alibaba / MNN

MNN is a blazing fast, lightweight deep learning framework, battle-tested by business-critical use cases in Alibaba
http://www.mnn.zone/
8.79k stars 1.67k forks source link

[Android]多次初始化和释放,概率性输出Tensor数据为nan值 #384

Closed GaussRieman closed 1 year ago

GaussRieman commented 5 years ago

请问有人遇到过类似的情况吗,在Android上强制多次初始化和释放,随机出现output从第一个开始全部为nan值,且之后每一帧结果都是nan值。

没有这个bug的时候,可以正常运行。

3q!

GaussRieman commented 5 years ago

MNN::Tensor *tensor_out = interpreter->getSessionOutput(session, nullptr); MNN::Tensor copy(tensor_out); tensor_out->copyToHostTensor(&copy);

打印的是 copy.host() 的值,为nan

GaussRieman commented 5 years ago

检查了初始化和输入tensor都没问题,但是Tensor_out的值都是-nan,且再也无法正确运行

GaussRieman commented 5 years ago

且有小概率第一次启动就出现这种情况

GaussRieman commented 5 years ago

MNN的纯CPU版本,4线程,机器是vivo R11 高通SDM660, 8核

rafezh commented 4 years ago

请问这个问题解决了不

wangzhaode commented 1 year ago

已解决