search

iamhankai / ghostnet.pytorch

[CVPR2020] GhostNet: More Features from Cheap Operations
https://arxiv.org/abs/1911.11907
523 stars 118 forks source link

关于cheap conv中relu的使用问题 #12

Closed pawopawo closed 4 years ago

pawopawo commented 4 years ago

https://github.com/iamhankai/ghostnet.pytorch/blob/9aabd96d8132aa05f2a3e40a35f5b67c019494a2/ghost_net.py#L88 https://github.com/iamhankai/ghostnet.pytorch/blob/9aabd96d8132aa05f2a3e40a35f5b67c019494a2/ghost_net.py#L94 为什么上面两处一个relu=True，一个relu=False呢

iamhankai commented 4 years ago

Inspired by ShuffleNet: https://arxiv.org/abs/1707.01083

deep-practice commented 4 years ago

能提供预训练模型么？

pawopawo commented 4 years ago

Inspired by ShuffleNet: https://arxiv.org/abs/1707.01083 image

ShuffleNet中，concat或add后是有relu的

iamhankai commented 4 years ago

Thanks for your question. We remove the 2nd ReLU like ShuffleNet and remove the 3rd ReLU like MobileNetV2. The setting is more suitable for GhostNet.

KnightWin123 commented 4 years ago

@pawopawo 你好 作者的tensorflow版本的代码是不是没有提供训练代码阿 这个怎么搞 刚刚入门求指教

pawopawo commented 4 years ago

@pawopawo 你好 作者的tensorflow版本的代码是不是没有提供训练代码阿 这个怎么搞 刚刚入门求指教

我用的是pytorch训练，没用过tensorflow～

pawopawo commented 4 years ago

Thanks for your question. We remove the 2nd ReLU like ShuffleNet and remove the 3rd ReLU like MobileNetV2. The setting is more suitable for GhostNet.

请问ghostnet 1.0的训练策略和ghostnet 1.3的训练策略一样么？ ghostnet1.0 可以复现出73.9，但是用相同的训练策略训练ghostnet 1.3，精度只有75.2。 论文中的ghostnet 1.3的精度是75.7

KnightWin123 commented 4 years ago

@pawopawo 你好 作者的tensorflow版本的代码是不是没有提供训练代码阿 这个怎么搞 刚刚入门求指教

我用的是pytorch训练，没用过tensorflow～

那你的训练代码是自己写的吗..可以分享一下不

recusant7 commented 4 years ago

image