Map2APS

Code release for the Map2APS benchmark: direct angle power spectrum (APS) prediction from urban geometry and transmitter/receiver locations.

This repository contains only the methods used in the paper. Data files, checkpoints, generated figures, and detailed result CSVs are not included.

Methods

Paper name	Code path
MS-AReg	baselines/ms_areg/
Adv-MLP	baselines/adv_mlp/
MS-MLP	baselines/ms_mlp/
RadioUNet	baselines/radiounet/
ResNet-MLP	baselines/resnet_mlp/
ResNet-Conv1D	baselines/resnet_conv1d/
ViT-Reg	baselines/vit_reg/

Data

Dataset link:

Baidu Netdisk: https://pan.baidu.com/s/1GHPOfXYwvDfO1zmtPpc0yw?pwd=p7sz
Extraction code: p7sz

Expected raw data layout for preprocessing:

<raw_root>/
  map_0/0_adps/aps/aps_*.npy
  map_1/1_adps/aps/aps_*.npy
  ...
<buildings_dir>/
  0.json
  1.json
  ...

Preprocessed samples are saved as .npz files under:

<processed_dir>/samples/

Each sample stores:

• aps_norm: normalized APS vector with shape (180,)
• cond_img: condition image with shape (3, 256, 256)

Setup

pip install -r requirements.txt

The experiments were developed with PyTorch and CUDA GPUs. Before training, edit each YAML config so that data.preprocessed_dir points to your preprocessed dataset and logging.checkpoint_dir points to your output directory.

Preprocess

python data/preprocess.py \
  --aps_root <raw_root> \
  --buildings_dir <buildings_dir> \
  --output_dir <processed_dir> \
  --aps_norm std

Optional map subset:

python data/preprocess.py \
  --aps_root <raw_root> \
  --buildings_dir <buildings_dir> \
  --output_dir <processed_dir> \
  --map_ids 0,1,2 \
  --aps_norm std

Train

python baselines/ms_areg/train_ms_areg.py --config baselines/ms_areg/config_ms_areg.yaml --gpu 0
python baselines/adv_mlp/train_adv_mlp.py --config baselines/adv_mlp/config_adv_mlp.yaml --gpu 0
python baselines/ms_mlp/train_ms_mlp.py --config baselines/ms_mlp/config_ms_mlp.yaml --gpu 0
python baselines/radiounet/train_radiounet.py --config baselines/radiounet/config_radiounet.yaml --gpu 0
python baselines/resnet_mlp/train_resnet_mlp.py --config baselines/resnet_mlp/config_resnet_mlp.yaml --gpu 0
python baselines/resnet_conv1d/train_resnet_conv1d.py --config baselines/resnet_conv1d/config_resnet_conv1d.yaml --gpu 0
python baselines/vit_reg/train_vit_reg.py --config baselines/vit_reg/config_vit_reg.yaml --gpu 0

Evaluate

python eval_compare.py \
  --data_dir <processed_dir> \
  --mode full \
  --output_dir <output_dir> \
  --models \
    ms_areg:<path_to_ms_areg_checkpoint> \
    adv_mlp:<path_to_adv_mlp_checkpoint> \
    ms_mlp:<path_to_ms_mlp_checkpoint> \
    radiounet:<path_to_radiounet_checkpoint> \
    resnet_mlp:<path_to_resnet_mlp_checkpoint> \
    resnet_conv1d:<path_to_resnet_conv1d_checkpoint> \
    vit_reg:<path_to_vit_reg_checkpoint>

The default train/test split is hard-coded in data/dataset.py: maps 0-45 for training and maps 46-50 for held-out evaluation.