mx.nd.contrib.BilinearResize2D (Inside HybridBlock)

Gluon
1

I’m trying to use BilinearResize2D, here is my code sample:

class DecoderBlock(nn.HybridBlock):
    def __init__(self, out_channels, in_channels):
        super(DecoderBlock, self).__init__()

        with self.name_scope():
            self.de_relu = nn.Activation(activation='relu')
            self.de_1_by_b_conv = nn.Conv2D(channels=out_channels, kernel_size=1, strides=1, padding=0)
            self.de_norm = nn.BatchNorm(momentum=0.1, in_channels=out_channels)

    def hybrid_forward(self, F, x):
        out = self.de_relu(x)
        out = self.de_1_by_b_conv(out)
        out = mx.nd.contrib.BilinearResize2D(out,
                                             out_height=x.shape[2] * 2,
                                             out_width=x.shape[3] * 2)
        out = self.de_norm(out)
        return out

I get the following error:

  File "/home/ubuntu/anaconda3/envs/mxnet_p27/lib/python2.7/site-packages/mxnet/gluon/parameter.py", line 297, in _init_impl
    self._data = [data.copyto(ctx) for ctx in self._ctx_list]
  File "/home/ubuntu/anaconda3/envs/mxnet_p27/lib/python2.7/site-packages/mxnet/ndarray/ndarray.py", line 2053, in copyto
    return _internal._copyto(self, out=hret)
  File "<string>", line 25, in _copyto
  File "/home/ubuntu/anaconda3/envs/mxnet_p27/lib/python2.7/site-packages/mxnet/_ctypes/ndarray.py", line 92, in _imperative_invoke
    ctypes.byref(out_stypes)))
  File "/home/ubuntu/anaconda3/envs/mxnet_p27/lib/python2.7/site-packages/mxnet/base.py", line 235, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [20:48:10] src/ndarray/ndarray.cc:1233: GPU is not enabled
2

Given the snippet, it’s not easy to detect what went wrong, but there’re couple of points that might lead to solving it;

  1. in hybrid blocks the use of F in hybrid_forward is for symbolic generation, so remove mx.nd and substitute it with F
  2. if you insist on using nd then remove the Hybrid from nn.HybridBlock and make it a simple nn.Block and implement forward

Hope that solves your problem

3

Hi Oron,

You can try the following code and let me know if it is working for you. I have checked in in cpu and gpu with hybridization enabled and disabled on 1.2.1.

Please, notice, that I:

  • Had to pass dimensions into the block as parameters, so hybridization would work
  • As @ehsanmok mentioned before, you need to use F.contrib for hybridization to work
  • in_channels argument of the block is actually not used (as it was in your example)
import mxnet as mx
from mxnet.gluon import nn


class DecoderBlock(nn.HybridBlock):
    def __init__(self, height, width, out_channels, in_channels, **kwargs):
        super(DecoderBlock, self).__init__(**kwargs)

        self._height = height
        self._width = width

        with self.name_scope():
            self.de_relu = nn.Activation(activation='relu')
            self.de_1_by_b_conv = nn.Conv2D(channels=out_channels, kernel_size=1, strides=1, padding=0)
            self.de_norm = nn.BatchNorm(momentum=0.1, in_channels=out_channels)

    def hybrid_forward(self, F, x, *args, **kwargs):
        out = self.de_relu(x)
        out = self.de_1_by_b_conv(out)
        out = F.contrib.BilinearResize2D(out, height=self._height * 2, width=self._width * 2)
        out = self.de_norm(out)
        return out


ctx = mx.cpu()
# batch size x number of filters x height x width
x = mx.random.uniform(shape=(5, 50, 400, 400), ctx=ctx)
block = DecoderBlock(400, 400, 3, 3)
block.initialize(ctx=ctx)
block.hybridize()
y = block(x)
print(y)
4

For mxnet-cu90mkl 1.1.0 it does not work.

I get the following error:
AttributeError: 'module' object has no attribute 'BilinearResize2D'

It works for me on mxnet-cu90mkl 1.2.1

thanks!