有关Precision的疑问

[StarvenLan]

我在训练的每一轮中都计算了相应的Precision，如图结果发现Precision是上下波动的，而不是稳步上升的，请问是什么情况？

[bubbliiiing]

emm，map应该比较上升，precision，看门限的呀，不过我感觉你这个数据集是不是比较少，而且损失这么低吗0 0

[StarvenLan]

数据集有1w多张，precision的门限一直都是一样的，我代码改动成每轮训练后计算一次map，门限这些参数都没有变，所以有点困惑为啥precision的波动这么大。

[bubbliiiing]

precsion……我也不是很确定，recall有稳定上升吗0 0。我感觉也不太合理，这种波动

[StarvenLan]

我又训练了一遍：

0.00% = face AP || score_threhold=0.5 : F1=0.00 ; Recall=0.00% ; Precision=0.00% 0.27% = face AP || score_threhold=0.5 : F1=0.00 ; Recall=0.00% ; Precision=0.00% 1.36% = face AP || score_threhold=0.5 : F1=0.00 ; Recall=0.04% ; Precision=47.37% 8.30% = face AP || score_threhold=0.5 : F1=0.09 ; Recall=4.60% ; Precision=65.25% 14.86% = face AP || score_threhold=0.5 : F1=0.17 ; Recall=9.69% ; Precision=68.13% 19.89% = face AP || score_threhold=0.5 : F1=0.24 ; Recall=14.57% ; Precision=69.10% 24.96% = face AP || score_threhold=0.5 : F1=0.33 ; Recall=23.70% ; Precision=56.91% 27.11% = face AP || score_threhold=0.5 : F1=0.32 ; Recall=20.87% ; Precision=73.27% 28.09% = face AP || score_threhold=0.5 : F1=0.33 ; Recall=20.57% ; Precision=78.81% 33.44% = face AP || score_threhold=0.5 : F1=0.41 ; Recall=29.61% ; Precision=65.90% 35.64% = face AP || score_threhold=0.5 : F1=0.40 ; Recall=26.83% ; Precision=79.49% 34.81% = face AP || score_threhold=0.5 : F1=0.42 ; Recall=30.97% ; Precision=67.29% 39.14% = face AP || score_threhold=0.5 : F1=0.46 ; Recall=34.40% ; Precision=70.24% 41.29% = face AP || score_threhold=0.5 : F1=0.48 ; Recall=38.01% ; Precision=65.89% 40.74% = face AP || score_threhold=0.5 : F1=0.48 ; Recall=36.73% ; Precision=67.44% 40.24% = face AP || score_threhold=0.5 : F1=0.48 ; Recall=36.29% ; Precision=70.14% 43.60% = face AP || score_threhold=0.5 : F1=0.50 ; Recall=38.97% ; Precision=71.11% 43.24% = face AP || score_threhold=0.5 : F1=0.50 ; Recall=39.24% ; Precision=70.14% 43.46% = face AP || score_threhold=0.5 : F1=0.50 ; Recall=37.01% ; Precision=75.31% 45.16% = face AP || score_threhold=0.5 : F1=0.51 ; Recall=39.55% ; Precision=73.53% 44.28% = face AP || score_threhold=0.5 : F1=0.51 ; Recall=38.24% ; Precision=74.93% 45.92% = face AP || score_threhold=0.5 : F1=0.53 ; Recall=40.93% ; Precision=73.61% 46.31% = face AP || score_threhold=0.5 : F1=0.53 ; Recall=41.69% ; Precision=71.50% 46.12% = face AP || score_threhold=0.5 : F1=0.52 ; Recall=40.74% ; Precision=73.60% 46.98% = face AP || score_threhold=0.5 : F1=0.53 ; Recall=40.64% ; Precision=76.28% 47.80% = face AP || score_threhold=0.5 : F1=0.54 ; Recall=43.81% ; Precision=71.02% 46.89% = face AP || score_threhold=0.5 : F1=0.53 ; Recall=40.48% ; Precision=76.31% 47.30% = face AP || score_threhold=0.5 : F1=0.53 ; Recall=39.96% ; Precision=79.16% 47.51% = face AP || score_threhold=0.5 : F1=0.54 ; Recall=42.65% ; Precision=72.63% 48.51% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=43.52% ; Precision=73.53% 48.99% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=43.09% ; Precision=75.67% 48.83% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=43.49% ; Precision=73.22% 48.97% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=42.92% ; Precision=75.49% 49.64% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=44.26% ; Precision=74.18% 49.55% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=44.75% ; Precision=72.79% 49.88% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=46.51% ; Precision=70.88% 50.16% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=43.42% ; Precision=77.78% 50.28% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=43.94% ; Precision=78.49% 50.19% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=44.41% ; Precision=75.63% 49.74% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=42.37% ; Precision=80.33% 49.78% = face AP || score_threhold=0.5 : F1=0.55 ; Recall=41.82% ; Precision=80.97% 50.29% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=45.19% ; Precision=74.05% 50.83% = face AP || score_threhold=0.5 : F1=0.57 ; Recall=46.78% ; Precision=72.25% 50.50% = face AP || score_threhold=0.5 : F1=0.56 ; Recall=44.49% ; Precision=76.69% 50.79% = face AP || score_threhold=0.5 : F1=0.57 ; Recall=45.28% ; Precision=75.30% 51.07% = face AP || score_threhold=0.5 : F1=0.57 ; Recall=45.21% ; Precision=76.54% 50.80% = face AP || score_threhold=0.5 : F1=0.57 ; Recall=45.79% ; Precision=74.05% 51.49% = face AP || score_threhold=0.5 : F1=0.57 ; Recall=46.30% ; Precision=75.75%

[bubbliiiing]

这个貌似正常很多，F1基本稳定上升