def _numerical_jacobian(self,
module,
input,
jacobian_input=True,
jacobian_parameters=True):
output = self._forward(module, input)
output_size = output.nelement()
if jacobian_parameters:
param, d_param = self._get_parameters(module)
def fw(input):
out = self._forward(module, input)
if isinstance(out, Variable):
return out.data
return out
res = tuple()
# TODO: enable non-contig tests
input = contiguous(input)
if jacobian_input:
res += get_numerical_jacobian(fw, input, input),
if jacobian_parameters:
res += torch.cat(
list(get_numerical_jacobian(fw, input, p) for p in param), 0),
return res
def do_test(self,
cls=cls,
constructor_args=constructor_args,
call_args=call_args,
test_name=test_name):
input = create_input(call_args)
output = cls(*constructor_args)(*input)
if not isinstance(output, tuple):
output = (output, )
for i, o in enumerate(output):
if not o.requires_grad:
continue
analytical = get_analytical_jacobian(input, o)
def fn(input):
tmp = cls(*constructor_args)(*input)
if not isinstance(tmp, tuple):
tmp = (tmp, )
return tmp[i].data
numerical = get_numerical_jacobian(fn, input, input)
self.assertLessEqual(
max(
a.add(-1, n).abs().max()
for a, n in zip(analytical, numerical)), PRECISION)
if test_name not in ignore_inplace and issubclass(
cls, InplaceFunction):
inplace_input = deepcopy(input)
inplace_input_copy = tuple(i + 0 for i in inplace_input)
fn = cls(*constructor_args, inplace=True)
inplace_output = fn(*inplace_input_copy)
if not isinstance(inplace_output, tuple):
inplace_output = (inplace_output, )
self.assertEqual(inplace_output, output)
# Check that gradient is the same
for inp_i, i in zip(inplace_input, input):
if inp_i.grad is not None:
inp_i.grad.data.zero_()
if i.grad is not None:
i.grad.data.zero_()
for io, o in zip(inplace_output, output):
grad = torch.randn(*io.size()).double()
io.backward(grad)
o.backward(grad)
for inp_i, i in zip(inplace_input, input):
self.assertEqual(inp_i.grad, i.grad)
