def __init__(self,
backbone,
cls_head,
neck=None,
train_cfg=None,
test_cfg=None):
super().__init__()
# The backbones in mmcls can be used by TSN
if backbone['type'].startswith('mmcls.'):
try:
import mmcls.models.builder as mmcls_builder
except (ImportError, ModuleNotFoundError):
raise ImportError('Please install mmcls to use this backbone.')
backbone['type'] = backbone['type'][6:]
self.backbone = mmcls_builder.build_backbone(backbone)
else:
self.backbone = builder.build_backbone(backbone)
if neck is not None:
self.neck = builder.build_neck(neck)
self.cls_head = builder.build_head(cls_head)
self.train_cfg = train_cfg
self.test_cfg = test_cfg
# aux_info is the list of tensor names beyond 'imgs' and 'label' which
# will be used in train_step and val_step, data_batch should contain
# these tensors
self.aux_info = []
if train_cfg is not None and 'aux_info' in train_cfg:
self.aux_info = train_cfg['aux_info']
# max_testing_views should be int
self.max_testing_views = None
if test_cfg is not None and 'max_testing_views' in test_cfg:
self.max_testing_views = test_cfg['max_testing_views']
assert isinstance(self.max_testing_views, int)
self.init_weights()
self.fp16_enabled = False
def __init__(self,
backbone,
cls_head,
neck=None,
train_cfg=None,
test_cfg=None):
super().__init__()
# record the source of the backbone
self.backbone_from = 'mmaction2'
if backbone['type'].startswith('mmcls.'):
try:
import mmcls.models.builder as mmcls_builder
except (ImportError, ModuleNotFoundError):
raise ImportError('Please install mmcls to use this backbone.')
backbone['type'] = backbone['type'][6:]
self.backbone = mmcls_builder.build_backbone(backbone)
self.backbone_from = 'mmcls'
elif backbone['type'].startswith('torchvision.'):
try:
import torchvision.models
except (ImportError, ModuleNotFoundError):
raise ImportError('Please install torchvision to use this '
'backbone.')
backbone_type = backbone.pop('type')[12:]
self.backbone = torchvision.models.__dict__[backbone_type](
**backbone)
# disable the classifier
self.backbone.classifier = nn.Identity()
self.backbone.fc = nn.Identity()
self.backbone_from = 'torchvision'
else:
self.backbone = builder.build_backbone(backbone)
if neck is not None:
self.neck = builder.build_neck(neck)
self.cls_head = builder.build_head(cls_head)
self.train_cfg = train_cfg
self.test_cfg = test_cfg
# aux_info is the list of tensor names beyond 'imgs' and 'label' which
# will be used in train_step and val_step, data_batch should contain
# these tensors
self.aux_info = []
if train_cfg is not None and 'aux_info' in train_cfg:
self.aux_info = train_cfg['aux_info']
# max_testing_views should be int
self.max_testing_views = None
if test_cfg is not None and 'max_testing_views' in test_cfg:
self.max_testing_views = test_cfg['max_testing_views']
assert isinstance(self.max_testing_views, int)
# mini-batch blending, e.g. mixup, cutmix, etc.
self.blending = None
if train_cfg is not None and 'blending' in train_cfg:
from mmcv.utils import build_from_cfg
from mmaction.datasets.builder import BLENDINGS
self.blending = build_from_cfg(train_cfg['blending'], BLENDINGS)
self.init_weights()
self.fp16_enabled = False