tensorflow的cross_entropy(loss)=nan问题

tensorflow

一个面向所有人的开源机器学习框架

项目地址：https://gitcode.com/gh_mirrors/te/tensorflow

免费下载资源

threadroc

6261人浏览 · 2017-01-01 18:33:41

threadroc · 2017-01-01 18:33:41 发布

解决方案1:

cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y_conv,1e-10,1.0)))

解决方案2:(推荐)

cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv + 1e-10))

原文参考:http://stackoverflow.com/questions/33712178/tensorflow-nan-bug

解释参考:http://blog.csdn.net/sinat_16823063/article/details/53699849

GitHub 加速计划 / te / tensorflow

下载

一个面向所有人的开源机器学习框架

最近提交(Master分支：1 个月前 )

4f64a3d5 Instead, check for this case in `ResolveUsers` and `ResolveOperand`, by querying whether the `fused_expression_root` is part of the `HloFusionAdaptor`. This prevents us from stepping into nested fusions. PiperOrigin-RevId: 724311958 1 个月前

aa7e952e Fix a bug in handling negative strides, and add a test case that exposes it. We can have negative strides that are not just -1, e.g. with a combining reshape. PiperOrigin-RevId: 724293790 1 个月前