Skip to content

separable_conv2

2D separable convolution.

This layer performs separable convolution operations on an input.
A separable convolution is composed of a depthwise convolution (3x3) and a point convolution (1x1).
The padding rule is compatible with Keras (TensorFlow) & Caffe.

Parameters

  • strides : Integer(s), stride values.
    Two integers mean [stride_y, stride_x].
    One integer means stride_y and stride_x have the same value.
    The default is 1.
  • padding : String,
    "save" and "valid" are compatible with Keras (TensorFlow).
    "general" is compatible with Caffe.
  • pads : Integer(s), pad values to the input.
    Four integers mean [pad_top, pad_left, pad_bottom, pad_right].
    Two integers mean [pad_top, pad_left] where pad_bottom and pad_right are same as pad_top and pad_left respectively.
    One integer means that pad_top, pad_left, pad_bottom and pad_right have the same value.
    The default is 0.
  • depthwise_has_relu : Boolean, whether the layer applies a ReLU function after depthwise convolution.
  • pointwise_has_relu : Boolean, whether the layer applies a ReLU function after pointwise convolution.
  • depthwise_relu_max_value : Number
  • pointwise_relu_max_value : Number