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 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 grad(self, inp, grads):
dy, sm = inp
g, = grads
tmp = g + tensor.neg(tensor.sum(g * sm, axis=1).dimshuffle((0, 'x')))
g_dy = tmp * sm
tmp2 = tensor.sum(dy * sm, axis=1).dimshuffle((0, 'x'))
g_sm = tmp * dy - g * tmp2
return g_dy, g_sm
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)