def fpn_feature_extractor(fpn_level, fea_channel):
layers = [
BasicConv(fea_channel, fea_channel, kernel_size=3, stride=1, padding=1)
]
for _ in range(fpn_level - 1):
layers.append(
BasicConv(fea_channel,
fea_channel,
kernel_size=3,
stride=2,
padding=1))
return nn.ModuleList(layers)
def __init__(self, channels, fea_channel):
super(CEM, self).__init__()
self.cv1 = BasicConv(channels[0],
fea_channel,
kernel_size=1,
padding=0)
self.cv2 = BasicConv(channels[1],
fea_channel,
kernel_size=1,
padding=0,
scale_factor=2)
self.gap = nn.AdaptiveAvgPool2d(1)
self.cv3 = BasicConv(channels[1],
fea_channel,
kernel_size=1,
padding=0)
def feature_transform_module(channels, fea_channel):
layers = []
for (i, channel) in enumerate(channels):
layers.append(
BasicConv(channel,
fea_channel,
kernel_size=1,
padding=0,
scale_factor=2**i))
return nn.ModuleList(layers)
def fpn_convs(fpn_level, fea_channel):
layers = []
for _ in range(fpn_level):
layers.append(
BasicConv(fea_channel,
fea_channel,
kernel_size=3,
stride=1,
padding=1))
return nn.ModuleList(layers)
def lateral_convs(fpn_level, fea_channel):
layers = []
for _ in range(fpn_level):
layers.append(BasicConv(fea_channel, fea_channel, kernel_size=1))
return nn.ModuleList(layers)