哪些权重将用于预测未来的响应？
我正在做时间序列数据预测的工作。
输入信号是空气中灰尘颗粒的每日浓度，格式为 (10x24)，10 = 天，每天有 24 个值，然后通过使用将其转换为 (1,240) 的行向量
input = imresize(dust, [1, 10*24]); % converts matrix into vector
为了训练我的网络，我有建立模型 (3:5:1)(tanh, tanh)(0.05)(1)(500)，
其中 3= 输入，5 个隐藏层神经元，1 个输出层，(tanh tanh) 输入隐藏层和隐藏输出层的传递函数，学习率为 0.05，1= 偏差，迭代次数为 500。我得到了经过训练的网络，并且跟踪是绝对的。
哪些层权重将用于预测未来响应（即输入隐藏层或隐藏输出层），因为它们的维度
为隐藏层的输入= inputweights(input,hidden)= 3 x 5 矩阵 隐藏到输出层 = outputwhts(output,hidden)= 1x5 行向量。
我想根据我的输入数据权重来预测 24 个值预测和 168 个值预测。
I am doing work on time series data prediction.
The input signal is the daily concentration of dust particles in the air and having format (10x24), 10 =days and for each day 24 values, then it is converted to row vector of (1,240) by using
input = imresize(dust, [1, 10*24]); % converts matrix into vector
For training my network, I have made the model (3:5:1)(tanh, tanh)(0.05)(1)(500),
where 3= inputs, 5 hidden layer neuron, 1 output layer, (tanh tanh) transfer function for input- hidden layer, and hidden-output layer, the learning rate is 0.05, 1= bias and iterations are 500. I get trained network and tracking was absolute.
Which layer weights will be used in prediction for future response (i.e input-hidden layer or hiden-output layer,) as their dimensions are
Input to hidden layer= inputweights(input,hidden)= 3 x 5 matrix
hidden to output layer = outputwhts(output,hidden)= 1x5 row vector.
I want to predict the 24 values prediction and 168 value prediction based on my input data weights.
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
输入隐藏权重和隐藏输出权重都参与预测。神经网络做出的预测始终依赖于网络中的所有权重。
现在，听起来好像您想要使用在大小为 24 的示例上训练的网络来预测大小为 168 的示例，这就是为什么您想知道要保留哪一组权重。
如果是这种情况（有点不清楚），那么坏消息是，一般来说，如果不完全重新训练网络，就无法更改网络输入或输出的大小。
我的建议是计算 168 个值的示例与每个 24 个值的示例相似，并根据这些示例训练网络。
Both the input-hidden and the hidden-output weights are involved in the prediction. Predictions made by neural networks always rely on all of the weights in the network.
Now, it sounds as though you want to use a network trained on examples of size 24 to predict examples of size 168, and that's why you want to know which set of weights to keep.
If that is the case (its a little unclear), then the bad news is that in general, you can't change the size of a network's input or output without completely retraining it.
My advice would be to calculate what 168 value example would be analogous to each 24 value example and train a network on those examples.