How to implement BilinearResize1D?

Gluon
1

I’m porting the Wave-U-Net from tensorflow to mxnet and I need to upsample the 3D audio features with bilinear interpolation. However, there’s no bilinear upsampling operator in mxnet for 1D audio data ( 3d feature map of shape [batch_size, channels, length]).

First, I use mx.nd.contrib.BilinearResize2D by reshaping the feature from [B, C, T] to [B, C, T, 1]. But the question is, the width and height of BilinearResize2D are in int16_t and my T is larger than 65535. Also the mx.init.Bilinear does not support weight less than 3 dimensions.

How to implement BilinearResize1D in mxnet gluon?

2

You can use mxnet.ndarray.UpSampling and set the type to bilinear https://beta.mxnet.io/api/ndarray/_autogen/mxnet.ndarray.UpSampling.html

3

Hi. Unfortunately, mxnet.ndarray.UpSampling does not support 3D tensor

MXNetError: [16:53:22] src/operator/nn/upsampling.cc:61: Check failed: dshape.ndim() == 4U (3 vs. 4) : UpSamplingBilinear: Input data should be 4D in (batch, channel, y, x)

Besides, there’s no doc indicating how to initialize the weight.