# Support batch_size > 1 in validation
val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
if val_samples_per_gpu > 1:
# Replace 'ImageToTensor' to 'DefaultFormatBundle'
cfg.data.val.pipeline = replace_ImageToTensor(
cfg.data.val.pipeline)
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
samples_per_gpu=val_samples_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
# user-defined hooks
if cfg.get('custom_hooks', None):
custom_hooks = cfg.custom_hooks
assert isinstance(custom_hooks, list), \
f'custom_hooks expect list type, but got {type(custom_hooks)}'
for hook_cfg in cfg.custom_hooks:
assert isinstance(hook_cfg, dict), \
'Each item in custom_hooks expects dict type, but got ' \
f'{type(hook_cfg)}'
hook_cfg = hook_cfg.copy()
priority = hook_cfg.pop('priority', 'NORMAL')
hook = build_from_cfg(hook_cfg, HOOKS)
runner.register_hook(hook, priority=priority)
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
else:
# if just swa training is performed, there should be a starting model
assert cfg.swa_resume_from is not None or cfg.swa_load_from is not None
# perform swa training
# build swa training runner
if not cfg.get('swa_training', False):
return
from mmdet.core import SWAHook
logger.info('Start SWA training')
swa_optimizer = build_optimizer(model, cfg.swa_optimizer)
swa_runner = EpochBasedRunner(model,
optimizer=swa_optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
swa_runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
swa_optimizer_config = Fp16OptimizerHook(**cfg.swa_optimizer_config,
**fp16_cfg,
distributed=distributed)
elif distributed and 'type' not in cfg.swa_optimizer_config:
swa_optimizer_config = OptimizerHook(**cfg.swa_optimizer_config)
else:
swa_optimizer_config = cfg.swa_optimizer_config
# register hooks
swa_runner.register_training_hooks(cfg.swa_lr_config, swa_optimizer_config,
cfg.swa_checkpoint_config,
cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
swa_runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
# Support batch_size > 1 in validation
val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
if val_samples_per_gpu > 1:
# Replace 'ImageToTensor' to 'DefaultFormatBundle'
cfg.data.val.pipeline = replace_ImageToTensor(
cfg.data.val.pipeline)
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
samples_per_gpu=val_samples_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
eval_hook = DistEvalHook if distributed else EvalHook
swa_runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
swa_eval = True
swa_eval_hook = eval_hook(val_dataloader, **eval_cfg)
else:
swa_eval = False
swa_eval_hook = None
# register swa hook
swa_hook = SWAHook(swa_eval=swa_eval, eval_hook=swa_eval_hook)
swa_runner.register_hook(swa_hook, priority='LOW')
# register user-defined hooks
if cfg.get('custom_hooks', None):
custom_hooks = cfg.custom_hooks
assert isinstance(custom_hooks, list), \
f'custom_hooks expect list type, but got {type(custom_hooks)}'
for hook_cfg in cfg.custom_hooks:
assert isinstance(hook_cfg, dict), \
'Each item in custom_hooks expects dict type, but got ' \
f'{type(hook_cfg)}'
hook_cfg = hook_cfg.copy()
priority = hook_cfg.pop('priority', 'NORMAL')
hook = build_from_cfg(hook_cfg, HOOKS)
swa_runner.register_hook(hook, priority=priority)
if cfg.swa_resume_from:
swa_runner.resume(cfg.swa_resume_from)
elif cfg.swa_load_from:
# use the best pretrained model as the starting model for swa training
if cfg.swa_load_from == 'best_bbox_mAP.pth':
best_model_path = os.path.join(cfg.work_dir, cfg.swa_load_from)
assert os.path.exists(best_model_path)
# avoid the best pretrained model being overwritten
new_best_model_path = os.path.join(cfg.work_dir,
'best_bbox_mAP_pretrained.pth')
os.rename(best_model_path, new_best_model_path)
cfg.swa_load_from = new_best_model_path
swa_runner.load_checkpoint(cfg.swa_load_from)
swa_runner.run(data_loaders, cfg.workflow, cfg.swa_total_epochs)
def _dist_train(model, dataset, validate, cfg, logger):
# prepare data loaders
if validate:
train_dataset = dataset[0]
val_dataset = dataset[1]
else:
train_dataset = dataset
batch_size = cfg.get('gpu_batch', 1)
data_loaders = [
build_dataloader(
dataset=train_dataset,
workers_per_gpu=cfg.workers_per_gpu,
batch_size=batch_size,
dist=True,
sampler=torch.utils.data.DistributedSampler(train_dataset))
]
# put model on gpus
rank, _ = get_dist_info()
num_gpus = torch.cuda.device_count()
# syn batchnorm warp
model = model.cuda(rank % num_gpus)
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = MMDistributedDataParallel(model,
device_ids=[rank % num_gpus],
find_unused_parameters=True)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = Runner(model,
batch_processor=None,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
extra=None)
# register hooks
optimizer_config = OptimizerHook(**cfg.optimizer_config)
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config)
runner.register_hook(DistSamplerSeedHook())
check4resume = cfg.get('check4resume', None)
if check4resume:
runner.register_hook(Ck4resumeHook(**check4resume))
if validate:
if cfg.dataset.dataname.lower() in 'kinetics':
print('kinetics regist eval hook')
interval = cfg.get('eval_interval', 1)
runner.register_hook(
DistEvalTopKRecallHook(val_dataset,
cfg,
interval=interval,
eval_bs=batch_size))
else:
print('downstream task regist eval hook')
interval = cfg.get('eval_interval', 1)
runner.register_hook(
DistEvalAccuracy(val_dataset,
cfg,
interval=interval,
eval_bs=batch_size))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
part = cfg.get("part", None)
runner.load_checkpoint(cfg.load_from, part=part)
elif cfg.get('autoresume', False):
ckpath = os.path.join(cfg.work_dir, 'resume_latest.pth')
if os.path.exists(ckpath):
runner.resume(ckpath)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (:obj:`Dataset`): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(log_level=cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', {}),
workers_per_gpu=cfg.data.get('workers_per_gpu', {}),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('train_dataloader', {}))
data_loaders = [
build_dataloader(ds, **dataloader_setting) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(
model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', {}),
workers_per_gpu=cfg.data.get('workers_per_gpu', {}),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
test=dict(test_best=False, test_last=False),
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (:obj:`Dataset`): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
test (dict): The testing option, with two keys: test_last & test_best.
The value is True or False, indicating whether to test the
corresponding checkpoint.
Default: dict(test_best=False, test_last=False).
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(log_level=cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('train_dataloader', {}))
if cfg.omnisource:
# The option can override videos_per_gpu
train_ratio = cfg.data.get('train_ratio', [1] * len(dataset))
omni_videos_per_gpu = cfg.data.get('omni_videos_per_gpu', None)
if omni_videos_per_gpu is None:
dataloader_settings = [dataloader_setting] * len(dataset)
else:
dataloader_settings = []
for videos_per_gpu in omni_videos_per_gpu:
this_setting = cp.deepcopy(dataloader_setting)
this_setting['videos_per_gpu'] = videos_per_gpu
dataloader_settings.append(this_setting)
data_loaders = [
build_dataloader(ds, **setting)
for ds, setting in zip(dataset, dataloader_settings)
]
else:
data_loaders = [
build_dataloader(ds, **dataloader_setting) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
Runner = OmniSourceRunner if cfg.omnisource else EpochBasedRunner
runner = Runner(
model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
if cfg.omnisource:
runner.register_hook(OmniSourceDistSamplerSeedHook())
else:
runner.register_hook(DistSamplerSeedHook())
# precise bn setting
if cfg.get('precise_bn', False):
precise_bn_dataset = build_dataset(cfg.data.train)
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=1, # save memory and time
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
data_loader_precise_bn = build_dataloader(precise_bn_dataset,
**dataloader_setting)
precise_bn_hook = PreciseBNHook(data_loader_precise_bn,
**cfg.get('precise_bn'))
runner.register_hook(precise_bn_hook)
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook(val_dataloader, **eval_cfg) if distributed \
else EvalHook(val_dataloader, **eval_cfg)
runner.register_hook(eval_hook)
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner_kwargs = dict()
if cfg.omnisource:
runner_kwargs = dict(train_ratio=train_ratio)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs, **runner_kwargs)
if test['test_last'] or test['test_best']:
best_ckpt_path = None
if test['test_best']:
if hasattr(eval_hook, 'best_ckpt_path'):
best_ckpt_path = eval_hook.best_ckpt_path
if best_ckpt_path is None or not osp.exists(best_ckpt_path):
test['test_best'] = False
if best_ckpt_path is None:
runner.logger.info('Warning: test_best set as True, but '
'is not applicable '
'(eval_hook.best_ckpt_path is None)')
else:
runner.logger.info('Warning: test_best set as True, but '
'is not applicable (best_ckpt '
f'{best_ckpt_path} not found)')
if not test['test_last']:
return
test_dataset = build_dataset(cfg.data.test, dict(test_mode=True))
gpu_collect = cfg.get('evaluation', {}).get('gpu_collect', False)
tmpdir = cfg.get('evaluation', {}).get('tmpdir',
osp.join(cfg.work_dir, 'tmp'))
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('test_dataloader', {}))
test_dataloader = build_dataloader(test_dataset, **dataloader_setting)
names, ckpts = [], []
if test['test_last']:
names.append('last')
ckpts.append(None)
if test['test_best']:
names.append('best')
ckpts.append(best_ckpt_path)
for name, ckpt in zip(names, ckpts):
if ckpt is not None:
runner.load_checkpoint(ckpt)
outputs = multi_gpu_test(runner.model, test_dataloader, tmpdir,
gpu_collect)
rank, _ = get_dist_info()
if rank == 0:
out = osp.join(cfg.work_dir, f'{name}_pred.pkl')
test_dataset.dump_results(outputs, out)
eval_cfg = cfg.get('evaluation', {})
for key in [
'interval', 'tmpdir', 'start', 'gpu_collect',
'save_best', 'rule', 'by_epoch', 'broadcast_bn_buffers'
]:
eval_cfg.pop(key, None)
eval_res = test_dataset.evaluate(outputs, **eval_cfg)
runner.logger.info(f'Testing results of the {name} checkpoint')
for metric_name, val in eval_res.items():
runner.logger.info(f'{metric_name}: {val:.04f}')
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
test=dict(test_best=False, test_last=False),
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (:obj:`Dataset`): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
test (dict): The testing option, with two keys: test_last & test_best.
The value is True or False, indicating whether to test the
corresponding checkpoint.
Default: dict(test_best=False, test_last=False).
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(log_level=cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
persistent_workers=cfg.data.get('persistent_workers', False),
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('train_dataloader', {}))
if cfg.omnisource:
# The option can override videos_per_gpu
train_ratio = cfg.data.get('train_ratio', [1] * len(dataset))
omni_videos_per_gpu = cfg.data.get('omni_videos_per_gpu', None)
if omni_videos_per_gpu is None:
dataloader_settings = [dataloader_setting] * len(dataset)
else:
dataloader_settings = []
for videos_per_gpu in omni_videos_per_gpu:
this_setting = cp.deepcopy(dataloader_setting)
this_setting['videos_per_gpu'] = videos_per_gpu
dataloader_settings.append(this_setting)
data_loaders = [
build_dataloader(ds, **setting)
for ds, setting in zip(dataset, dataloader_settings)
]
else:
data_loaders = [
build_dataloader(ds, **dataloader_setting) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(model, device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
Runner = OmniSourceRunner if cfg.omnisource else EpochBasedRunner
runner = Runner(model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg,
distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config,
optimizer_config,
cfg.checkpoint_config,
cfg.log_config,
cfg.get('momentum_config', None),
custom_hooks_config=cfg.get(
'custom_hooks', None))
# multigrid setting
multigrid_cfg = cfg.get('multigrid', None)
if multigrid_cfg is not None:
from mmaction.utils.multigrid import LongShortCycleHook
multigrid_scheduler = LongShortCycleHook(cfg)
runner.register_hook(multigrid_scheduler)
logger.info('Finish register multigrid hook')
# subbn3d aggregation is HIGH, as it should be done before
# saving and evaluation
from mmaction.utils.multigrid import SubBatchNorm3dAggregationHook
subbn3d_aggre_hook = SubBatchNorm3dAggregationHook()
runner.register_hook(subbn3d_aggre_hook, priority='VERY_HIGH')
logger.info('Finish register subbn3daggre hook')
# precise bn setting
if cfg.get('precise_bn', False):
precise_bn_dataset = build_dataset(cfg.data.train)
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=1, # save memory and time
persistent_workers=cfg.data.get('persistent_workers', False),
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
data_loader_precise_bn = build_dataloader(precise_bn_dataset,
**dataloader_setting)
precise_bn_hook = PreciseBNHook(data_loader_precise_bn,
**cfg.get('precise_bn'))
runner.register_hook(precise_bn_hook, priority='HIGHEST')
logger.info('Finish register precisebn hook')
if distributed:
if cfg.omnisource:
runner.register_hook(OmniSourceDistSamplerSeedHook())
else:
runner.register_hook(DistSamplerSeedHook())
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
persistent_workers=cfg.data.get('persistent_workers', False),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook(val_dataloader, **eval_cfg) if distributed \
else EvalHook(val_dataloader, **eval_cfg)
runner.register_hook(eval_hook)
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner_kwargs = dict()
if cfg.omnisource:
runner_kwargs = dict(train_ratio=train_ratio)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs, **runner_kwargs)
if distributed:
dist.barrier()
time.sleep(5)
if test['test_last'] or test['test_best']:
best_ckpt_path = None
if test['test_best']:
ckpt_paths = [x for x in os.listdir(cfg.work_dir) if 'best' in x]
ckpt_paths = [x for x in ckpt_paths if x.endswith('.pth')]
if len(ckpt_paths) == 0:
runner.logger.info('Warning: test_best set, but no ckpt found')
test['test_best'] = False
if not test['test_last']:
return
elif len(ckpt_paths) > 1:
epoch_ids = [
int(x.split('epoch_')[-1][:-4]) for x in ckpt_paths
]
best_ckpt_path = ckpt_paths[np.argmax(epoch_ids)]
else:
best_ckpt_path = ckpt_paths[0]
if best_ckpt_path:
best_ckpt_path = osp.join(cfg.work_dir, best_ckpt_path)
test_dataset = build_dataset(cfg.data.test, dict(test_mode=True))
gpu_collect = cfg.get('evaluation', {}).get('gpu_collect', False)
tmpdir = cfg.get('evaluation', {}).get('tmpdir',
osp.join(cfg.work_dir, 'tmp'))
dataloader_setting = dict(
videos_per_gpu=cfg.data.get('videos_per_gpu', 1),
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
persistent_workers=cfg.data.get('persistent_workers', False),
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('test_dataloader', {}))
test_dataloader = build_dataloader(test_dataset, **dataloader_setting)
names, ckpts = [], []
if test['test_last']:
names.append('last')
ckpts.append(None)
if test['test_best'] and best_ckpt_path is not None:
names.append('best')
ckpts.append(best_ckpt_path)
for name, ckpt in zip(names, ckpts):
if ckpt is not None:
runner.load_checkpoint(ckpt)
outputs = multi_gpu_test(runner.model, test_dataloader, tmpdir,
gpu_collect)
rank, _ = get_dist_info()
if rank == 0:
out = osp.join(cfg.work_dir, f'{name}_pred.pkl')
test_dataset.dump_results(outputs, out)
eval_cfg = cfg.get('evaluation', {})
for key in [
'interval', 'tmpdir', 'start', 'gpu_collect',
'save_best', 'rule', 'by_epoch', 'broadcast_bn_buffers'
]:
eval_cfg.pop(key, None)
eval_res = test_dataset.evaluate(outputs, **eval_cfg)
runner.logger.info(f'Testing results of the {name} checkpoint')
for metric_name, val in eval_res.items():
runner.logger.info(f'{metric_name}: {val:.04f}')
def train_detector(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
logger = get_root_logger(log_level=cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
if 'imgs_per_gpu' in cfg.data:
logger.warning('"imgs_per_gpu" is deprecated in MMDet V2.0. '
'Please use "samples_per_gpu" instead')
if 'samples_per_gpu' in cfg.data:
logger.warning(
f'Got "imgs_per_gpu"={cfg.data.imgs_per_gpu} and '
f'"samples_per_gpu"={cfg.data.samples_per_gpu}, "imgs_per_gpu"'
f'={cfg.data.imgs_per_gpu} is used in this experiments')
else:
logger.warning(
'Automatically set "samples_per_gpu"="imgs_per_gpu"='
f'{cfg.data.imgs_per_gpu} in this experiments')
cfg.data.samples_per_gpu = cfg.data.imgs_per_gpu
runner_type = 'EpochBasedRunner' if 'runner' not in cfg else cfg.runner[
'type']
data_loaders = [
build_dataloader(
ds,
cfg.data.samples_per_gpu,
cfg.data.workers_per_gpu,
# `num_gpus` will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed,
runner_type=runner_type) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', False)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
if 'runner' not in cfg:
cfg.runner = {
'type': 'EpochBasedRunner',
'max_epochs': cfg.total_epochs
}
warnings.warn(
'config is now expected to have a `runner` section, '
'please set `runner` in your config.', UserWarning)
else:
if 'total_epochs' in cfg:
assert cfg.total_epochs == cfg.runner.max_epochs
runner = build_runner(
cfg.runner,
default_args=dict(
model=model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta))
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
if isinstance(runner, EpochBasedRunner):
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
# Support batch_size > 1 in validation
val_samples_per_gpu = cfg.data.val.pop('samples_per_gpu', 1)
if val_samples_per_gpu > 1:
# Replace 'ImageToTensor' to 'DefaultFormatBundle'
cfg.data.val.pipeline = replace_ImageToTensor(
cfg.data.val.pipeline)
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
val_dataloader = build_dataloader(
val_dataset,
samples_per_gpu=val_samples_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
eval_cfg = cfg.get('evaluation', {})
eval_cfg['by_epoch'] = cfg.runner['type'] != 'IterBasedRunner'
eval_hook = DistEvalHook if distributed else EvalHook
# In this PR (https://github.com/open-mmlab/mmcv/pull/1193), the
# priority of IterTimerHook has been modified from 'NORMAL' to 'LOW'.
runner.register_hook(
eval_hook(val_dataloader, **eval_cfg), priority='LOW')
# user-defined hooks
if cfg.get('custom_hooks', None):
custom_hooks = cfg.custom_hooks
assert isinstance(custom_hooks, list), \
f'custom_hooks expect list type, but got {type(custom_hooks)}'
for hook_cfg in cfg.custom_hooks:
assert isinstance(hook_cfg, dict), \
'Each item in custom_hooks expects dict type, but got ' \
f'{type(hook_cfg)}'
hook_cfg = hook_cfg.copy()
priority = hook_cfg.pop('priority', 'NORMAL')
hook = build_from_cfg(hook_cfg, HOOKS)
runner.register_hook(hook, priority=priority)
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow)