keras2ncnn过程中的错误总结

刺骨的言语ヽ痛彻心扉 2022-03-31 13:14 179阅读 0赞

1、getops使用，参见我的博客：

[https://blog.csdn.net/weixin\_41770169/article/details/85695457][https_blog.csdn.net_weixin_41770169_article_details_85695457]

2、不支持的层

> (tensorflow) bash-3.2$ python /Users/xuqiong/PycharmProjects/keras2ncnn/Keras2NCNN.py -i /Users/xuqiong/PycharmProjects/keras2ncnn/ssrnet\_3\_3\_3\_64.h5
> 
> Traceback (most recent call last):
> 
>   File "/Users/xuqiong/PycharmProjects/keras2ncnn/Keras2NCNN.py", line 392, in <module>
> 
>     text\_net, binary\_weights = k2n\_converter.convert()
> 
>   File "/Users/xuqiong/PycharmProjects/keras2ncnn/Keras2NCNN.py", line 370, in convert
> 
>     layer\_ncnn\_lst = self.\_dump\_layer\_weights(layer)
> 
>   File "/Users/xuqiong/PycharmProjects/keras2ncnn/Keras2NCNN.py", line 339, in \_dump\_layer\_weights
> 
>     raise ValueError("non supported layer:\{\}".format(layer\_name))
> 
> ValueError: non supported layer:t

猜测是h5文件中不含有结构，只有权重信息。因此，先将纯权值的模型改成结构+权值。

3、模型保存为：结构+权值

[https://blog.csdn.net/weixin\_41770169/article/details/85697281][https_blog.csdn.net_weixin_41770169_article_details_85697281]

[https://blog.csdn.net/weixin\_41770169/article/details/85705489][https_blog.csdn.net_weixin_41770169_article_details_85705489]

4、再次运行keras2ncnn

还是步骤2同样的问题，跟进去后发现，如下网址给的脚本有很多问题，由于项目第一周期时间紧，决定放弃直接从keras转ncnn。而是改成keras改成caffe，再转ncnn。

[https://github.com/yanghanxy/Keras2NCNN][https_github.com_yanghanxy_Keras2NCNN]

[https_blog.csdn.net_weixin_41770169_article_details_85695457]: https://blog.csdn.net/weixin_41770169/article/details/85695457
[https_blog.csdn.net_weixin_41770169_article_details_85697281]: https://blog.csdn.net/weixin_41770169/article/details/85697281
[https_blog.csdn.net_weixin_41770169_article_details_85705489]: https://blog.csdn.net/weixin_41770169/article/details/85705489
[https_github.com_yanghanxy_Keras2NCNN]: https://github.com/yanghanxy/Keras2NCNN