def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = opt.scalarconsts_rest(
inv_arg.owner.inputs, only_process_constants=True)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and np.allclose(np.sum(scalars), 1):
out = opt._fill_chain(
sigmoid(tensor.neg(nonconsts[0].owner.inputs[0])),
scalar_inputs,
)
# keep combined stack traces of
# exp(x): nonconsts[0],
# 1 + exp(x): inv_arg,
# 1 / (1 + exp(x)): node.outputs[0]
copy_stack_trace(
[nonconsts[0], inv_arg, node.outputs[0]], out)
return out
def is_1pexp(t):
"""
If 't' is of the form (1+exp(x)), return (False, x).
Else return None.
"""
if t.owner and t.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(t.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
maybe_exp = nonconsts[0]
if maybe_exp.owner and maybe_exp.owner.op == tensor.exp:
# Verify that the constant terms sum to 1.
if scalars:
scal_sum = scalars[0]
for s in scalars[1:]:
scal_sum = scal_sum + s
if numpy.allclose(scal_sum, 1):
return False, maybe_exp.owner.inputs[0]
# Before 7987b51 there used to be a bug where *any* constant
# was considered as if it was equal to 1, and thus this
# function would incorrectly identify it as (1 + exp(x)).
if config.warn.identify_1pexp_bug:
warnings.warn(
'Although your current code is fine, please note that '
'Theano versions prior to 0.5 (more specifically, '
'prior to commit 7987b51 on 2011-12-18) may have '
'yielded an incorrect result. To remove this warning, '
'either set the `warn.identify_1pexp_bug` config '
'option to False, or `warn.ignore_bug_before` to at '
'least \'0.4.1\'.')
return None
def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(inv_arg.owner.inputs, only_process_constants=True)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and numpy.allclose(numpy.sum(scalars), 1):
out = opt._fill_chain(
sigmoid(
tensor.neg(nonconsts[0].owner.inputs[0])),
scalar_inputs)
# keep combined stack traces of
# exp(x): nonconsts[0],
# 1 + exp(x): inv_arg,
# 1 / (1 + exp(x)): node.outputs[0]
copy_stack_trace(
[nonconsts[0], inv_arg, node.outputs[0]], out)
return out
def is_1pexp(t):
# if t is of form (1+exp(x)), return x
# else return None
if t.owner and t.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(t.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
maybe_exp = nonconsts[0]
if maybe_exp.owner and maybe_exp.owner.op == tensor.exp:
return False, maybe_exp.owner.inputs[0]
return None
def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = opt.scalarconsts_rest(inv_arg.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and numpy.allclose(numpy.sum(scalars), 1):
return opt._fill_chain(sigmoid(tensor.neg(nonconsts[0].owner.inputs[0])), scalar_inputs)
def local_inv_1_plus_exp(node):
"""
1/(1+exp(x)) -> sigm(-x)
"""
# this optimization should be done for numerical stability
# so we don't care to check client counts
if node.op == tensor.inv:
inv_arg = node.inputs[0]
if inv_arg.owner and inv_arg.owner.op == tensor.add:
scalars, scalar_inputs, nonconsts = \
opt.scalarconsts_rest(inv_arg.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars
if len(nonconsts) == 1:
if nonconsts[0].owner and nonconsts[0].owner.op == tensor.exp:
if scalars and numpy.allclose(numpy.sum(scalars), 1):
return opt._fill_chain(
sigmoid(tensor.neg(nonconsts[0].owner.inputs[0])),
scalar_inputs)
