Preparation
The eighth assignment is Anomaly Detection, which requires the teaching assistant code and data set. Keeping connected to the Internet during the running of the code can automatically download the data set. If there is already a data set, you can close the download data part in the teaching assistant code. Follow this official account to get the code and data set (there are methods at the end of the article).
submit address
Kaggle: www.kaggle.com/competitions/ml2022spring-hw8, students who want to discuss communication can enter the QQ group: 156013866. The following is the job analysis, and the detailed code is at the end of the article.
Simple Baseline (AUC>0.53150)
Method : Run the TA code directly. Note that when running locally or on kaggle, you need to adjust the corresponding file name or path. The score submitted to kaggle is: 0.53158.
Medium Baseline (AUC>0.73171)
Method: CNN model + reduce latent dim . The teaching assistant code uses the VAE model. Here I use the CNN model and add a linear layer to reduce the intermediate data dimension. The socre submitted to kaggle is: 0.74087 . In addition, I also fine-tuned the FCN model and VAE model. Using the FCN model can easily get a result above 0.75, but the VAE model can only hover around 0.60. It may be caused by the powerful 'brain supplement' ability of VAE, and it may be possible by adding models. size to solve.
Strong Baseline (AUC>0.77196)
Method: ResNet model + more epochs + small batch size . The encoder of the model was replaced by ResNet, the epoch was increased to 100, and the batch size was reduced to 128. The socre submitted to kaggle is: 0.77437 . Here I first referred to the teaching assistant assignment PPT, and experimented with the fusion model: cnn + fcn, and found that the results were not ideal.
Boss Baseline (Acc>0.79506)
Method: ResNet model + auxiliary network . On the basis of strong baseline, an additional decoder auxiliary network is used. The ResNet network is consistent with the original training method. The loss function of the decoder network is controlled by resnet, and the result is stronger than resnet. The submitted file is also calculated by this decoder. It came out, the specific process can be seen in the code, here I will not explain why I did it, let’s see who can realize it, it is a small Easter egg, this method has not been taught in Mr. Li’s class, and there are few existing articles. When you see it, those who are interested can communicate in the communication group or private message me. The socre submitted to kaggle by this method is: 0.79557 .
How to get the answer to homework 8:
-
Follow the WeChat public account " Machine Learning Craftsman "
-
Background reply keywords: 202208