torch-scatter

library2.1.2pypypiunverified

PyTorch Extension Library of Optimized Scatter Operations. Provides efficient scatter, gather, segment, and softmax operations for PyTorch tensors, commonly used in graph neural networks. Current version 2.1.2, released Oct 2023. Maintained by Rusty1s, part of the PyTorch Geometric ecosystem.

ai-ml data

pip install torch-scatter

github ↗package ↗homepage ↗

INSTALL

IMPORT

SIG · TORCH-SCATTER

torch-scatter

ai-mlpythonv2.1.2

Install

—

Import

—

Disk

—

Pass rate

0/ 10

Env Coverage0 / 10

glibc

3.9–3.13

musl

3.9–3.13

Install & Compatibility

Where this runs

tested against v? · pip install

Install × environment matrix

Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux

musl

py 3.10–3.910 runs

build_error

glibc

py 3.10–3.910 runs

build_error

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

scatter

✓ from torch_scatter import scatter

scatter_add

✓ from torch_scatter import scatter_add

scatter_max

✓ from torch_scatter import scatter_max

scatter_min

✓ from torch_scatter import scatter_min

scatter_mean

✓ from torch_scatter import scatter_mean

scatter_mul

✓ from torch_scatter import scatter_mul

scatter_softmax

✓ from torch_scatter import scatter_softmax

scatter_std

✓ from torch_scatter import scatter_std

segment_coo

✓ from torch_scatter import segment_coo

segment_csr

✓ from torch_scatter import segment_csr

Compute max over indices using scatter_max.

import torch from torch_scatter import scatter_max src = torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0]) index = torch.tensor([0, 0, 1, 1, 1]) out, argmax = scatter_max(src, index, dim=0) print(out) # tensor([2., 5.])

Debug

Known issues

breakingVersion 2.1.0 removed `scatter_logsumexp` and changed `scatter_softmax` to no longer require an `eps` argument. Code using these must update calls.

fix

Use `torch_scatter.scatter_logsumexp` (new function) or remove `eps` from `scatter_softmax`.

affects: >=2.1.0

deprecated`scatter_` functions (e.g., `scatter_add_`) are deprecated in favor of the out-of-place versions. They may be removed in a future release.

fix

Replace `scatter_add_(src, index, out)` with `scatter_add(src, index, dim, out=out)`.

affects: >=2.0.0

gotchaWhen using CUDA, the installed wheel must match the PyTorch CUDA version exactly. Installing from PyPI may pull a CPU-only build if the CUDA version mismatch is detected.

fix

Install using the PyTorch Geometric wheel index: `pip install torch-scatter -f https://data.pyg.org/whl/torch-{torch_version}+{cu_version}.html`.

affects: all

gotcha`scatter_std` with `unbiased=False` uses Bessel's correction? Actually it uses population std. Check documentation: the flag defaults to `False` which computes the biased standard deviation.

fix

Set `unbiased=True` to compute sample standard deviation.

affects: >=2.0.0

Upgrade

Version history

2.1.2latest on PyPI · released Oct 6, 2023

Audit

Dependencies

torchrequiredCore dependency, torch-scatter compiles against specific PyTorch versions

Agent activity

24 hits · last 30 days

node

OpenAI (training)

Resources

githubhttps://github.com/rusty1s/pytorch_scatter ↗

packagehttps://pypi.org/project/torch-scatter/ ↗

homepagehttps://rusty1s.github.io/pytorch_scatter/ ↗