def __init__(self, **kwargs):
super(retina_fpn, self).__init__(**kwargs)
self.conv5 = conv_layer(256, 1, 1, name='P5_conv')
self.up_pool = nn.UpSampling2D(size=(2, 2), interpolation='bilinear')
self.conv4 = conv_layer(256, 1, 1, name='P4_conv')
self.conv3 = conv_layer(256, 1, 1, name='P3_conv')
self.conv6 = conv_layer(256, 3, 2, padding='same', name='P6_conv')
self.conv7 = conv_layer(256, 3, 2, padding='same', name='P7_conv')
self.add = nn.Add()
def __init__(self, filters3x3, filters1x1, **kwargs):
super(OSA_module, self).__init__(**kwargs)
self.conv_layers = []
for i in range(5):
self.conv_layers.append(
conv_layer(filters3x3, 3, 1, name='conv{}'.format(i + 1)))
self.concat = nn.Concatenate(axis=-1, name='concat')
self.conv1x1 = conv_layer(filters1x1, 1, 1, name='conv1x1')
def __init__(self, **kwargs):
super(regression, self).__init__(**kwargs)
self.conv1 = conv_layer(256, 3, 1, 'same', name='conv1')
self.conv2 = conv_layer(256, 3, 1, 'same', name='conv2')
self.conv3 = conv_layer(256, 3, 1, 'same', name='conv3')
self.conv4 = conv_layer(256, 3, 1, 'same', name='conv4')
self.regression = nn.Conv2D(4,
1,
1,
activation=None,
name='regression')
def __init__(self, out_channel, **kwargs):
'''
fcos classification head
:param num_classes:
:param kwargs:
'''
super(classification, self).__init__(**kwargs)
self.conv1 = conv_layer(256, 3, 1, 'same', name='conv1')
self.conv2 = conv_layer(256, 3, 1, 'same', name='conv2')
self.conv3 = conv_layer(256, 3, 1, 'same', name='conv3')
self.conv4 = conv_layer(256, 3, 1, 'same', name='conv4')
# 加1表示center_ness
self.classification = nn.Conv2D(out_channel,
1,
1,
activation='sigmoid',
name='classification')
def __init__(self, nun_classes, **kwargs):
'''
model unit
:param nun_classes: output classes
:param kwargs:
'''
super(FcosHead, self).__init__(**kwargs)
self.conv1 = conv_layer(256, 3, 1, 'same', name='conv1')
self.conv2 = conv_layer(256, 3, 1, 'same', name='conv2')
self.conv3 = conv_layer(256, 3, 1, 'same', name='conv3')
self.conv4 = conv_layer(256, 3, 1, 'same', name='conv4')
# 论文中未强调 classification和regression不能共享权重
self.classification = nn.Conv2D(
nun_classes + 1, 1, 1, activation='sigmoid',
name='classification') # num_classes: classification
# 1: center-ness
self.regression = nn.Conv2D(4, 1, 1,
name='regression') #linear activation
self.concat = nn.Concatenate(axis=-1) # stack all result
def __init__(self, **kwargs):
super(stem, self).__init__(**kwargs)
self.conv1 = conv_layer(64, 3, 2, name='conv1')
self.conv2 = conv_layer(64, 3, 1, name='conv2')
self.conv3 = conv_layer(128, 3, 1, name='conv3')