Repository navigation
Add outward-rounded tanh second-derivative interval bounds - #190
MoritzMaibaum wants to merge 1 commit into
Conversation
Compute the scalar range from endpoints and included stationary extrema. Bound elementary-function and arithmetic errors with Decimal, including cancellation, saturation and subnormal cases. Expose tanh_double_prime_bounds and a compatible PyTorch adapter; document the rounding argument. Validation: full main test suite, 117 passed.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5b7b7c0011
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| """Interval arithmetic utilities for neural network evaluation.""" | ||
|
|
||
| from .interval import Interval | ||
| from .activations import tanh_double_prime_bounds |
There was a problem hiding this comment.
For users installing only the base package, from intervalnets import tanh_double_prime_bounds still fails: __init__.py proceeds to import .pytorch, that module catches the missing torch import and sets nn = None, and its class IntervalAdd(nn.Module) declaration then raises AttributeError, which this initializer does not catch. This makes the newly documented dependency-free API unusable unless the optional PyTorch extra is installed; the optional import must fail cleanly or be skipped when PyTorch is unavailable.
Useful? React with 👍 / 👎.
1 participant
The main branch has no scalar tanh'' enclosure. Add
tanh_double_prime_bounds(Interval(a, b)), computing its range from the endpoints and any included stationary extrema at ±atanh(1/√3).The implementation uses directed Decimal arithmetic and explicit brackets for elementary-function rounding, including near-zero cancellation, saturation, subnormal values, and uncertain critical-point membership. It needs no additional dependency. A PyTorch adapter preserves float32 outward padding and the known sign; this does not introduce a network Hessian API on main.
Includes the rounding argument, API documentation, and high-precision regression tests. Example: [0.5, 1] gives approximately [-0.769800359, -0.639700008].
Validation: full main test suite passed (117 tests). The committed tree matches the locally tested tree exactly.