项目仅依赖scipy的gaussian_filter1d函数，需安装完整scipy包吗？

Hey there! Let's tackle this question head-on—your intuition is actually right, but let's break down why and explore practical alternatives:

The Short Answer #

Unfortunately, there's no feasible way to install just gaussian_filter1d and its dependencies without installing the full SciPy package.

Why This Isn't Possible #

• SciPy's Distribution Model: SciPy is published as a single, monolithic package on PyPI. There are no official split packages (like scipy-ndimage) that let you install only specific submodules. Pip operates at the package level, not the submodule or individual function level.
• Interconnected Underlying Dependencies: Even though gaussian_filter1d lives in scipy.ndimage.filters, it relies on a web of tightly integrated SciPy components—core array utilities, low-level image processing primitives, and configuration modules that can't be cleanly stripped out and used in isolation.
• Packaging Limitations: SciPy's wheel and source distributions are built to deliver the entire library. There's no official mechanism to cherry-pick individual functions or submodules during installation.

Practical Alternatives #

If minimizing dependency bloat is a top priority, here are two solid options:

1. Use a Pure NumPy Implementation #

You can replicate the core functionality of gaussian_filter1d with just NumPy, avoiding SciPy entirely. Here's a functional, production-ready example:

import numpy as np

def custom_gaussian_filter1d(arr, sigma):
    # Calculate kernel size (captures ~99.7% of the Gaussian curve with 6*sigma)
    kernel_length = int(6 * sigma) + 1
    # Ensure kernel is odd-sized for symmetric convolution
    if kernel_length % 2 == 0:
        kernel_length += 1
    
    # Generate and normalize Gaussian kernel
    x = np.linspace(-3*sigma, 3*sigma, kernel_length)
    kernel = np.exp(-x**2 / (2 * sigma**2))
    kernel /= kernel.sum()  # Preserve signal amplitude after convolution
    
    # Apply convolution with same-length output
    return np.convolve(arr, kernel, mode="same")

This implementation behaves similarly to SciPy's version for most use cases, and only requires NumPy as a dependency.

2. Accept the Full SciPy Installation #

While SciPy is a large package, it's a standard dependency in most Python data science environments—many of your users might already have it installed. Additionally, pip's caching system means users won't have to re-download the package if they've installed it before. For most projects, this is the simplest and most reliable approach.

内容的提问来源于stack exchange，提问作者HexxNine