How many iterations of training？

NIEQINQIN commented 6 years ago

您好，非常感谢大神的代码！在voc2007 数据集上训练，但是mAP 只有72，请问一下你迭代了多少次呢？实验中迭代次数是7.5万次，如果换成COCO 的数据集？应该在哪里改anchor 和ratio 呢？是在配置文件中吗？

unsky commented 6 years ago

readme已经写了，用voc2007+voc2012训练训练八万次即可

NIEQINQIN commented 6 years ago

谢谢大神的回复，我对您的readme 里面有句话不是很理解 in the paper the anchor setting is Ratios： [0.5,1,2],scales :[8,]

With the setting and P2~P6, all anchor sizes are [32,64,128,512,1024],but this setting is suit for COCO dataset which has so many small targets.

But the voc dataset targets are range [128,256,512].

So, we desgin the anchor setting:Ratios： [0.5,1,2],scales :[8,16], this is very import for voc dataset. 我看了config.py 文件里面好像只有 C.FPNRATIOS = [0.5,1,2] C.FPNRSCALES = 2 ** np.arange(4,6) 这里的scale 是16、32 为什么是这个呢？我对VOC 2007的变成进行了聚类，预测的聚类中心个数：7 [[ 75.89890956] [ 180.74493869] [ 310.83273847] [ 125.66768035] [ 386.90149049] [ 242.98615928] [ 34.32092669]] 0: scale:75.8989095611243 1: scale:180.74493868691016 2: scale:310.83273846745465 3: scale:125.66768034606157 4: scale:386.90149049240097 5: scale:242.98615927978665 6: scale:34.32092668661748 如果实验中采用的是scales :16、32，那么在conv_2层的anchor 就会比较大基础的anchor 会不会都是由前面conv_2 conV_3 输出的呢？

unsky commented 6 years ago

聚类中心不大能说明啥，应该说属于一个聚类的数量才是有意义的，真正的anchor大小是scale*stride 而voc以经验起见是在128,256,512中的较多。使用聚类策略分析是一种很好的方法，你可以接着分析，找到最合适的anchor设置。

unsky / FPN

How many iterations of training？ #75