added erf op to math.py #908
Conversation
keras_core/backend/torch/math.py
Outdated
|
|
||
| def erf(x): | ||
| if not isinstance(x, torch.Tensor): | ||
| x = torch.tensor(x) |
There was a problem hiding this comment.
You can just do x = convert_to_tensor(x) unconditionally
keras_core/ops/math.py
Outdated
| """Computes the error function of x element-wise. | ||
|
|
||
| Args: | ||
| input_tensor: A tensor of type `float32` or `float64`. |
There was a problem hiding this comment.
It can have more types, no?
There was a problem hiding this comment.
Edited the comments based on that
keras_core/ops/math.py
Outdated
|
|
||
| @keras_core_export("keras_core.ops.erf") | ||
| def erf(x): | ||
| """Functional interface to the `Erf` operation.""" |
There was a problem hiding this comment.
This is where the docstring should be, not the op above, since this is the public symbol.
keras_core/ops/math.py
Outdated
| >>> y_large = Erf()(x_large) | ||
| """ | ||
|
|
||
| def __init__(self): |
keras_core/ops/math.py
Outdated
| def compute_output_spec(self, input_tensor): | ||
| return KerasTensor(shape=input_tensor.shape, dtype=input_tensor.dtype) | ||
|
|
||
| def call(self, input_tensor): |
keras_core/ops/math.py
Outdated
| return KerasTensor(shape=x.shape, dtype=x.dtype) | ||
|
|
||
| def call(self, x): | ||
| return backend.erf(x) |
There was a problem hiding this comment.
You need to call backend.math. Tests are failing.
keras_core/ops/math.py
Outdated
|
|
||
| Examples: | ||
|
|
||
| # Basic usage |
There was a problem hiding this comment.
Since you're not printing any outputs, just use a fenced code block for the code example.
|
Hi @fchollet , Thanks so much for the guidance. The Tests are failing because the errors are greater than the tolerance level (1×10^-5) set as you can see below.
Kindly guide me on how to resolve this issue. Thanks & Regards |
|
You can just lower the precision of this particular check to 1e-4. |
lowered the precision tolerance to 1e-4 for erf function
|
Hi @fchollet , Thanks for the reply, Lowered the precision to 1e-4 such that it would resolve the issue but the maximum difference is around 0.12 which is way higher than 0.0001 which still leads to failing tests. Kindly advise. |
keras_core/ops/math_test.py
Outdated
|
|
||
| def test_erf_operation_edge_cases(self): | ||
| # Test for edge cases | ||
| edge_values = np.array([1e10, -1e10, 1e-10, -1e-10], dtype=np.float64) |
There was a problem hiding this comment.
Your test values are too large. Try 1e5. This the source of the large discrepancy IMO.
There was a problem hiding this comment.
I have implemented the changes, but I can see from the tests that it is failing for the below array examples. I wonder if there is anything wrong in the implementation function itself.
- x: array([ 1.128379e+00, -1.128379e+00, 1.273240e-05, -1.273240e-05])
- x: array([-1.128354, -1.123101, -0.950886, 0. , 0.950886, 1.123101, 1.128354])
lowered the test values to 1e5
|
Keras Core is becoming Keras 3, and we're switching development to the main repository! Please reopen this PR in the keras-team/keras repository. Unfortunately we aren't able to automatically transfer PRs (but we have transferred all issues). |
Hi @fchollet ,
Hope you are doing well. As discussed in issue keras-team/keras#18442, I have raised this PR to work on the ERF function. Kindly review the changes I have currently made. Once approved, I will go ahead with the implementation for JAX, torch, and numpy backends.
Thanks & Regards