Closed carlosgmartin closed 2 months ago
@vroulet Here's the output I get from running tests on my machine (after applying https://github.com/google-deepmind/optax/pull/1068 and https://github.com/google-deepmind/optax/pull/1071):
Do you get the same output from running tests on your machine?
Indeed, I get the error too. I don't understand why it's not been caught on the tests in github. Anyway, I'd prefer using small relative accuracies rather than slightly loose absolute differences (the functions that these tests verify are supposed to be use for high precision purposes). So I'd prefer to use for the lbfgs test:
chex.assert_trees_all_close(
plain_precond_vec, expected_precond_vec, rtol=1e-5
)
and for the linesearch test:
chex.assert_trees_all_close(scipy_res[0], stepsize, rtol=1e-5)
chex.assert_trees_all_close(scipy_res[3], final_value, rtol=1e-5)
I checked, the tests pass with these changes.
The doctest of _dog also fails on my end. As you are on it, could you also use ellipses on the doctest of dog, i.e., here use
Objective function: 13.99...
Objective function: 13.99...
Objective function: 13.99...
Objective function: 13.99...
Objective function: 13.99...
(This algorithm needs to be completed with a proper evaluation function and fuse with a similar one in the alias/transform files.)
Thanks @carlosgmartin ! I don't see when or where these tests broke. Could you point out the reason for this change?