Custom interpolation layer. Gluon.

[ashkanaev]

How to realize custom layer (bilinear interpolation) for gluon without input tensor shape?
The issue is hardlink to target size. I want to realize layer for resize arbitrary tensor with some factor.
I found, that SpatialTransformer is what i really need, but target size makes this code inflexible.

Could i realize custom symbol layer with symbol layer?

class Inter(Block):
    def __init__(self, factor, target_size, **kwargs):
        super(Inter, self).__init__(**kwargs)

        self._affine = mx.nd.reshape(mx.nd.array([[1. * factor, 0, 0],
                                 [0, 1. * factor, 0]]), shape=(1, 6))

        self.target_size = target_size



    def forward(self, x):
        out = mx.nd.SpatialTransformer(x,
                                   self.affine_matrix,
                                   self.target_size,
                                   'affine',
                                   'bilinear'
                                   )

[zhanghang1989]

I guess this is what you are looking for #9688

import mxnet as mx
x1 = mx.nd.ones(shape=(2,3,4,4))
y1 = mx.nd.contrib.BilinearResize2D(x1, out_height=5, out_width=5)

@szha

[zhanghang1989]

class BilinearInterp(Block):
    def __init__(self, factor, target_height, target_width, **kwargs):
        super(BilinearInterp, self).__init__(**kwargs)
        self.height = target_height
        self.width = target_height
    def forward(self, x):
        return mx.nd.contrib.BilinearResize2D(x, self.height, self.width)