(c) Sharan Mourya, email: [email protected]
This repository holds the pytorch implementaion of the original models described in the paper
Sharan Mourya, Sai Dhiraj Amuru, "A Spatially Separable Attention Mechanism For Massive MIMO CSI Feedback"
- Python >= 3.7
- PyTorch >= 1.2
- Scipy >= 1.8.0
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.
Once dataset is downloaded, we recommend to organize the folders as follows
├── STNet # The cloned STNet repository
│ ├── stnet.py
├── data # The data folder
│ ├── DATA_Htestin.mat
│ ├── ...
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...
Normalized Mean Square Error (NMSE) and Floating-Point Operations per second (FLOPS) achieved by STNet for different compression ratios under different scenarios are tabulated below.
| S.No | Compression Ratio | indoor | outdoor | Flops |
|---|---|---|---|---|
| 1 | 1/4 | -31.81 | -12.91 | 5.22M |
| 2 | 1/8 | -21.28 | -8.53 | 4.38M |
| 3 | 1/16 | -15.28 | -5.72 | 3.96M |
| 4 | 1/32 | -9.42 | -3.51 | 3.75M |
| 5 | 1/64 | -7.81 | -2.46 | 3.65M |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent