keras 的 example 文件 mnist_hierarchical_rnn.py 解析

最新推荐文章于 2023-08-07 00:23:07 发布

zhqh100

最新推荐文章于 2023-08-07 00:23:07 发布

阅读量193

点赞数

分类专栏： TensorFlow python

本文链接：https://blog.youkuaiyun.com/zhqh100/article/details/105144650

版权

python 同时被 2 个专栏收录

51 篇文章

订阅专栏

TensorFlow

48 篇文章

订阅专栏

很显然，我没有看懂 HRNN 是啥意思，没有去看论文，应该就是一种RNN结构的变形吧

网络结构如下：

__________________________________________________________________________________________
Layer (type)                            Output Shape                        Param #
==========================================================================================
input_1 (InputLayer)                    (None, 28, 28, 1)                   0
__________________________________________________________________________________________
time_distributed_1 (TimeDistributed)    (None, 28, 128)                     66560
__________________________________________________________________________________________
lstm_2 (LSTM)                           (None, 128)                         131584
__________________________________________________________________________________________
dense_1 (Dense)                         (None, 10)                          1290
==========================================================================================
Total params: 199,434
Trainable params: 199,434
Non-trainable params: 0
__________________________________________________________________________________________

输入是图片，输出是分类

类似的，mnist_irnn.py 的网络结构如下：

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
simple_rnn_1 (SimpleRNN)     (None, 100)               10200
_________________________________________________________________
dense_1 (Dense)              (None, 10)                1010
_________________________________________________________________
activation_1 (Activation)    (None, 10)                0
=================================================================
Total params: 11,210
Trainable params: 11,210
Non-trainable params: 0
_________________________________________________________________

——————————————————————

总目录