The introductory route of deep learning based on personal experience (simple and fast)
https://blog.csdn.net/weixin_44414948/article/details/109704871
Deep learning, a small demo exercise of the first stage of the introductory stage of deep learning
https://blog.csdn.net/weixin_44414948/article/details/109864551
Recently, I have been instructing the introductory of deep learning for juniors and sisters. After they completed the first stage of demo exercises, they arranged the following two stages of demo exercises.
(Mainly exercise their ability to build neural networks, read data sets, export training parameter data, and draw charts)
1. Use numpy to manually build a convolutional neural network (without calling the deep learning framework), and deepen your understanding of the convolutional operation, parameter update and other rules and implementation of the underlying CNN.
Data set : Boston housing price data set, the accuracy rate is higher than 70%.
Additional requirements : draw the loss and acc curves of the training process on the same graph, and intercept the graph of the accuracy of the test set.
Note: The previous mnist data set is not used because the reaction of the younger brother is a bit difficult. . .
The better implementation code is as follows:
https://blog.csdn.net/weixin_44414948/article/details/111460618
#############################分隔符#######################################
2. Use a deep learning framework (tensorflow2.0 and above, pytorch1.0 and above) to build a convolutional neural network.
Data set : Skeleton behavior data set, the accuracy rate is higher than 60%.
Introduction to the Skeleton Behavior Data Set:
This data set is collected by students from the behavior detection direction of this research group. There are 6 types of labels for running, jumping, leaning on products, smoking, playing with mobile phones, and background behaviors. The data is the human skeleton node (25 points). ) A txt file of the sequence, with a total of 600,000 frames. The example is shown in the following figure. Each line has a total of 75 points (xyz coordinates of 25 joint points), and one line is one frame.
Note: The author himself is the research direction of behavior detection
Additional requirements : draw the loss and acc curves of the training process on the same graph, and intercept the graph of the accuracy of the test set.
A better implementation code is as follows:
https://blog.csdn.net/weixin_44414948/article/details/110940712
#############################分隔符#######################################
3. Use deep learning frameworks (tensorflow2.0 and above, pytorch1.0 and above) to build VGG, inception, and resnet networks .
Data set : mnist, accuracy rate is higher than 90%.
Additional requirements : training process loss curve (a canvas draws 3 network loss curves at the same time ), training process acc curve diagram (a canvas draws 3 network acc curves at the same time ).
The better implementation code is as follows:
https://blog.csdn.net/weixin_44414948/article/details/110674054