简体繁体 English

那些喀拉拉邦损失和准确性怪异吗？

[英]Are those keras loss and accuracy weird?

原文 2018-10-21 20:27:00 3 1 python/ tensorflow/ keras/ training-data/ loss

I have a relatively small mri dataset and I'm trying to do a binary segmentation. 我有一个相对较小的mri数据集，我正在尝试进行二进制分割。 I have built an ordinary U-Net structure and trained it. 我已经建立了一个普通的U-Net结构并对其进行了培训。

But the output seems a bit weird to me. 但是输出对我来说似乎有点奇怪。 Both train and validation accuracies stucked at a value first, but then both accuracies made a sudden big jump at 27th or 28th epoch. 训练和验证精度首先都固定在一个值上，但是随后在第27或28个时代，这两个精度突然大幅度提高。

Loss graph looks more acceptable, next is the graphs: 损耗图看起来更容易接受，接下来是这些图：

Accuracy Graph: 精度图：

精度图

Loss Graph: 损耗图：

损失图

I have another issue that even if I have an %97-98 accuracy on training data, when I tested it on some images from training data, results converted to binary mask were not that good. 我还有另一个问题，即使我对训练数据的准确性达到%97-98 ，当我在训练数据的某些图像上对其进行测试时，将结果转换为二进制掩码的效果也不是很好。

Then I have decreased the threshold from 0.5 to 0.35 while retrieving output images and the results were almost perfect. 然后我在检索输出图像时将阈值从0.5降低到0.35，结果几乎是完美的。

What do you think about that? 您对此有何看法？ thanks in advance. 提前致谢。

1 个解决方案

They seem a little off with those stuck epochs, it really means the model isn't learning (weights are not changing, new cases are not providing useful information) but that is totally plausible. 它们似乎与那些停滞的时代有些偏离，这确实意味着该模型没有学习（权重没有变化，新案例没有提供有用的信息），但这完全是合理的。

Just to be sure. 只是要确定。 What optimizer are you using? 您正在使用什么优化程序？ and did you try with another one? 您尝试了另一种吗？