31、SensorNet：低功耗教育神经网络框架全解析

SensorNet：低功耗教育神经网络框架全解析

一、SensorNet训练算法

在对SensorNet进行深入研究时，首先要了解其训练算法。以下是训练SensorNet以预测标签（动作）的算法：

Algorithm 5 Train SensorNet to predict labels (actions)
Input: The Network N as defined in Fig. 10.6. An input dataset D of size k (d1..dk) sampled from
various sensors with each point having M attributes.
Output: Predict the class label li for a datapoint di
# Consider the training batch size to be b, the learning rate LR and reshape() changes the shape
of the tensor.
# W is the size of the sliding window.
# epochs is the number of epochs for which the model is trained.
# For categorical crossentropy refer to Eq.10.3
# xtrain is a list of images and ytrain has the expected labels.
for i ← W to D do # After reshape the ten