最终层的输出结果是NAN

yiran-THU commented 4 years ago

同学你好，感谢你开源EAST-caffe，很漂亮的工作。但是我在调试的时候，用caffe作为框架，发现最终层（f_score和geo_concat）的值全部都是NAN。于是我逐层输出每一层的feature map，发现conv_11_pw_hswish开始的输出中包含INF，并且之后的层里INF越来越多，导致最终无法得到想要的结果。

我的系统是ubuntu16，caffe是weiliu89/SSD的caffe，opencv版本4.1.1，RELU6层直接用普通的RELU代替了。请问下你的RELU6层自己重新实现的，还是也用的普通RELU代替？

yiran-THU commented 4 years ago

我在dnn框架下，测试了未用RELU代替RELU6，测试结果正常；但是如果用RELU代替RELU6，则也会出现NAN情况；

yiran-THU commented 4 years ago

所以up主能不能顺便也开源下你RELU6层的.cu和.cpp啊，谢谢

SURFZJY commented 4 years ago

The ReLU6 implementation can be found here.

yiran-THU commented 4 years ago

@SURFZJY

I recompile caffe with relu6 from here last night. It is also working.

Thanks again.

ghost commented 4 years ago

@yiran-THU Compiling caffe is a nightmare to me. How did you compile caffe to work with east. is there a guide you followed

SURFZJY / EAST-caffe

最终层的输出结果是NAN #3