Selecting for Task Requirements: Exploring the Effectiveness of CPU-based Datasets and Optimizer Technique Selections on Informer
Zhiruo Qiu · 2023
Informer experiments often ignore the effects of calculation conditions, data size and parameter setting on model performance. Recently, GPU-based Transformers are blindly adapted for tasks in different fields. GPU has a high degree of parallelism, resulting in a lack of intuitive evaluation of model's own performance. In this CPU-based experiment, the sample capacity is reduced to 1/10 and 3 optimizer strategies are used in the model training. The dataset was mainly derived from Informer experiments, supplemented with a public dataset from Kaggle to investigate the generalization capabilities of Informer. MSE (Mean Squared Error) and MAE (Mean Absolute Error) were used as evaluation indexes. The results show that CPU meet the needs of study design. MSE increased by 0.342 at most, indicating that there was no significant loss in model performance while calculation reduction. Informer has stability, with MSE and MAE of 0.452 and 0.341 respectively on public dataset. However, MSE and MAE increased by 1.271 and 0.547, which means that the model performance decreased a lot. Reasonable selection of optimizer can effectively improve model performance, especially the use of Informerstack+Adagrad to achieve the reduction of MSE 0.166 and MAE 0.056. Models and parameters' selection based on task condition are proposed in this paper to obtain desired results.