Autograd not working for torch.exp(1j * phase)

[jonashaag]

🐛 Bug

Complex tensor construction from magnitude and phase does not seem to support autograd when using mag * torch.exp(1j * phase) notation:

import torch
mag, phase = torch.tensor(5., requires_grad=True), torch.tensor(3., requires_grad=True)

complex_good = torch.view_as_complex(torch.stack([mag * torch.cos(phase), mag * torch.sin(phase)], dim=-1))
complex_good.backward()  # works

complex_bad = mag * torch.exp(1j * phase)
complex_bad.backward()

=>

.../torch/autograd/__init__.py:125: UserWarning: Complex backward is not fully supported yet and could lead to wrong gradients for functions we have not fixed yet (Triggered internally at  /opt/conda/conda-bld/pytorch_1597820903894/work/torch/csrc/autograd/python_engine.cpp:172.)
  Variable._execution_engine.run_backward(
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../torch/tensor.py", line 214, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File ".../torch/autograd/__init__.py", line 125, in backward
    Variable._execution_engine.run_backward(
RuntimeError: Expected isFloatingType(grad.scalar_type()) || (input_is_complex == grad_is_complex) to be true, but got false.  (Could this error message be improved?  If so, please report an enhancement request to PyTorch.)
...

Torch version: 1.7.0.dev20200819

cc @ezyang @gchanan @zou3519 @ssnl @albanD @gqchen @anjali411 @dylanbespalko

[albanD]

As mentioned in the warning there, complex autograd is not fully supported right now.

My guess here is that the formula for one of these ops has not been updated yet and does not produce a complex gradient like it should.

@anjali411 you might want to open a single umbrella issue that tracks these as I expect there will be more...

[anjali411]

@jonashaag thanks for reporting the issue! It's because the MulBackward's behavior for complex is problematic. According to the current formula, when a real tensor is multiplied by a complex scalar, a complex valued gradient is returned instead of the expected real valued gradient.

>>> import torch
x=t>>> x=torch.randn(4, dtype=torch.cdouble)
>>> x=torch.randn(4, dtype=torch.cdouble, requires_grad=True)
>>> x=torch.randn(4, requires_grad=True)
>>> y=x*1j
>>> y
tensor([-0.-1.1449j, 0.+0.5593j, 0.+1.7524j, 0.+0.1101j],
       grad_fn=<MulBackward0>)
>>> y.sum().backward()
/home/chourdiaanjali/pytorch2/torch/autograd/__init__.py:127: UserWarning: Complex backward is not fully supported yet and could lead to wrong gradients for functions we have not fixed yet (Triggered internally at  ../torch/csrc/autograd/python_engine.cpp:172.)
  allow_unreachable=True)  # allow_unreachable flag
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/chourdiaanjali/pytorch2/torch/tensor.py", line 214, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/chourdiaanjali/pytorch2/torch/autograd/__init__.py", line 127, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: Expected isFloatingType(grad.scalar_type()) || (input_is_complex == grad_is_complex) to be true, but got false.  (Could this error message be improved?  If so, please report an enhancement request to PyTorch.)

As Alban mentioned, complex backward is not fully supported yet but many of these functions will be fixed soon.

@anjali411 you might want to open a single umbrella issue that tracks these as I expect there will be more...

I have been adding these issues in the Autograd Tasks subsection in #33152

[cshewmake2]

Having the same issue as above. How's this going? Any way I can help?

[mountain]

same issue +1

[IvanYashchuk]

This was fixed in #43208.

[rjkilpatrick]

This is still an issue on nightly

>>> import torch
>>> print(torch.__version__)
>>> mag = torch.tensor(5., requires_grad=True, dtype=torch.complex128)
>>> phase = torch.tensor(3., requires_grad=True, dtype=torch.complex128)
>>>
>>> complex_good = mag * (torch.cos(phase) + 1.j * torch.sin(phase))
>>> complex_good.backward() # works
>>>
>>> complex_bad = mag * torch.exp(1j * phase)
>>> complex_bad.backward()

1.8.0.dev20201101
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-20-98d908b62931> in <module>
      8 complex_good.backward()  # works
      9 
---> 10 complex_bad = mag * torch.exp(1j * phase)
     11 complex_bad.backward()

RuntimeError: exp does not support automatic differentiation for outputs with complex dtype.

pytorch/test/test_autograd.py

Lines 4935 to 4941 in 1cc1da5

	complex_list = ['t', 'view', 'reshape', 'reshape_as', 'view_as', 'roll', 'clone',
	'repeat', 'expand', 'flip', 'fliplr', 'flipud', 'rot90', 'transpose',
	'permute', 'squeeze', 'unsqueeze', 'resize', 'resize_as', 'tril', 'triu',
	'chunk', 'split', 'split_with_sizes', 'repeat', 'expand', 'zero_',
	'eq_', 'ne_', 'add', '__radd__', 'sum', 'conj', 'sin', 'cos', 'mul', 'sinh',
	'cosh', '__rmul__', 'sgn', 'abs', 'dot', 'vdot', 'tensor_split', 'matmul',
	'bmm', 'mv', 'ger', 'diagonal', 'atan', 'angle', 'tanh', 'fill_', 'sub'] + separate_complex_tests

torch.exp is not being tested for autograd backwards on master