LSTM doesn't work with Symbols - "'Symbol' object has no attribute 'shape'"

[aodhan-domhnaill]

I am using the following versions,

$ pip list | grep mxnet
mxnet (1.0.0.post1)
$ python --version
Python 3.6.3

When I run,

import mxnet as mx
from mxnet import gluon

class Net(gluon.Block):
    def __init__(self, recurrent_width, forward_width, **kwargs):
        super(Net, self).__init__(**kwargs)
        with self.name_scope():
            self._rnn = gluon.rnn.LSTM(recurrent_width)
            self._nn  = gluon.nn.Dense(forward_width)

    def forward(self, x):
        return self._nn(self._rnn(x))

net = Net(10, 10)
net(mx.sym.var('data'))

I get,

$ python shape_error.py 
Traceback (most recent call last):
  File "shape_error.py", line 15, in <module>
    net(mx.sym.var('data'))
  File "/usr/local/lib/python3.6/site-packages/mxnet/gluon/block.py", line 304, in __call__
    return self.forward(*args)
  File "shape_error.py", line 12, in forward
    return self._nn(self._rnn(x))
  File "/usr/local/lib/python3.6/site-packages/mxnet/gluon/block.py", line 304, in __call__
    return self.forward(*args)
  File "/usr/local/lib/python3.6/site-packages/mxnet/gluon/rnn/rnn_layer.py", line 173, in forward
    batch_size = inputs.shape[self._layout.find('N')]
AttributeError: 'Symbol' object has no attribute 'shape'

So the problem is occuring here because RNN Layer checks the shape right away.

How can I fix this?

[eric-haibin-lin]

@szha

[szha]

Block doesn't support Symbol.

[aodhan-domhnaill]

@szha Is there a way to fix this? I am trying to use SageMaker, and the examples do this save call. For example, MNIST saves the symbolic graph.

[aodhan-domhnaill]

Never mind. I just needed to use HybridSequential.