Description
This issue is to track the progression. Help is welcome, we have a lot of work to do but it's fairly easy. You can take example on the pull requests made before.
The idea is that before, using the @run_all_in_graph_and_eager_mode meant that we needed to make the tests with the tensorflow 1.x style, with things like tf.compat.v1.initialize_global_variables, self.evaluateand other scary functions like that.
Now we use the @pytest.mark.usefixtures("maybe_run_functions_eagerly") which will run the test function twice. Once normally, and once with tf.config.experimental_run_functions_eagerly(True). Which means the tests are eager-first. No need to initialize variables, get the result with .numpy(). To test the values, use the numpy testing module instead of the methods of tf.test.TestCase like self.AssertAllClose and such.
You can take as reference the two pull request already made:
#1288
#1327
When doing a pull request, please do not do more than 2 tests at once.
The policy is that when we're testing a simple function (for example, using a custom op), no need to use tf.functions and no need to use @pytest.mark.usefixtures("maybe_run_functions_eagerly") because we're not afraid of some python side effects.
When working with complex functions (for loops, if/else with tensors...) we need to add tf.function and to add the @pytest.mark.usefixtures("maybe_run_functions_eagerly") to make sure we don't rely on some pythonic behavior. To avoid having to make complex input_signature in tf.function we can isolate the sensitive part (if/else/for loop with tensors) in a separate tf.function and not decorate the main one.