LookupTable followed by Conv layer causes error

[ylhsieh]

A small example like:

layers = [LookupTable(vocab_size=1000, embedding_dim=100),
          Conv((2, 2, 64), strides=1)]

would produce the following error message using gpu backend

File "/usr/local/lib/python2.7/dist-packages/neon/backends/convolution.py", line 758, in _magic32
    nc = ((nmax + 1) // d) * d - 1
ZeroDivisionError: integer division or modulo by zero

and using cpu backend would produce

File "/usr/local/lib/python2.7/dist-packages/neon/backends/nervanacpu.py", line 1343, in fprop_pool
    array_O = O._tensor.reshape(layer.dimO)
ValueError: can only specify one unknown dimension

Is there a proper way to shape outputs of LookupTable for use in the Conv layer? Thanks.

[yinyinl]

This layer combination is not supported in neon currently. Conv layer is not able to handle the out_shape from lookuptable layer, since lookuptable layer out_shape is a tuple with a time dimension as the 2nd element.
We are working on adding a reshape layer, which allows more flexible combination of various layer types, and hopefully release that soon. And we will make sure your example works as a test case. Thanks.

[jennifermyers]

Closing this issue as reshape layer is now available in neon 1.6.0.

[yinyinl]

Please take a look at the test for reshape layer for layer combination examples.