def __init__(self, rpool, whiten=None, L=3, eps=1e-6):
super(Rpool, self).__init__()
self.rpool = rpool
self.L = L
self.whiten = whiten
self.norm = L2N()
self.eps = eps
def __init__(self, features, pool, whiten, meta):
super(ImageRetrievalNet, self).__init__()
self.features = nn.Sequential(*features)
self.pool = pool
self.whiten = whiten
self.norm = L2N()
self.meta = meta
def __init__(self, features, lwhiten, pool, whiten, meta):
super(ImageRetrievalNet, self).__init__()
self.features = nn.Sequential(*features)
# hxq added
# self.features = nn.DataParallel(self.features)
self.lwhiten = lwhiten
self.pool = pool
self.whiten = whiten
self.norm = L2N()
self.meta = meta
def __init__(
self,
n_classes: int,
model_name: str = 'resnet50',
pretrained: bool = True,
pooling_name: str = 'adaptive', # 'GeM',
args_pooling: dict = {},
normalize: bool = True,
use_fc: bool = False,
fc_dim: int = 512,
dropout: float = 0.0,
loss_module: str = 'softmax'):
super().__init__()
self.backbone, final_in_features = self.get_backbone(
model_name, pretrained, num_classes=fc_dim)
if pooling_name in ('AdaptiveAvgPool2d', 'adaptive'):
self.pooling = nn.AdaptiveAvgPool2d(1)
elif pooling_name in ('MAC', 'SPoC', 'GeM', 'GeMmp', 'RMAC', 'Rpool'):
self.pooling = getattr(cirtorch.pooling,
pooling_name)(**args_pooling)
elif pooling_name is None or pooling_name.lower() == 'identity':
self.pooling = nn.Identity()
else:
raise ValueError("Incorrect pooling name")
if normalize:
self.norm = L2N()
else:
self.norm = None
self.use_fc = use_fc
if use_fc:
self.final_block = nn.Sequential(
OrderedDict([('bn1', nn.BatchNorm1d(final_in_features)),
('dropout', nn.Dropout(p=dropout)),
('fc2', nn.Linear(final_in_features, fc_dim)),
('bn2', nn.BatchNorm1d(fc_dim))]))
self._init_params()
final_in_features = fc_dim
self.loss_module = loss_module
if loss_module == 'arcface':
self.final = ArcMarginProduct(final_in_features, n_classes)
else:
self.final = nn.Linear(final_in_features, n_classes)
def __init__(self, d_in, d_out):
super(Whiten_layer, self).__init__()
self.w = nn.Linear(d_in, d_out)
self.norm = L2N()