NewModuleTest: Don't call both check_jacobian and gradcheck (#49566)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/49566
Fixes #49422.
check_jacobian and gradcheck do roughly the same thing: they both
compute an analytic jacobian and a numeric jacobian and check that
they are equivalent. Furthermore, NewModuleTest will (by default) call
both check_jacobian and gradcheck, leading to some redundant checks that
waste CI resources.
However, there is one subtle difference: `check_jacobian` can handle the
special case where a Module takes in dense inputs and dense parameters
but returns sparse gradients, but that is not something gradcheck can
handle. This is only used in the tests for nn.Embedding and
nn.EmbeddingBag.
This PR does the following:
- have NewModuleTest call gradcheck instead of check_jacobian by default
- add a new "has_sparse_gradients" flag to NewModuleTest. These are True
for the nn.Embedding and nn.EmbeddingBag sparse gradient tests. If
`has_sparse_gradients` is True, then we call check_jacobian, otherwise,
we call gradcheck.
- Kills the "jacobian_input" flag. This flag was used to tell
NewModuleTest to not attempt to compute the jacobian for the inputs to
the module. This is only desireable if the input to the module isn't
differentiable and was only set in the case of nn.Embedding /
nn.EmbeddingBag that take a LongTensor input. `gradcheck` handles these
automatically by not checking gradients for non-differentiable inputs.
Test Plan:
- Code reading
- run test_nn.py
Reviewed By: albanD
Differential Revision: D25622929
Pulled By: zou3519
fbshipit-source-id: 8d831ada98b6a95d63f087ea9bce1b574c996a22
