Adding: Numpy Backend #483
Conversation
…as-core into aritra-np-backend OK
|
Todos:
|
There was a problem hiding this comment.
Thanks for the PR!
I suppose if we're going to add Numpy as a backend we should also add CI for it -- this just means we'll need to add numpy to the backend list in actions.yaml
| def fit(self): | ||
| raise NotImplementedError("Trainer not implemented for NumPy backend.") | ||
|
|
||
| def predict(self): |
There was a problem hiding this comment.
Long term it could be possible to implement predict and evaluate.
| @@ -49,3 +49,11 @@ def __getattr__(self, name): | |||
| from keras_core.backend import torch as torch_backend | |||
|
|
|||
| return getattr(torch_backend, name) | |||
| if self._backend == "numpy": | |||
| # TODO (ariG23498): | |||
| # The import `from keras_core.backend import numpy as numpy_backend` | |||
There was a problem hiding this comment.
I am usure about this. It might have to do with numpy backend having a numpy file on its own (which somehow creates the error).
|
Please run |
|
I am usure why the JAX tests fail. |
|
Are you able to run them locally? They're timing out rather than failing. This is possibly normal. |
Yep! Tried running the test suite with JAX and it ran fine. |
|
Removed the debug statements! Sorry about that 😅 |
|
I think we should be good to merge. Thanks for the great contribution! |
* chore: adding numpy backend * creview comments * review comments * chore: adding math * chore: adding random module * chore: adding ranndom in init * review comments * chore: adding numpy and nn for numpy backend * chore: adding generic pool, max, and average pool * chore: adding the conv ops * chore: reformat code and using jax for conv and pool * chore: added self value * chore: activation tests pass * chore: adding post build method * chore: adding necessaity methods to the numpy trainer * chore: fixing utils test * chore: fixing losses test suite * chore: fix backend tests * chore: fixing initializers test * chore: fixing accuracy metrics test * chore: fixing ops test * chore: review comments * chore: init with image and fixing random tests * chore: skipping random seed set for numpy backend * chore: adding single resize image method * chore: skipping tests for applications and layers * chore: skipping tests for models * chore: skipping testsor saving * chore: skipping tests for trainers * chore:ixing one hot * chore: fixing vmap in numpy and metrics test * chore: adding a wrapper to numpy sum, started fixing layer tests * fix: is_tensor now accepts numpy scalars * chore: adding draw seed * fix: warn message for numpy masking * fix: checking whether kernel are tensors * chore: adding rnn * chore: adding dynamic backend for numpy * fix: axis cannot be None for normalize * chore: adding jax resize for numpy image * chore: adding rnn implementation in numpy * chore: using pytest fixtures * change: numpy import string * chore: review comments * chore: adding numpy to backend list of github actions * chore: remove debug print statements
In this PR I will continue working on the Numpy Backend for Keras Core