def __init__(self, args, **kwargs):
super(ArcMarginBlock, self).__init__(**kwargs)
self.margin_s = args.margin_s
self.margin_m = args.margin_m
self.margin_a = args.margin_a
self.margin_b = args.margin_b
self.num_classes = args.num_classes
self.emb_size = args.emb_size
# self.weight = gluon.Parameter(name = 'fc7_weight', shape = (self.num_classes, self.emb_size))
# self.weight.initialize()
# self._weight = nd.empty(shape = (self.num_classes, self.emb_size))
# if self.margin_a>0.0:
with self.name_scope():
self.fc7_weight = self.params.get('fc7_weight',
shape=(self.num_classes,
self.emb_size))
# else:
# self.dense = nn.Dense(self.num_classes, prefix='fc7')
self.body = nn.HybridSequential(prefix='')
feat = fresnet.get(args.num_layers,
version_unit=args.version_unit,
version_act=args.version_act)
self.body.add(feat)
self.body.add(
EmbeddingBlock(args.emb_size, args.version_output, prefix=''))
def __init__(self, args, **kwargs):
super(TestBlock, self).__init__(**kwargs)
feat_with_bn = False
use_dropout = True
if args.num_classes >= 20000:
use_dropout = False
with self.name_scope():
self.feature_scale = None
self.feat_net = nn.HybridSequential(prefix='')
self.feat_net.add(
fresnet.get(args.num_layers, args.emb_size, use_dropout))
if feat_with_bn:
#self.feat_net.add(nn.BatchNorm(scale=True, epsilon=2e-5))
#self.bn1_gamma = self.params.get('bn1_gamma', shape=(1, args.emb_size))
#self.bn1_gamma.initialize(init=mx.init.Constant(1.0))
self.bn1_beta = self.params.get('bn1_beta',
shape=(1, args.emb_size))
self.bn1_beta.initialize(init=mx.init.Constant(0.0))
#else:
# self.bn1_beta = self.params.get('bn1_beta', shape=(1, args.emb_size))
# self.bn1_beta.initialize(init=mx.init.Constant(0.0))
#self.feat_net = fresnet.get(args.num_layers, args.emb_size, use_dropout)
#initializer = mx.init.Xavier(rnd_type='gaussian', factor_type="out", magnitude=2) #resnet style
#self.feat_net.initialize(init=initializer)
self.feat_with_bn = feat_with_bn
def __init__(self, args, **kwargs):
super(ArcMarginTestBlock, self).__init__(**kwargs)
self.body = nn.HybridSequential(prefix='')
feat = fresnet.get(args.num_layers,
version_unit=args.version_unit,
version_act=args.version_act)
self.body.add(feat)
self.body.add(EmbeddingBlock(args.emb_size, args.version_output))
def __init__(self, args, num_classes, **kwargs):
super(_GABlock, self).__init__(**kwargs)
with self.name_scope():
self.body = nn.HybridSequential(prefix='')
feat = fresnet.get(args.num_layers,
version_unit=args.version_unit,
version_act=args.version_act)
self.body.add(feat)
self.body.add(EmbeddingBlock(mode=args.version_output))
self.body.add(nn.Dense(num_classes))
def __init__(self, args, **kwargs):
super(DenseBlock, self).__init__(**kwargs)
self.num_classes = args.num_classes
self.emb_size = args.emb_size
self.body = nn.HybridSequential(prefix='')
feat = fresnet.get(args.num_layers,
version_unit=args.version_unit,
version_act=args.version_act)
self.body.add(feat)
self.body.add(EmbeddingBlock(args.emb_size, args.version_output, prefix=''))
self.dense = nn.Dense(self.num_classes, prefix='fc7')
def __init__(self, args, is_train, **kwargs):
super(FeatBlock, self).__init__(**kwargs)
with self.name_scope():
self.feat_net = nn.HybridSequential(prefix='')
self.feat_net.add(
fresnet.get(args.num_layers, args.emb_size, args.use_dropout))
self.is_train = is_train
if is_train:
initializer = mx.init.Xavier(rnd_type='gaussian',
factor_type="out",
magnitude=2) #resnet style
self.feat_net.initialize(init=initializer)