恢复模型时，摘要写入以避免在张量板上重叠的正确方法是什么 [英] What's the right way for summary write to avoid overlapping on tensorboard when restore a model

问题描述

我用 tensorflow 编写了一个 CNN 图像分类器，并使用 tensorboard 来监控训练.但是，当我停止并从检查点恢复时，会出现如下重叠:

I wrote a CNN image classifier with tensorflow and use tensorboard to monitor the training. However when I stop and restore from a checkpoint, there are overlaps like:

我遵循了 Tensorboard README 将 SessionStatus.START 消息写入摘要文件，但它似乎不起作用.

I followed the instruction on the Tensorboard README to write a SessionStatus.START message to the summary file, but it doesn't seem to work.

这是我的代码:

summary_writer.add_session_log(SessionLog(status=SessionLog.START),global_step=step)

推荐答案

我不知道这是否是一个答案，但是如果您将 global_step 变量放入 tensorflow 变量(以将其与您的数据一起存储)，然后当您恢复时模型，global_step 变量也恢复到它的旧值，它会从那里在 tensorboard 中继续，还是?抱歉，我没有测试过.

i dont know if its an answer, but if you put the global_step variable into a tensorflow variable (to store it with your data) and then, when you restore the model, the global_step variable is also restored to its old value, it will go on from there in tensorboard, or? Sorry,i have not tested it.

这篇关于恢复模型时，摘要写入以避免在张量板上重叠的正确方法是什么的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！