（fast-reid）visualize_result.py中的distmat = 1 - torch.mm(q_feat, g_feat.t())理解

柔光的暖阳◎ 2023-02-26 11:21 12阅读 0赞

## 0.疑问  ##

# compute cosine distance
    distmat = 1 - torch.mm(q_feat, g_feat.t())
    distmat = distmat.numpy()

为什么此处是计算余弦距离？

## 1.相关与理解 ##

（1）余弦距离计算（[https://www.cnblogs.com/chaosimple/p/3160839.html][https_www.cnblogs.com_chaosimple_p_3160839.html]）：

![20200715143127332.png][]

（2）torch.mm()

torch.mm()含义到底是什么（[https://blog.csdn.net/Real\_Brilliant/article/details/85756477][https_blog.csdn.net_Real_Brilliant_article_details_85756477]）：

这篇博客中说：

torch.mm(a, b)是矩阵a和b矩阵相乘

（3）代码中做了norm操作

在github中搜索torch.mm，找到有意义相关的issue。参考此处[https://github.com/layumi/AICIty-reID-2020/issues/17][https_github.com_layumi_AICIty-reID-2020_issues_17]，去代码中搜索，发现特征出来的时候，用了norm操作：

\fast-reid-master\fastreid\engine\defaults.py line155
    with torch.no_grad():  # https://github.com/sphinx-doc/sphinx/issues/4258
        image = image.to(self.device)
        inputs = {"images": image}
        predictions = self.model(inputs)
        # Normalize feature to compute cosine distance
        pred_feat = F.normalize(predictions)
        pred_feat = pred_feat.cpu().data
        return pred_feat

cosine similarity--torch.mm(q\_feat, g\_feat.t())。 方向接近，夹角小，similarity高。

## 参考 ##

[1.余弦距离、欧氏距离和杰卡德相似性度量的对比分析][https_www.cnblogs.com_chaosimple_p_3160839.html]

[2.torch.mul() 和 torch.mm() 的区别][https_blog.csdn.net_Real_Brilliant_article_details_85756477]

[3.AICIty-reID-2020/issues/17][https_github.com_layumi_AICIty-reID-2020_issues_17]

[https_www.cnblogs.com_chaosimple_p_3160839.html]: https://www.cnblogs.com/chaosimple/p/3160839.html
[20200715143127332.png]: /images/20230209/30ca76aa71bd4a2994409080d4277a6f.png
[https_blog.csdn.net_Real_Brilliant_article_details_85756477]: https://blog.csdn.net/Real_Brilliant/article/details/85756477
[https_github.com_layumi_AICIty-reID-2020_issues_17]: https://github.com/layumi/AICIty-reID-2020/issues/17