LoMACS-SVDNet: Orthogonality without Decompositions

第四届“华为杯”无线通信算法大赛

AI-Enabled Robust SVD Operator for Wireless Communication
End-to-end network that predicts (U, S, V) for large MIMO channels without any QR/SVD/EVD/inversion inside the network.
Final score aligns with the official metric: Score = 100 × AE + MACs.

Competition Result: Final Score 63.0 — 4th place (Third Prize) at Huawei Tech Arena 2025.

1. Highlights

• Axial Low-rank Frequency Gate (ALF): 2D-FFT front-end gating keeps informative angle/delay bands; hidden size is structurally prunable.
• Grouped Projected Attention (GPA): K/V column projection with length k_len → complexity O(T·k_len); interpretable column dictionary; prunable by energy.
• Gated Depthwise-Conv: local smoothing with very low MACs.
• Neural Ortho Refiner (NOR): one-step, decomposition-free orthogonality refinement on the Stiefel manifold; learns step sizes (a, b).
• Spectral Self-Calibration (SSC): annealed blend of predicted S and measured |diag(UᴴHV)|.
• AEPlus Loss: L_rec + λ L_ortho + w_E L_energy + w_D L_diag + w_S L_smatch with smooth schedules.
• Step-3 Structured Pruning: rebuild a smaller isomorphic network (no masks) → real MACs drop under PyTorch Profiler; short finetune recovers AE.
• Train/Prune/Infer Unified: one model.py driving the full pipeline.

2. Repository Layout

.
├── model.py          # train / prune_ft / infer
├── scripts/          # convenience shell scripts
├── docs/             # analysis plots used in README
├── ckpts/            # checkpoints (ignored by git)
├── submissions/      # npz outputs (ignored by git)
├── requirements.txt
└── .gitignore

3. Installation

conda create -n svdnet python=3.10 -y
conda activate svdnet
pip install -r requirements.txt

4. Data Preparation

Place the official data (e.g. CompetitionData2) under ./data2/, following the organizer’s naming:

data2/
  Round2CfgData1.txt
  Round2CfgData2.txt  
  Round2CfgData3.txt  
  Round2CfgData4.txt  
  Round2TrainData1.npy...
  Round2TrainLabel1.npy...
  Round2TestData1.npy...
  

5. Quick Start

5.1 Train (Step-1/2 → Step-3 prune → finetune → best.pth)

python model.py \
  --mode train \
  --data_dir ./data2 \
  --out_dir ./submissions

During training, the code measures real Mega MACs using torch.profiler and prints val_AE, MACs and Score.

5.2 Prune + Finetune existing Step-1/2 checkpoint

python model.py \
  --mode prune_ft \
  --data_dir ./data2 \
  --src_ckpt ./ckpts/preprune.pth \
  --keep_klen 28 --keep_hidden 28 \
  --ft_epochs 60 --ft_lr 2e-4

5.3 Inference & Packaging for Submission

python model.py \
  --mode infer \
  --data_dir ./data2 \
  --ckpt ./ckpts/best.pth \
  --out_dir ./submissions/round2

This produces 1.npz…4.npz under --out_dir, each containing keys U, S, V, C.
C is the Mega MACs measured by PyTorch Profiler, used in the final score.

6. Model Notes

• Compliance: No QR/SVD/EVD/inversion is used inside the network. Orthogonality is enforced by loss + NOR.
• Normalization: per-sample Frobenius normalization during train/test; we rescale S back at inference.
• Augmentation: power-aware complex noise and random antenna (row/column) dropout.
• Complexity: get_avg_flops() calls PyTorch Profiler to compute real MACs / sample, avoiding proxy FLOPs.

7. Reproduce Our Setting

The defaults in make_cfg() reflect our final setting:

• DIM=64, DEPTH=2, GROUPS=4, K_LEN=32 → 28 (after pruning), GATE_HIDDEN=32 → 28
• Schedules: λ ramp; τ: 0.90 → 0.60; w_E, w_D, w_S smoothly decay.
• Finetune: epochs=60, lr=2e-4.

8. Visualization

9. License

This repository is for the competition/research purpose. Commercial use is subject to the organizer’s policy.

10. Acknowledgements

Thanks to the organizers and the open-source community.