关于实现和paper的区别，请指点一下 about person-reid-triplet-loss-baseline

[Yitian-Li]

噢，不好意思，我的训练是stride2，写错了。

我用的命令：
python script/experiment/train.py -d '(0,)'
--only_test false
--dataset market1501
--last_conv_stride 2
--normalize_feature false
--trainset_part trainval
--exp_dir TRAIN-S2-gpu0
--steps_per_log 50
--epochs_per_val 10

结果：
=========> Test on dataset: market1501 <=========

Extracting feature...
1000/1000 batches done, +1.73s, total 87.83s
Done, 88.01s
Computing distance...
Done, 0.83s
Computing scores...
User Warning: Version 0.18.1 is required for package scikit-learn, your current version is 0.20.4. As a result, the mAP score may not be t uninstall scikit-learnand thenpip install scikit-learn==0.18.1 Done, 15.43s Single Query: [mAP: 71.78%], [cmc1: 86.73%], [cmc5: 94.83%], [cmc10: 96.97%] Multi Query, Computing distance... Done, 0.74s Multi Query, Computing scores... User Warning: Version 0.18.1 is required for package scikit-learn, your current version is 0.20.4. As a result, the mAP score may not be t uninstall scikit-learn and then pip install scikit-learn==0.18.1
Done, 14.95s
Multi Query: [mAP: 79.54%], [cmc1: 91.42%], [cmc5: 96.94%], [cmc10: 98.16%]
Re-ranking distance...
./tri_loss/utils/distance.py:35: RuntimeWarning: invalid value encountered in sqrt
dist = np.sqrt(squared_dist)
Done, 91.14s

[huanghoujing]

您好，感谢关注！
这个代码和文章有一些细节不同，而且有TensorFlow和Pytorch的差异，可能复现不出来文章的分数。

1. 文章按照iteration数来算的话，那确实相当于训了很多epoch
2. 数据增强在这里，这份代码只用了镜像对称
3. 是的

[Yitian-Li]

那么请问除了stride不一样之外，还有哪些差别呢

[huanghoujing]

1. 训练迭代次数
2. PK采样的时候，一个batch内有多少人，每个人多少张图片
3. 网络backbone后面有一个、或两个全连接层

其它地方的差异我也已经记不清楚了，得翻出In defense那篇论文仔细找找才知道了。

[Yitian-Li]

非常感谢！