Skip to content



This layer changes the shape of input without changing the contents.


  • shape : Integers, target shape.
    It supports Caffe style parameters, e.g. [0, 16, -1] means [128, 16, 16] when the input is [128, 256].
    The default is [0, -1]; i.e. flatten without changing batch.
  • origin : String, original channel-position when importing from other deep-learnig frameworks.
    'chl' for channel last or 'chf' for channel first.
    The default is 'chl'.