def __init__(self,
embed_dims,
num_heads,
attn_drop=0.,
proj_drop=0.,
dropout_layer=None,
init_cfg=None,
batch_first=True,
qkv_bias=False,
norm_cfg=dict(type='LN'),
sr_ratio=1):
super().__init__(
embed_dims,
num_heads,
attn_drop,
proj_drop,
dropout_layer=dropout_layer,
init_cfg=init_cfg,
batch_first=batch_first,
bias=qkv_bias)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = Conv2d(
in_channels=embed_dims,
out_channels=embed_dims,
kernel_size=sr_ratio,
stride=sr_ratio)
# The ret[0] of build_norm_layer is norm name.
self.norm = build_norm_layer(norm_cfg, embed_dims)[1]
# handle the BC-breaking from https://github.com/open-mmlab/mmcv/pull/1418 # noqa
from mmseg import digit_version, mmcv_version
if mmcv_version < digit_version('1.3.17'):
warnings.warn('The legacy version of forward function in'
'EfficientMultiheadAttention is deprecated in'
'mmcv>=1.3.17 and will no longer support in the'
'future. Please upgrade your mmcv.')
self.forward = self.legacy_forward
def test_digit_version():
assert digit_version('0.2.16') == (0, 2, 16, 0, 0, 0)
assert digit_version('1.2.3') == (1, 2, 3, 0, 0, 0)
assert digit_version('1.2.3rc0') == (1, 2, 3, 0, -1, 0)
assert digit_version('1.2.3rc1') == (1, 2, 3, 0, -1, 1)
assert digit_version('1.0rc0') == (1, 0, 0, 0, -1, 0)
assert digit_version('1.0') == digit_version('1.0.0')
assert digit_version('1.5.0+cuda90_cudnn7.6.3_lms') == digit_version('1.5')
assert digit_version('1.0.0dev') < digit_version('1.0.0a')
assert digit_version('1.0.0a') < digit_version('1.0.0a1')
assert digit_version('1.0.0a') < digit_version('1.0.0b')
assert digit_version('1.0.0b') < digit_version('1.0.0rc')
assert digit_version('1.0.0rc1') < digit_version('1.0.0')
assert digit_version('1.0.0') < digit_version('1.0.0post')
assert digit_version('1.0.0post') < digit_version('1.0.0post1')
assert digit_version('v1') == (1, 0, 0, 0, 0, 0)
assert digit_version('v1.1.5') == (1, 1, 5, 0, 0, 0)
def main():
args = parse_args()
assert args.out or args.eval or args.format_only or args.show \
or args.show_dir, \
('Please specify at least one operation (save/eval/format/show the '
'results / save the results) with the argument "--out", "--eval"'
', "--format-only", "--show" or "--show-dir"')
if args.eval and args.format_only:
raise ValueError('--eval and --format_only cannot be both specified')
if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
raise ValueError('The output file must be a pkl file.')
cfg = mmcv.Config.fromfile(args.config)
if args.cfg_options is not None:
cfg.merge_from_dict(args.cfg_options)
# set multi-process settings
setup_multi_processes(cfg)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
if args.aug_test:
# hard code index
cfg.data.test.pipeline[1].img_ratios = [
0.5, 0.75, 1.0, 1.25, 1.5, 1.75
]
cfg.data.test.pipeline[1].flip = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
if args.gpu_id is not None:
cfg.gpu_ids = [args.gpu_id]
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
cfg.gpu_ids = [args.gpu_id]
distributed = False
if len(cfg.gpu_ids) > 1:
warnings.warn(f'The gpu-ids is reset from {cfg.gpu_ids} to '
f'{cfg.gpu_ids[0:1]} to avoid potential error in '
'non-distribute testing time.')
cfg.gpu_ids = cfg.gpu_ids[0:1]
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
rank, _ = get_dist_info()
# allows not to create
if args.work_dir is not None and rank == 0:
mmcv.mkdir_or_exist(osp.abspath(args.work_dir))
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
if args.aug_test:
json_file = osp.join(args.work_dir,
f'eval_multi_scale_{timestamp}.json')
else:
json_file = osp.join(args.work_dir,
f'eval_single_scale_{timestamp}.json')
elif rank == 0:
work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
mmcv.mkdir_or_exist(osp.abspath(work_dir))
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
if args.aug_test:
json_file = osp.join(work_dir,
f'eval_multi_scale_{timestamp}.json')
else:
json_file = osp.join(work_dir,
f'eval_single_scale_{timestamp}.json')
# build the dataloader
# TODO: support multiple images per gpu (only minor changes are needed)
dataset = build_dataset(cfg.data.test)
# The default loader config
loader_cfg = dict(
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
# The overall dataloader settings
loader_cfg.update({
k: v
for k, v in cfg.data.items() if k not in [
'train', 'val', 'test', 'train_dataloader', 'val_dataloader',
'test_dataloader'
]
})
test_loader_cfg = {
**loader_cfg,
'samples_per_gpu': 1,
'shuffle': False, # Not shuffle by default
**cfg.data.get('test_dataloader', {})
}
# build the dataloader
data_loader = build_dataloader(dataset, **test_loader_cfg)
# build the model and load checkpoint
cfg.model.train_cfg = None
model = build_segmentor(cfg.model, test_cfg=cfg.get('test_cfg'))
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
checkpoint = load_checkpoint(model, args.checkpoint, map_location='cpu')
if 'CLASSES' in checkpoint.get('meta', {}):
model.CLASSES = checkpoint['meta']['CLASSES']
else:
print('"CLASSES" not found in meta, use dataset.CLASSES instead')
model.CLASSES = dataset.CLASSES
if 'PALETTE' in checkpoint.get('meta', {}):
model.PALETTE = checkpoint['meta']['PALETTE']
else:
print('"PALETTE" not found in meta, use dataset.PALETTE instead')
model.PALETTE = dataset.PALETTE
# clean gpu memory when starting a new evaluation.
torch.cuda.empty_cache()
eval_kwargs = {} if args.eval_options is None else args.eval_options
# Deprecated
efficient_test = eval_kwargs.get('efficient_test', False)
if efficient_test:
warnings.warn(
'``efficient_test=True`` does not have effect in tools/test.py, '
'the evaluation and format results are CPU memory efficient by '
'default')
eval_on_format_results = (args.eval is not None
and 'cityscapes' in args.eval)
if eval_on_format_results:
assert len(args.eval) == 1, 'eval on format results is not ' \
'applicable for metrics other than ' \
'cityscapes'
if args.format_only or eval_on_format_results:
if 'imgfile_prefix' in eval_kwargs:
tmpdir = eval_kwargs['imgfile_prefix']
else:
tmpdir = '.format_cityscapes'
eval_kwargs.setdefault('imgfile_prefix', tmpdir)
mmcv.mkdir_or_exist(tmpdir)
else:
tmpdir = None
if not distributed:
warnings.warn(
'SyncBN is only supported with DDP. To be compatible with DP, '
'we convert SyncBN to BN. Please use dist_train.sh which can '
'avoid this error.')
if not torch.cuda.is_available():
assert digit_version(mmcv.__version__) >= digit_version('1.4.4'), \
'Please use MMCV >= 1.4.4 for CPU training!'
model = revert_sync_batchnorm(model)
model = MMDataParallel(model, device_ids=cfg.gpu_ids)
results = single_gpu_test(model,
data_loader,
args.show,
args.show_dir,
False,
args.opacity,
pre_eval=args.eval is not None
and not eval_on_format_results,
format_only=args.format_only
or eval_on_format_results,
format_args=eval_kwargs)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
results = multi_gpu_test(model,
data_loader,
args.tmpdir,
args.gpu_collect,
False,
pre_eval=args.eval is not None
and not eval_on_format_results,
format_only=args.format_only
or eval_on_format_results,
format_args=eval_kwargs)
rank, _ = get_dist_info()
if rank == 0:
if args.out:
warnings.warn(
'The behavior of ``args.out`` has been changed since MMSeg '
'v0.16, the pickled outputs could be seg map as type of '
'np.array, pre-eval results or file paths for '
'``dataset.format_results()``.')
print(f'\nwriting results to {args.out}')
mmcv.dump(results, args.out)
if args.eval:
eval_kwargs.update(metric=args.eval)
metric = dataset.evaluate(results, **eval_kwargs)
metric_dict = dict(config=args.config, metric=metric)
mmcv.dump(metric_dict, json_file, indent=4)
if tmpdir is not None and eval_on_format_results:
# remove tmp dir when cityscapes evaluation
shutil.rmtree(tmpdir)
def train_segmentor(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
"""Launch segmentor training."""
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
# The default loader config
loader_cfg = dict(
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed,
drop_last=True)
# The overall dataloader settings
loader_cfg.update({
k: v
for k, v in cfg.data.items() if k not in [
'train', 'val', 'test', 'train_dataloader', 'val_dataloader',
'test_dataloader'
]
})
# The specific dataloader settings
train_loader_cfg = {**loader_cfg, **cfg.data.get('train_dataloader', {})}
data_loaders = [build_dataloader(ds, **train_loader_cfg) 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:
if not torch.cuda.is_available():
assert digit_version(mmcv.__version__) >= digit_version('1.4.4'), \
'Please use MMCV >= 1.4.4 for CPU training!'
model = MMDataParallel(model, device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
if cfg.get('runner') is None:
cfg.runner = {'type': 'IterBasedRunner', 'max_iters': cfg.total_iters}
warnings.warn(
'config is now expected to have a `runner` section, '
'please set `runner` in your config.', UserWarning)
runner = build_runner(cfg.runner,
default_args=dict(model=model,
batch_processor=None,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta))
# register hooks
runner.register_training_hooks(cfg.lr_config, cfg.optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
# when distributed training by epoch, using`DistSamplerSeedHook` to set
# the different seed to distributed sampler for each epoch, it will
# shuffle dataset at each epoch and avoid overfitting.
if isinstance(runner, EpochBasedRunner):
runner.register_hook(DistSamplerSeedHook())
# an ugly walkaround to make the .log and .log.json filenames the same
runner.timestamp = timestamp
# register eval hooks
if validate:
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
# The specific dataloader settings
val_loader_cfg = {
**loader_cfg,
'samples_per_gpu': 1,
'shuffle': False, # Not shuffle by default
**cfg.data.get('val_dataloader', {}),
}
val_dataloader = build_dataloader(val_dataset, **val_loader_cfg)
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 is None and cfg.get('auto_resume'):
resume_from = find_latest_checkpoint(cfg.work_dir)
if resume_from is not None:
cfg.resume_from = resume_from
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)