我同意 @cyniikal ，您的网络似乎对于此数据集来说太复杂了.使用单层模型，我可以在训练数据上达到93.75％的精度，在测试数据上可以达到86.7％的精度.

I agree with @cyniikal, your network seems too complex for this dataset. With a single layer model, I was able to achieve 93.75% accuracy on the training data and 86.7% accuracy on the test data.

在我的模型中，我使用了 GradientDescentOptimizer ，它与您一样将 cross_entropy 最小化.我还使用了 16 批量大小.

In my model, I used GradientDescentOptimizer that minimized cross_entropy just as you did. I also used a size 16 batch-size.

您看到的与我的主要区别在于:

The main difference I see between your approach and mine is that I:

1. OneHot编码标签
2. 使用单层网络而不是VGG-16

请参阅此带有我的单层模型代码示例的笔记本.

如果您想在神经网络中添加图层(网络将收敛，将遇到更多困难)，我强烈建议您阅读消失梯度问题.请参见此页面以解决消失梯度.

If you would like to add layers to your neural network (the network will converge with more difficulties), I highly recommend reading this article on neural nets. Specifically, since you added sigmoid as your last activation function, I believe you are suffering from a vanishing gradient problem. See this page to address the vanishing gradient.