pytorch/vision

Classification references does not work without distributed setup

Open

#6,529 创建于 2022年9月1日

在 GitHub 查看
 (2 评论) (0 反应) (0 负责人)Python (15,050 star) (6,858 fork)batch import
bughelp wantedmodule: reference scripts

描述

If you don't set the respective env vars

https://github.com/pytorch/vision/blob/d5bd8b728f14c33b339fc45c90ca39be339bce3f/references/classification/utils.py#L255-L258

training will not be distributed and in turn the backend will not be initialized. However, during evaluation we check

https://github.com/pytorch/vision/blob/d5bd8b728f14c33b339fc45c90ca39be339bce3f/references/classification/train.py#L88

unguarded, which then fails with

RuntimeError: Default process group has not been initialized, please make sure to call init_process_group.

cc @datumbox

贡献者指南

技术栈
pythonpytorch
领域
machine learningbackend
议题类型
bug
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-3 hours
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
basic Pythonbasic PyTorchunderstanding of torch.distributed
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
70
研究方向
The issue is in PyTorch Vision's classification reference scripts. The bug occurs when distributed training env vars are not set, causing an unguarded `dist.*` call in `references/classification/train.py` line 88. The fix should add a check like `if dist.is available() and dist.is initialized()` before using distributed utilities. Also ensure that `utils.py` lines 255-258 properly guard the distributed initialization. See the linked lines in the issue for exact locations.