Allow old-style shape in blobproto_to_array

[lukeyeager]

Fixes #3199
Bug introduced in #3170

[jeffdonahue]

Thanks @lukeyeager, sorry for neglecting to think about this in #3170. I think the logic should be reversed though -- i.e., should check if any of the legacy fields are present and use them if so, othereise use the shape. This way, a BlobProto with neither the legacy fields nor a shape will be interpreted as having an empty shape (a scalar). This is consistent with the behavior of Blob::FromProto.

[lukeyeager]

I've updated the logic and added a third test.

[jeffdonahue]

This wasn't quite what I meant -- this will give a 1D array, not a 0D array (scalar). You'd still want to return data.reshape(blob.shape.dim), or equivalently return data.reshape(()). The invariant we want is len(data.shape) == len(blob.shape.dim) (unless any of the legacy fields are present). It will only work in the case where len(data) == 1, but that's okay since a BlobProto with no shape field (and no legacy dims) is kind of an odd corner case.

[lukeyeager]

Well I'm not letting it fall back to this worthless error:

  File "/home/tranlaman/BLVC-caffe/python/caffe/io.py", line 26, in blobproto_to_array
    return np.array(blob.data).reshape(*blob.shape.dim)
TypeError: function takes exactly 1 argument (0 given)

How about raising a ValueError?

          if len(data) != 1:
              raise ValueError('blob has no shape')
          # Return scalar
          return data[0]

[jeffdonahue]

If you remove the * (reshape(blob.shape.dim)) it should work? ValueError seems okay to me too, though.

[lukeyeager]

Oh yeah, that works. Good call.

[lukeyeager]

Updated logic again and added a fourth test.

[jeffdonahue]

Thanks for the fixes and tests @lukeyeager! LGTM.