def get_or_compute_grads(loss_or_grads, params):
"""Helper function returning a list of gradients
Parameters
----------
loss_or_grads : symbolic expression or list of expressions
A scalar loss expression, or a list of gradient expressions
params : list of shared variables
The variables to return the gradients for
Returns
-------
list of expressions
If `loss_or_grads` is a list, it is assumed to be a list of
gradients and returned as is, unless it does not match the length
of `params`, in which case a `ValueError` is raised.
Otherwise, `loss_or_grads` is assumed to be a cost expression and
the function returns `theano.grad(loss_or_grads, params)`.
Raises
------
ValueError
If `loss_or_grads` is a list of a different length than `params`, or if
any element of `params` is not a shared variable (while we could still
compute its gradient, we can never update it and want to fail early).
"""
if any(not theano.compat.is_shared(p) for p in params):
raise ValueError(
"params must contain shared variables only. If it "
"contains arbitrary parameter expressions, then "
"lasagne.utils.collect_shared_vars() may help you."
)
if isinstance(loss_or_grads, list):
if not len(loss_or_grads) == len(params):
raise ValueError("Got %d gradient expressions for %d parameters" % (len(loss_or_grads), len(params)))
return loss_or_grads
else:
return theano.grad(loss_or_grads, params)
def get_output_for(self, inputs, **kwargs):
input, layer_out, layer_in = inputs
return theano.grad(None, wrt=layer_in, known_grads={layer_out: input})
