def __init__()

in utils/gluon/utils/resnetv1.py [0:0]

• 46 lines of code
• 16 McCabe index (conditional complexity)

    def __init__(self, in_planes, mid_planes, out_planes, groups=1, strides=1,
                 norm_kwargs=None, last_gamma=False, name_prefix=None,
                 down_pos=0, use_se=False, se_planes=-1, extra_bn=False,
                 **kwargs):
        super(_BottleneckV1, self).__init__(prefix=name_prefix)
        assert down_pos in [0, 1, 2], \
            "down_pos value({}) is unknown.".format(down_pos)
        strides1 = strides if down_pos == 0 else 1
        strides2 = strides if down_pos == 1 else 1
        strides3 = strides if down_pos == 2 else 1
        with self.name_scope():
            if extra_bn:
                # note: se-block seems not converage well on ResNet152
                self.bn1a = nn.BatchNorm(in_channels=in_planes, prefix='bn1a',
                                  center=False, scale=False,
                                  **({} if norm_kwargs is None else norm_kwargs))
            # extract information
            self.conv1 = nn.Conv2D(channels=mid_planes, in_channels=in_planes,
                                  kernel_size=1, use_bias=False, strides=strides1,
                                  prefix='conv1')
            self.bn1 = nn.BatchNorm(in_channels=mid_planes, prefix='bn1',
                                  **({} if norm_kwargs is None else norm_kwargs))
            self.relu1 = nn.Activation('relu')
            # capture spatial relations
            self.conv2 = nn.Conv2D(channels=mid_planes, in_channels=mid_planes,
                                  kernel_size=3, padding=1, groups=groups,
                                  strides=strides2, use_bias=False, prefix='conv2')
            self.bn2 = nn.BatchNorm(in_channels=mid_planes, prefix='bn2',
                                  **({} if norm_kwargs is None else norm_kwargs))
            self.relu2 = nn.Activation('relu')
            # embeding back to information highway
            self.conv3 = nn.Conv2D(channels=out_planes, in_channels=mid_planes,
                                  kernel_size=1, use_bias=False, strides=strides3,
                                  prefix='conv3')
            self.bn3 = nn.BatchNorm(in_channels=out_planes, prefix='bn3',
                                  gamma_initializer='zeros' if last_gamma else 'ones',
                                  **({} if norm_kwargs is None else norm_kwargs))

            self.se_block = nn.SE(in_channels=out_planes, channels=se_planes,
                                  prefix='se') if use_se else None

            # this relue is added after fusion
            self.relu3 = nn.Activation('relu')

            if strides != 1 or in_planes != out_planes:
                if extra_bn:
                    # note: se-block seems not converage well on ResNet152
                    self.bn4a = nn.BatchNorm(in_channels=in_planes, prefix='bn4a',
                                  center=False, scale=False,
                                  **({} if norm_kwargs is None else norm_kwargs))
                self.conv4 = nn.Conv2D(channels=out_planes, in_channels=in_planes,
                                  kernel_size=1, strides=strides, use_bias=False,
                                  prefix='conv4')
                self.bn4 = nn.BatchNorm(in_channels=out_planes, prefix='bn4',
                                  **({} if norm_kwargs is None else norm_kwargs))