PACE-Net: Point-Aware Conditional Enhancement Network for Radio Map Reconstruction from Sparse Observations

Official PyTorch implementation of PACE-Net.

High-resolution radio maps must be reconstructed from extremely sparse field measurements — on a 256×256 grid the measured pixels cover barely 0.5% of the image. These are the only pixels with ground-truth anchoring, yet their contribution is diluted in global losses and smoothed away by convolution.

PACE-Net is a point-aware conditional enhancement network that anchors on measured points and radiates their information across the feature stream:

• PFE (Point Feature Enhancement) module — uses the point mask and the measured values as a two-channel condition to drive zero-initialized gated convolutions at nine decoder positions. Adds only +0.139% parameters (18.4k on a 13.28M backbone) and starts as an exact identity mapping, so training always improves on the baseline.
• Point-aware loss — point weighting (with hard-example up-weighting)
  • a spatial gradient constraint on the whole-map reconstruction.
• Two-phase training — freezes the coarse U-Net for 80 epochs, then briefly fine-tunes the whole network for 20 epochs.

Results

On 990 RadioMapSeer test samples (99 maps × 10 transmitters, 10–300 randomly drawn sparse measurements per sample), PACE-Net is compared with nine methods under identical conditions (Kriging, RBF, DeepAE, UNet, cGAN, RadioUNet, RME-GAN, ACT-GAN, TAN) and ranks first on global RMSE, held-out RMSE and NMSE:

Model	Point RMSE ↓	Held-out RMSE ↓	Global RMSE ↓	Global NMSE ↓
RadioUNet (backbone)	0.0271	0.0397	0.0394	0.0286
PACE-Net (ours)	0.0126	0.0318	0.0316	0.0183

Relative to RadioUNet: global RMSE −19.8%, NMSE −36.0%, measurement-point RMSE −53.5%, with a negligible +0.139% model-size increase. On a real campus with at most 300 field measurements, the predictions reach 0.901 correlation and 3.42 dBm RMSE at the measurement points (within the 5 dBm engineering margin).

Repository layout

PACE-Net/
├── pace_net.py          # PACE-Net model (PointAwareWNetV2 + PFE v2)
├── modules.py           # RadioWNet backbone (from the RadioUNet repository)
├── point_losses.py      # Point-aware loss + evaluation metrics
├── loaders.py           # RadioUNet_s_points dataset (derived from RadioUNet)
├── data.py              # Data pipeline: 501/100/99 split, 10~300 points
├── train.py             # Two-phase training recipe
├── evaluate.py          # Verify the trained model on the test split
├── requirements.txt
├── LICENSE              # CC BY-NC 4.0
└── assets/              # Architecture figure used in this README

No pre-trained checkpoints are distributed; train.py writes the trained weights to weights/ (created automatically).

1. Setup

git clone https://github.com/xcljn/PACE-Net.git
cd PACE-Net
pip install -r requirements.txt

2. Dataset: RadioMapSeer

Download the RadioMapSeer dataset (free from IEEE DataPort after registration; DPM simulation variant). Extract it so that the directory layout is:

PACE-Net/RadioMapSeer/
├── png/buildings_complete/   # 700 building maps
├── png/antennas/             # 700×80 transmitter maps
└── gain/DPM/                 # 700×80 gain maps

Alternatively, set the environment variable RADIOMAPSEER_DIR or pass --data-dir to point to the dataset location.

The 501/100/99 train/val/test map split is the deterministic standard RadioUNet split (np.random.seed(42) inside loaders.py), and the random sparse-point sampling is governed by the seed in data.set_seed(20260922), applied after loader construction — both are fixed, so evaluate.py is deterministic and bit-for-bit reproducible across runs and machines.

3. Train

python train.py

• Phase 1: first-level U-Net frozen, second level trained for 80 epochs (Adam 1e-4, cosine schedule).
• Phase 2: full model fine-tuned for 20 epochs (Adam 2e-5).
• Loss: L = L_global + λ1·L_point + λ2·L_spatial (λ1=1.5, λ2=2.0) with hard-example point weighting (β=1, τ=0.02).
• EMA (decay 0.999); validation/model selection use the EMA weights.
• Batch 15 fits an 8 GB GPU.

Smoke test on a subset:

python train.py --epochs 2 --stage2-epochs 1 --max-train-batches 20 --max-val-batches 10

Starting weights: by default train.py expects an optional starting second-level checkpoint via --init-weights; when neither it nor --baseline-weights is supplied, it exits with an explicit message instead of silently starting from random weights. Two supported options:

• --baseline-weights with the official RadioWNet_s second-level checkpoint (Trained_Model_SecondU_s.pt) from the RadioUNet repository, placed at init/Trained_Model_SecondU_s.pt; or
• any compatible second-level checkpoint via --init-weights init/<file>.pt.

Non-PFE keys are loaded strictly (missing or shape-mismatched keys abort with an error); PFE v2 parameters always start from their zero-initialized identity mapping.

4. Evaluate

After training (weights/model_best.pt exists):

python evaluate.py

The four metrics are the acceptance criteria for training quality. Expected output on the full test set (990 samples) when the model is trained with the full recipe:

point RMSE    : 0.01263   (reference: 0.0126)
held-out RMSE : 0.03178   (reference: 0.0318)
map RMSE      : 0.03157   (reference: 0.0316)
map NMSE      : 0.01829   (reference: 0.0183)

Acknowledgments

This implementation builds on the RadioUNet codebase and the RadioMapSeer dataset by Levie et al. We thank the authors for releasing their code and data.

License

Released under CC BY-NC 4.0 for non-commercial research use. modules.py and loaders.py are derived from RadioUNet and keep their original license. The RadioMapSeer dataset is governed by its own license and is not distributed with this repository.