Skip to content

[Bug]: ivy.vector_norm and ivy.lgamma when backend=numpy return rounded values #28844

Open

Description

@34j

Bug Explanation

> python -c "import ivy; ivy.set_backend('numpy'); print(ivy.lgamma(5)); print(ivy.lgamma(5.))"
ivy.array(3)
ivy.array(3.1780539)
> python -c "import ivy; ivy.set_backend('numpy'); print(ivy.vector_norm([1, 1])); print(ivy.vector_norm([1., 1.]))"
ivy.array(1)
ivy.array(1.4142135)

https://data-apis.org/array-api/latest/extensions/generated/array_api.linalg.vector_norm.html

According to array-api, there is no specified behavior when a non-floating point type is given for vector_norm, but I would like to suggest, with all due respect, to stop this casting or to output an error, for the following reasons. (For lgamma, not even defined in array-api.)

  • There are very limited situations where such behavior is needed.
  • It does not match the behavior of other functions in numpy; np.sin(1) returns 0.8414709848078965 instead of 0.
  • This behavior can be easily achieved by ivy.astype(ivy.lgamma(x), dtype=x.dtype)
  • Package maintainers using ivy need to worry about the type of input, and casts need to be made.

Thanks for this great project anyway.

Steps to Reproduce Bug

python -c "import ivy; ivy.set_backend('numpy'); print(ivy.lgamma(5)); print(ivy.lgamma(5.))"
python -c "import ivy; ivy.set_backend('numpy'); print(ivy.vector_norm([1, 1])); print(ivy.vector_norm([1., 1.]))"

Environment

NixOS 24.05, Python 3.11

Ivy Version

1.0.0.1

Backend

  • NumPy
  • TensorFlow
  • PyTorch
  • JAX

Device

CPU

Activity

  1. 34j commented on Nov 29, 2024

    @34j
    Author

    (Temporary workaround would be to add + 0. to the argument)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

Bug ReportReport bugs detected in Ivy.

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions