CSI Prediction

This repository explores and compares deep learning models for Channel State Information (CSI) prediction in massive MIMO systems, with a focus on improving spectral and temporal predictions using various architectures like STEMGNN, Transformer, BiLSTM, and STNet.

✅ Credits: This work extends the contributions of Sharan Mourya et al. and is intended for comparative research and educational purposes.

For questions, collaborations, or feedback:

(a) Mailapalli Purushotham, Email: purus15987@gmail.com

---

📌 Reference Projects

It builds upon and integrates prior work from the following repositories:

• microsoft/StemGNN
  • Defu Cao, Yujing Wang, Juanyong Duan, Ce Zhang, Xia Zhu, Conguri Huang, Yunhai Tong, Bixiong Xu, Jing Bai, Jie Tong, Qi Zhang
  • Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting
• sharanmourya/CSI-Prediction
  • Sharan Mourya, Pavan Reddy, Sai Dhiraj Amuru
  • Spectral Temporal Graph Neural Network for massive MIMO CSI Prediction.
• sharanmourya/Pytorch_STNet
  • Sharan Mourya, Sai Dhiraj Amuru
  • A Spatially Separable Attention Mechanism For Massive MIMO CSI Feedback

---

🚀 Project Overview

The goal of this project is to:

• Predict future CSI values in MIMO systems using deep learning models.
• Evaluate performance across varying channel dimensions and mobility scenarios.
• Use STNet for compression/decompression and integrate with STEMGNN pipeline.
• Models Implemented
  • STEMGNN – Graph neural network for time-series prediction.
  • Transformer – Sequence modeling architecture.
  • BiLSTM – Bidirectional recurrent neural network.
  • STNet – Attention-based CSI compression network (encoder-decoder).

---

⚙️ Requirements

• Core

  • Python >= 3.7
  • numpy >= 1.19.5
  • scipy >= 1.5.4
  • pandas >= 1.1.5
  • torch >= 1.9.0
  • torchvision
  • torchaudio
  • CUDA, install the appropriate torch version for your CUDA setup

• Plotting and visualization

  • matplotlib
  • seaborn

• Utility

  • tqdm
  • scikit-learn
  • tensorboard/wandb (optional)

Install dependencies with:

pip install -r requirements.txt

---

📂 Directory Structure

    CSI-Prediction/
    │
    ├── CSI-Prediction/X/                 # compressed channel matrices for Urban Macro (UMa) scenario
    │
    ├── data/                           # COST 2100 dataset
    │   ├── DATA_Htestin.mat
    │   └── ...
    │
    ├── assets/                          # Visualizations (NMSE, RMSE plots)
    │   ├── NMSE_comparison.png
    │   └── RMSE_comparison.png
    │
    ├── Results/                      # Stores output results and evaluation metrics
    ├── STNet Pretrained/             # Contains pretrained STNet models for CSI compression/decompression
    │
    ├── data_loader/                    # Data loading and preprocessing utilities
    │   └── data_utils.py
    │
    ├── models/                         # Deep learning model implementations
    │   ├── base_model.py               # STEMGNN architecture
    │   ├── transformer.py              # Transformer model for CSI prediction
    │   └── LSTM_model.py               # BiLSTM model for sequence modeling
    │   └── handler.py                  # BiLSTM model for sequence modeling
    │
    ├── utils/                          # Miscellaneous helper functions
    │   └── evaluation_metrics.py       # NMSE, RMSE, etc.
    │
    ├── Spectral_Temporal_Graph_Neural_Network_for_Massive_MIMO_CSI_Prediction.pdf  # Research paper
    ├── auto_compare.py               # Script for automated comparison of models
    ├── auto_run.py                   # Script to automate training and evaluation runs
    ├── compare.py                    # Script for comparing different model performances
    ├── practice.py                   # Script for practice or experimentation
    ├── requirements.txt              # Lists required Python packages
    ├── spectral_efficiency.py        # Calculates spectral efficiency metrics
    ├── stnet.py                      # Implementation of STNet for CSI compression
    └── train.py                      # Script to train the models

---

📊 Experimental Setup

STNet

• A Spatially Separable Attention Mechanism For Massive MIMO CSI Feedback

1. Dataset
   • For simulation purposes, we generate channel matrices from COST2100 model. Chao-Kai Wen and Shi Jin group provides a ready-made version of COST2100 dataset in Dropbox.
   • After downloading, place it in the data/ directory
2. Training STNet
   • Firstly, choose the compression ratio 1/4, 1/8, 1/16, 1/32 or 1/64 by populating the variable encoded_dim with 512, 256, 128, 64 or 32 respectively.
   • Secondly, choose a scenario "indoor" or "outdoor" by assiging the variable envir the same.
   • Finally run the file STNet.py to begin training...

python stnet.py

- pretrained weights of STNet encoder, decoder are saved in STNet [Pretrained/]() directory or you can use pretrained weights and use it in STEMGNN CSI-Prediction

STEMGNN CSI-Prediction -

1. Dataset

   • We have provided the compressed channel matrices for Urban Macro (UMa) scenario with code word dimensions of 128, 256 and 512 is provided in the google folder shared here due to its large size.
   • Once dataset (train, test) is downloaded, we recommend to organize the folders according to project directory CSI-Prediction/X/

2. Training and Evaluation

   • Set the window size, horizon, dataset name, and other training parameters in train.py before running it.
   • you can use auto_run.py to run all models with different scenarios and different channel dimensions.
   • After training is finished, the predicted channel matrices of train and test are stored in (train, test) [{args.model_name}output/X{args.channel}_{args.kmph}kmph/] automatically.

Train CSI predictors using:

python train.py

Or run batch jobs with:

python auto_run.py

3. STNet Decoder
   • Try to Reconstruct the original dataset (train, test) [CSI-Prediction/X/] and predicted dataset (train, test) [{args.model_name}output/X{args.channel}_{args.kmph}kmph/] using compare.py
   • you can use auto_compare.py to reconstruct (train, test) for all models with different scenarios and different channel dimensions.

python compare.py

Run batch reconstructions with:

python auto_compare.py

4. Plotting Results
   • Run spectral_eficiency.py by importing the decompressed channel matrices obtained in the previous step. This step produces the spectral efficiency plots of STEM GNN for various mobilities.
   • for comparative analysis - use models = ['stem', 'transformer', 'LSTM'] line no. 14 and cr_values = [128, 256, 512] line no. 11
   • the results saved with predifined name at Results/

python spectral_efficiency.py

Configure:

• models = ['stem', 'transformer', 'LSTM']
• cr_values = [128, 256, 512]

---

📈 Comparative Analysis

The repository includes practice.py to evaluate model performance and visualize results:

• Metrics: Normalized Mean Square Error (NMSE), Root Mean Square Error (RMSE)
• Plots: NMSE_comparison.png, RMSE_comparison.png

---

---

📁 Full Results

Download the complete project output from: 📦 Google Drive Folder

---

🤝 Contact

For questions, collaborations, or feedback:

📧 Mailapalli Purushotham

✉️ Email: purus15987@gmail.com

🔗 LinkedIn

---

📄 License

This repository is available for educational and non-commercial research purposes only. For academic or commercial usage, please refer to the respective licenses of the base repositories.