Why the bcast is needed in create_param()?

[david-leon]

In create_param(), these two lines confuse me:

        bcast = tuple(s == 1 for s in shape)
        spec = theano.shared(spec, broadcastable=bcast)

I'm wondering what the bcast is for, especially these two lines will cause theano exception TypeError: Cannot convert Type TensorType(float32, vector) (of Variable Elemwise{add,no_inplace}.0) into Type TensorType(float32, (True,)) when shape arg is set to [1]

[nouiz]

It is there to work around the fact that Theano broadcastable isn't perfect. In fact, in some case, it is not possible to do it without user help, but in other case, Theano could be improved too, but those case are complicated.

…

On Thu, Aug 2, 2018 at 2:21 AM David Leon ***@***.***> wrote: In create_param() <https://github.com/Lasagne/Lasagne/blob/a61b76fd991f84c50acdb7bea02118899b5fefe1/lasagne/utils.py#L392>, these two lines confuse me: bcast = tuple(s == 1 for s in shape) spec = theano.shared(spec, broadcastable=bcast) I'm wondering what the bcast is for, especially these two lines will cause theano exception TypeError: Cannot convert Type TensorType(float32, vector) (of Variable Elemwise{add,no_inplace}.0) into Type TensorType(float32, (True,)) when shape arg is set to [1] — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#912>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AALC-yjdjrg4-WumTsZWzoF6xVfASuOCks5uMpptgaJpZM4VrspB> .

[f0k]

Sorry for the late reply. Lasagne has partial information on the input shapes and full information on the parameter shapes, and it uses this to set the correct broadcast patterns on Theano variables (which do not have any shape information at all). This allows some expressions using the outputs of Lasagne networks to work out of the box that would fail otherwise (see #715 for why this was introduced in the first place). On the downside, it requires extra care about broadcast patterns in user-provided variables.
Closing this issue since it was only a question, but please feel free to post how you got this error message and we can look into improving the documentation or failing earlier with a more useful message (and double-check it's not a bug in Lasagne).