// the find
bubbliiiing/yolox-pytorch
这是一个yolox-pytorch的源码，可以用于训练自己的模型。
A PyTorch re-implementation of YOLOX by a Chinese developer who has done the same for most of the YOLO family (v3 through v7). Targets people who want to train object detection models on their own datasets without wrestling with the official Megvii codebase. The README is in Chinese.
Pretrained weights for all six model sizes (nano through x) are provided with COCO benchmarks that match the paper closely. The training script exposes a reasonable set of knobs — cosine LR decay, Adam/SGD selection, multi-GPU support, per-batch LR scaling — without burying them. VOC-format training pipeline is straightforward: run one annotation script, then train.py, done. Part of a consistent series of YOLO repos from the same author, so if you already know one of them the patterns transfer directly.
Pinned to pytorch==1.2.0, which is ancient — released in 2019, predates AMP, modern DataLoader improvements, and anything using torch.compile. The last commit was August 2023 and the repo is effectively unmaintained. Pretrained weights are hosted on Baidu Pan, which is inaccessible without a Chinese account; COCO weights are on GitHub releases but VOC weights aren't. No ONNX or TensorRT export path included, so getting this to production inference requires you to wire that up yourself. The decoupled head and SimOTA label assignment are implemented but there's no serious test suite, so you're trusting that the numbers in the README aren't cherry-picked.