Pix2PixHD for Wind Comfort ML

This repository contains a heavily stripped-down and specialized version of the original NVIDIA pix2pixHD architecture, adapted specifically for Wind Comfort Machine Learning Research.

Unlike traditional image-to-image translation (which operates on 3-channel RGB images), this pipeline has been modified to map 8-channel physical geometry inputs mapping to 1-channel wind deficit predictions.

Dataset Specification

The model expects inputs and outputs formatted as .npy arrays tightly normalized around [-1, 1] for the generator.

Input Channels (8)

Idx	Channel	Description
0	SDF	Signed Distance Field (distance to nearest wall)
1	Bldg_height	Height of the building at the specific pixel
2	Z_relative	Height slice of the CFD simulation
3	U_over_Uref	Background wind ratio (inlet profile)
4	X_local	X distance from building center
5	Y_local	Y distance from building center
6	dir_sin	Sine of Wind Direction
7	dir_cos	Cosine of Wind Direction

Output Target (1)

Idx	Channel	Description
0	mag_U	Target Wake Deficit

---

Quickstart & Scripts

We use uv for lightning-fast Python dependency management. Make sure uv is installed, and the environment will auto-sync.

1. Data Preprocessing

If you have new CFD raw .csv results in input_csv/, run the preprocessing script to generate the proper 8-channel input and 1-channel output arrays:

uv run preprocess_csv.py

This will populate the datasets/wind/ directory and compile a stats.json for normalization.

2. Local Training

To test the model architecture locally (defaulting to CPU if no CUDA is available):

# Trains for 25 epochs, saving a checkpoint every 5 epochs
uv run python train.py --name pix2pix --dataset_mode wind --dataroot datasets/wind --input_nc 8 --output_nc 1 --niter 25 --niter_decay 0 --save_epoch_freq 5 --label_nc 0 --no_instance --display_freq 30 --gpu_ids -1 --nThreads 0

3. Training on HPC (PACE)

To submit a full-scale job on an HPC Slurm cluster (leveraging H200 GPUs):

sbatch slurm/train_PACE.sbatch

4. Visualizing Progress

To generate a side-by-side comparison GIF (Ground Truth vs. Model Prediction) of your training progress, run:

uv run python make_gif.py

Outputs to training_progress_comparison.gif.

---

Repository Cleanup

Note: All legacy dataloaders for Cityscapes/Faces, unused TensorRT inference endpoints, and 1024p bash scripts have been purged to keep this repository clean and strictly focused on wind engineering datasets.