Nnls solvers fail when weights=True

[tcstewar]

The non-negative solvers don't seem to work on full weight matrices. Here's a quick example:

model = nengo.Network()
with model:
    a = nengo.Ensemble(100, 1)
    b = nengo.Ensemble(100, 1)
    nengo.Connection(a, b, solver=nengo.solvers.Nnls(weights=True))
sim = nengo.Simulator(model)

The error is:

---------------------------------------------------------------------------
BuildError                                Traceback (most recent call last)
<ipython-input-22-b1fdc30f942d> in <module>()
----> 1 sim = nengo.Simulator(model)
      2 sim.run(2)

/home/tcstewar/github/nengo/nengo/simulator.pyc in __init__(self, network, dt, seed, model)
    115         if network is not None:
    116             # Build the network into the model
--> 117             self.model.build(network)
    118 
    119         self.model.decoder_cache.shrink()

/home/tcstewar/github/nengo/nengo/builder/builder.pyc in build(self, obj, *args, **kwargs)
     83 
     84     def build(self, obj, *args, **kwargs):
---> 85         return Builder.build(self, obj, *args, **kwargs)
     86 
     87     def add_op(self, op):

/home/tcstewar/github/nengo/nengo/builder/builder.pyc in build(cls, model, obj, *args, **kwargs)
    128                 "Cannot build object of type %r" % obj.__class__.__name__)
    129 
--> 130         return cls.builders[obj_cls](model, obj, *args, **kwargs)

/home/tcstewar/github/nengo/nengo/builder/network.pyc in build_network(model, network)
     68         # TODO: Except perhaps if the connection being learned
     69         # is in a subnetwork?
---> 70         model.build(conn)
     71 
     72     logger.debug("Network step 4: Building probes")

/home/tcstewar/github/nengo/nengo/builder/builder.pyc in build(self, obj, *args, **kwargs)
     83 
     84     def build(self, obj, *args, **kwargs):
---> 85         return Builder.build(self, obj, *args, **kwargs)
     86 
     87     def add_op(self, op):

/home/tcstewar/github/nengo/nengo/builder/builder.pyc in build(cls, model, obj, *args, **kwargs)
    128                 "Cannot build object of type %r" % obj.__class__.__name__)
    129 
--> 130         return cls.builders[obj_cls](model, obj, *args, **kwargs)

/home/tcstewar/github/nengo/nengo/builder/connection.pyc in build_connection(model, conn)
    206                     in_signal,
    207                     signal,
--> 208                     tag="%s.weights_elementwiseinc" % conn))
    209 
    210     # Add operator for filtering

/home/tcstewar/github/nengo/nengo/builder/builder.pyc in add_op(self, op)
     90         signals = SignalDict()
     91         op.init_signals(signals)
---> 92         op.make_step(signals, self.dt, np.random)
     93 
     94     def has_built(self, obj):

/home/tcstewar/github/nengo/nengo/builder/operator.pyc in make_step(self, signals, dt, rng)
    386         A = signals[self.A]
    387         Y = signals[self.Y]
--> 388         reshape = reshape_dot(A, X, Y, self.tag)
    389 
    390         def step_dotinc():

/home/tcstewar/github/nengo/nengo/builder/operator.pyc in reshape_dot(A, X, Y, tag)
    350     if (badshape or incshape != Y.shape) and incshape != ():
    351         raise BuildError("shape mismatch in %s: %s x %s -> %s"
--> 352                          % (tag, A.shape, X.shape, Y.shape))
    353 
    354     # Reshape to handle case when np.dot(A, X) and Y are both scalars

BuildError: shape mismatch in <Connection from <Ensemble (unlabeled) at 0x7fda44dc74d0> to <Ensemble (unlabeled) at 0x7fda44edd610>>.weights_elementwiseinc: (1, 100) x (100,) -> (100,)

If I try it with the standard solvers, it runs fine.