torch-sci 1.0.0

Differentiable PyTorch functions to calculate scientific notation.

torch-sci

Differentiable PyTorch functions to calculate scientific notation.

Installation

Bring your own PyTorch, then install this package:

pip install torch-sci

Usage

import torch

import torch_sci


x = torch.tensor(1234.0, requires_grad=True)
sig, exp = torch_sci.to_sci(x)
sig, exp
# => tensors: 1.234, 3.0

# some made up loss function, to test differentiability
loss = sig * exp
loss.backward()
assert x.grad != 0.0

torch_sci.from_sci(sig, exp)
# => tensor: 1234

# more test cases
torch_sci.to_sci(torch.tensor(-1234.0, requires_grad=True))
# => -1.234, 3
torch_sci.to_sci(torch.tensor(-0.1234, requires_grad=True))
# => -1.234, -1
torch_sci.to_sci(torch.tensor(0.0, requires_grad=True))
# => 0, -8 (common trick of + 1e-8 to avoid log 0)
torch_sci.to_sci(torch.tensor(0.1234, requires_grad=True))
# => 1.234, -1
torch_sci.to_sci(torch.tensor(1234.0, requires_grad=True))
# => 1.234, 3

Development

Setup

Install uv for dependency management if you haven't already. Then run:

# setup virtualenv
venv sync

Unit Tests

uv run pytest