def init_weights(self):
if (isinstance(self.init_cfg, dict)
and self.init_cfg.get('type') == 'Pretrained'):
logger = get_root_logger()
checkpoint = CheckpointLoader.load_checkpoint(
self.init_cfg['checkpoint'], logger=logger, map_location='cpu')
if 'state_dict' in checkpoint:
state_dict = checkpoint['state_dict']
else:
state_dict = checkpoint
if 'pos_embed' in state_dict.keys():
if self.pos_embed.shape != state_dict['pos_embed'].shape:
logger.info(msg=f'Resize the pos_embed shape from '
f'{state_dict["pos_embed"].shape} to '
f'{self.pos_embed.shape}')
h, w = self.img_size
pos_size = int(
math.sqrt(state_dict['pos_embed'].shape[1] - 1))
state_dict['pos_embed'] = self.resize_pos_embed(
state_dict['pos_embed'],
(h // self.patch_size, w // self.patch_size),
(pos_size, pos_size), self.interpolate_mode)
load_state_dict(self, state_dict, strict=False, logger=logger)
elif self.init_cfg is not None:
super(VisionTransformer, self).init_weights()
else:
# We only implement the 'jax_impl' initialization implemented at
# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py#L353 # noqa: E501
trunc_normal_(self.pos_embed, std=.02)
trunc_normal_(self.cls_token, std=.02)
for n, m in self.named_modules():
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if m.bias is not None:
if 'ffn' in n:
nn.init.normal_(m.bias, mean=0., std=1e-6)
else:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Conv2d):
kaiming_init(m, mode='fan_in', bias=0.)
elif isinstance(m, (_BatchNorm, nn.GroupNorm, nn.LayerNorm)):
constant_init(m, val=1.0, bias=0.)
def main():
parser = argparse.ArgumentParser(
description='Convert keys in official pretrained swin models to'
'MMSegmentation style.')
parser.add_argument('src', help='src model path or url')
# The dst path must be a full path of the new checkpoint.
parser.add_argument('dst', help='save path')
args = parser.parse_args()
checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
if 'state_dict' in checkpoint:
state_dict = checkpoint['state_dict']
elif 'model' in checkpoint:
state_dict = checkpoint['model']
else:
state_dict = checkpoint
weight = convert_swin(state_dict)
mmcv.mkdir_or_exist(osp.dirname(args.dst))
torch.save(weight, args.dst)
def main():
parser = argparse.ArgumentParser(
description='Convert keys in official pretrained STDC1/2 to '
'MMSegmentation style.')
parser.add_argument('src', help='src model path')
# The dst path must be a full path of the new checkpoint.
parser.add_argument('dst', help='save path')
parser.add_argument('type', help='model type: STDC1 or STDC2')
args = parser.parse_args()
checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
if 'state_dict' in checkpoint:
state_dict = checkpoint['state_dict']
elif 'model' in checkpoint:
state_dict = checkpoint['model']
else:
state_dict = checkpoint
assert args.type in ['STDC1',
'STDC2'], 'STD type should be STDC1 or STDC2!'
weight = convert_stdc(state_dict, args.type)
mmcv.mkdir_or_exist(osp.dirname(args.dst))
torch.save(weight, args.dst)
def test_checkpoint_loader():
from mmcv.runner import _load_checkpoint, save_checkpoint, CheckpointLoader
import tempfile
import os
checkpoint_path = os.path.join(tempfile.gettempdir(), 'checkpoint.pth')
model = Model()
save_checkpoint(model, checkpoint_path)
checkpoint = _load_checkpoint(checkpoint_path)
assert 'meta' in checkpoint and 'CLASSES' not in checkpoint['meta']
# remove the temp file
os.remove(checkpoint_path)
filenames = [
'http://xx.xx/xx.pth', 'https://xx.xx/xx.pth',
'modelzoo://xx.xx/xx.pth', 'torchvision://xx.xx/xx.pth',
'open-mmlab://xx.xx/xx.pth', 'openmmlab://xx.xx/xx.pth',
'mmcls://xx.xx/xx.pth', 'pavi://xx.xx/xx.pth', 's3://xx.xx/xx.pth',
'ss3://xx.xx/xx.pth', ' s3://xx.xx/xx.pth'
]
fn_names = [
'load_from_http', 'load_from_http', 'load_from_torchvision',
'load_from_torchvision', 'load_from_openmmlab', 'load_from_openmmlab',
'load_from_mmcls', 'load_from_pavi', 'load_from_ceph',
'load_from_local', 'load_from_local'
]
for filename, fn_name in zip(filenames, fn_names):
loader = CheckpointLoader._get_checkpoint_loader(filename)
assert loader.__name__ == fn_name
@CheckpointLoader.register_scheme(prefixes='ftp://')
def load_from_ftp(filename, map_location):
return dict(filename=filename)
# test register_loader
filename = 'ftp://xx.xx/xx.pth'
loader = CheckpointLoader._get_checkpoint_loader(filename)
assert loader.__name__ == 'load_from_ftp'
def load_from_ftp1(filename, map_location):
return dict(filename=filename)
# test duplicate registered error
with pytest.raises(KeyError):
CheckpointLoader.register_scheme('ftp://', load_from_ftp1)
# test force param
CheckpointLoader.register_scheme('ftp://', load_from_ftp1, force=True)
checkpoint = CheckpointLoader.load_checkpoint(filename)
assert checkpoint['filename'] == filename
# test print function name
loader = CheckpointLoader._get_checkpoint_loader(filename)
assert loader.__name__ == 'load_from_ftp1'
# test sort
@CheckpointLoader.register_scheme(prefixes='a/b')
def load_from_ab(filename, map_location):
return dict(filename=filename)
@CheckpointLoader.register_scheme(prefixes='a/b/c')
def load_from_abc(filename, map_location):
return dict(filename=filename)
filename = 'a/b/c/d'
loader = CheckpointLoader._get_checkpoint_loader(filename)
assert loader.__name__ == 'load_from_abc'
def mmpose2torchserve(config_file: str,
checkpoint_file: str,
output_folder: str,
model_name: str,
model_version: str = '1.0',
force: bool = False):
"""Converts MMPose model (config + checkpoint) to TorchServe `.mar`.
Args:
config_file:
In MMPose config format.
The contents vary for each task repository.
checkpoint_file:
In MMPose checkpoint format.
The contents vary for each task repository.
output_folder:
Folder where `{model_name}.mar` will be created.
The file created will be in TorchServe archive format.
model_name:
If not None, used for naming the `{model_name}.mar` file
that will be created under `output_folder`.
If None, `{Path(checkpoint_file).stem}` will be used.
model_version:
Model's version.
force:
If True, if there is an existing `{model_name}.mar`
file under `output_folder` it will be overwritten.
"""
mmcv.mkdir_or_exist(output_folder)
config = mmcv.Config.fromfile(config_file)
with TemporaryDirectory() as tmpdir:
model_file = osp.join(tmpdir, 'config.py')
config.dump(model_file)
handler_path = osp.join(osp.dirname(__file__), 'mmpose_handler.py')
model_name = model_name or osp.splitext(
osp.basename(checkpoint_file))[0]
# use mmcv CheckpointLoader if checkpoint is not from a local file
if not osp.isfile(checkpoint_file):
ckpt = CheckpointLoader.load_checkpoint(checkpoint_file)
checkpoint_file = osp.join(tmpdir, 'checkpoint.pth')
with open(checkpoint_file, 'wb') as f:
torch.save(ckpt, f)
args = Namespace(
**{
'model_file': model_file,
'serialized_file': checkpoint_file,
'handler': handler_path,
'model_name': model_name,
'version': model_version,
'export_path': output_folder,
'force': force,
'requirements_file': None,
'extra_files': None,
'runtime': 'python',
'archive_format': 'default'
})
manifest = ModelExportUtils.generate_manifest_json(args)
package_model(args, manifest)
def init_weights(self):
logger = get_root_logger()
if self.init_cfg is None:
logger.warn(f'No pre-trained weights for '
f'{self.__class__.__name__}, '
f'training start from scratch')
if self.use_abs_pos_embed:
trunc_normal_(self.absolute_pos_embed, std=0.02)
for m in self.modules():
if isinstance(m, nn.Linear):
trunc_normal_init(m, std=.02, bias=0.)
elif isinstance(m, nn.LayerNorm):
constant_init(m, val=1.0, bias=0.)
else:
assert 'checkpoint' in self.init_cfg, f'Only support ' \
f'specify `Pretrained` in ' \
f'`init_cfg` in ' \
f'{self.__class__.__name__} '
ckpt = CheckpointLoader.load_checkpoint(
self.init_cfg['checkpoint'], logger=logger, map_location='cpu')
if 'state_dict' in ckpt:
_state_dict = ckpt['state_dict']
elif 'model' in ckpt:
_state_dict = ckpt['model']
else:
_state_dict = ckpt
state_dict = OrderedDict()
for k, v in _state_dict.items():
if k.startswith('backbone.'):
state_dict[k[9:]] = v
else:
state_dict[k] = v
# strip prefix of state_dict
if list(state_dict.keys())[0].startswith('module.'):
state_dict = {k[7:]: v for k, v in state_dict.items()}
# reshape absolute position embedding
if state_dict.get('absolute_pos_embed') is not None:
absolute_pos_embed = state_dict['absolute_pos_embed']
N1, L, C1 = absolute_pos_embed.size()
N2, C2, H, W = self.absolute_pos_embed.size()
if N1 != N2 or C1 != C2 or L != H * W:
logger.warning('Error in loading absolute_pos_embed, pass')
else:
state_dict['absolute_pos_embed'] = absolute_pos_embed.view(
N2, H, W, C2).permute(0, 3, 1, 2).contiguous()
# interpolate position bias table if needed
relative_position_bias_table_keys = [
k for k in state_dict.keys()
if 'relative_position_bias_table' in k
]
for table_key in relative_position_bias_table_keys:
table_pretrained = state_dict[table_key]
table_current = self.state_dict()[table_key]
L1, nH1 = table_pretrained.size()
L2, nH2 = table_current.size()
if nH1 != nH2:
logger.warning(f'Error in loading {table_key}, pass')
elif L1 != L2:
S1 = int(L1**0.5)
S2 = int(L2**0.5)
table_pretrained_resized = F.interpolate(
table_pretrained.permute(1, 0).reshape(1, nH1, S1, S1),
size=(S2, S2),
mode='bicubic')
state_dict[table_key] = table_pretrained_resized.view(
nH2, L2).permute(1, 0).contiguous()
# load state_dict
load_state_dict(self, state_dict, strict=False, logger=logger)