Use the Keras set_random_seed in tests #30504
Conversation
| from transformers.pipelines import SUPPORTED_TASKS | ||
|
|
||
| set_seed_fn = lambda: tf.random.set_seed(0) # noqa: E731 | ||
| set_seed_fn = lambda: keras.utils.set_random_seed(0) # noqa: E731 |
There was a problem hiding this comment.
So tf.random.set_seed doesn't actually set the seed for the whole TF environment?
Does keras.utils.set_random_seed control just keras objects, or the whole TF env too?
There was a problem hiding this comment.
As far as I can tell, keras.utils.set_random_seed() basically setstf.random.set_seed() as before, but also sets the Python and Numpy random seeds. It's the currently recommended approach for reproducible initialization in the Keras docs.
There was a problem hiding this comment.
Hope there is no other tests having the same init issue but no failure by luck 🤞 .
(otherwise we might get different results for them after this PR 😅 )
There was a problem hiding this comment.
I'm not surprised it's the recommended one in the keras docs ;)
OK, well I just we just shrug and put it in the list of other TF quirks
There was a problem hiding this comment.
thanks @Rocketknight1
(is this a API changes causing the problem?)
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Some tests load checkpoints where weights are missing (e.g. the
poolerweights forViT). To get reproducible results, we need to ensure that seeds are set consistently. This PR useskeras.set_random_seed()instead oftf.set_seed(), which fixes some issues the CI was seeing where models mismatched because of differing weight inits.