Infinite Gradient Handling #582
Conversation
There was a problem hiding this comment.
Thanks for the PR!
This looks pretty good already. I have added a few review comments to your changes.
Additionally: Can you create a new test file tests/optimagic/optimization/test_invalid_function_value.py, and add two test cases where you check the cases of
- Infinite values and
- NaN values
These tests should fail without your changes, and pass with your changes; i.e., you will have to check that your code raises the correct error! We have many such tests in our test suite -- I would just search for "pytest.raises".
Thanks again and let me know if you have any questions!
| # Check for infinite values in the user-provided gradient | ||
| self._check_infinite_gradients(params, out_jac, jac_value) |
There was a problem hiding this comment.
The comment is not necessary. Should be self-explanatory given the methods name. Also applies to the instances below.
| params: PyTree, | ||
| out_jac: NDArray[np.float64], | ||
| jac_value: PyTree, |
There was a problem hiding this comment.
I think the order should be (out_jac, jac_value, params)
| """Check for infinite values in gradients and raise an error if found. | ||
|
|
||
| Args: | ||
| x: internal parameter vector at which the gradient was evaluated. |
There was a problem hiding this comment.
| x: internal parameter vector at which the gradient was evaluated. |
| params: user-facing parameter representation at evaluation point. | ||
| out_jac: internal processed gradient to check for infinities. | ||
| jac_value: original gradient value as returned by the user function, | ||
| included in error messages for debugging. |
| @@ -704,6 +717,33 @@ def _pure_evaluate_fun_and_jac( | |||
|
|
|||
| return (algo_fun_value, out_jac), hist_entry, log_entry | |||
|
|
|||
| def _check_infinite_gradients( | |||
There was a problem hiding this comment.
I prefer
def _assert_finite_jac(...)
...instead of the negative form.
I would also speak of jac and not gradient, because we are dealing with the jacobian (it is only the gradient in the special case of scalar-valued function).
| InvalidFunctionError: If any infinite values are found in the gradient. | ||
|
|
||
| """ | ||
| if np.any(np.isinf(out_jac)) or np.any(np.isnan(out_jac)): |
There was a problem hiding this comment.
Given the adjusted function name, I would also use
if not np.all(np.isfinite(out_jac))
...here. For this you should then adjust the docstring and add a comment in the docstring that np.isfinite checks for infinity and NaN values (see np.isfinite docstring).
Added support for infinite gradient check as per this
Note that the method added also checks for
nanvalues in grads along withinfinitevalues, since both lead to the same errors in optimization and are therefore, related.Let me know if any further changes/polishing is required. All the CI tests seem to pass.
Thanks