Performance and backbone

MioChiu / TF_CenterNet

A pure TensorFlow implementation of CenterNet(Objects as Points)

MIT License

89 stars 10 forks source link

Open phucnhs opened 4 years ago

phucnhs commented 4 years ago

Thanks for source code ,

Can you tell me the performance on coco or voc on backbone mobilenetv2?
I want to add backbone mobilenetv3, I have to do it.
I think the backbone DLA is quite good, why don't you add it in this code ! Thanks you

dl8207531 commented 4 years ago

我也想知道不同backbone map多少，我自己用mobilenetv3训练过

luqioayun commented 4 years ago

我也想知道不同backbone map多少，我自己用mobilenetv3训练过

请问你有跑backbone是resnet18的吗？mAP是多少呢？

dl8207531 commented 4 years ago

voc map mobilenetv3 384x384 69+ 512x512 70+ resnet18 384x384 70+ 512x512 没测 @luqioayun

luqioayun commented 4 years ago

voc map mobilenetv3 384x384 69+ 512x512 70+ resnet18 384x384 70+ 512x512 没测 @luqioayun

好的，谢谢！我刚好搞了512x512的，只有60+ 。。。。。。

luqioayun commented 4 years ago

voc map mobilenetv3 384x384 69+ 512x512 70+ resnet18 384x384 70+ 512x512 没测 @luqioayun

1、您好，请问在resnet18 384x384 70+mAP这种情况下您的一些参数设置是什么呢？比如说learning rate的衰减方式、batchsize的大小设置~ 这个是我目前的参数设置~

dl8207531 commented 4 years ago

lr=0.000125 batchsize=24 70000, 110000, 140000 对应步数除10 你根据你的batchsize 按比例设置我认为最关键的是在imagenet预训练的模型本身不能太差