A Flow Field Prediction Program (based on OpenSTL)

Overview

Code Structures

• core/ core training plugins and metrics.
• methods/ contains training methods for various prediction methods.
• Model/ contains the control files for various prediction methods and is used to store the model and training results.
• models/ contains the main network architectures of various video prediction methods.
• modules/ contains network modules and layers.
• runfiles/ contains various startup scripts.
• tool/ contains the font files tool/font/, pre-processing file tool/pre-data.pyand standardized file tool/comput_norm.py
• utils/ contains a variety of utilities for model training, model modules, plots, parsers, etc.
• DataDefine.py is used to get the flow field dataset and make a dataloader.
• modelbuild.py is used to build and initialize the model.
• modeltrain.py is used to train, validate, test and inference about model.
• main.py is the main function that runs the program.
• inference.py is used for model inference.
• test.py is used for model test.

Multiple Spatio_temporal Attention (MSTA) Network

The code for Multiple Spatio_temporal Attention (MSTA) Layer can be found in MSTA/MSTA.py

The detailed MSTA code (Spatial Encoder/Decoder & MSTA module) can be found in models/simvp_model.py and modules/simvp_modules.py

• Overview architecture

• MSTA Block

• MSTA Layer

• Large Kernel Attention (LKA) & Multiple Fusion Attention (MFA)