run_demo_inference.py

#!/usr/bin/env python3
"""Run ZoomLDM-NAIP demo inference using local demo assets."""

from pathlib import Path

import numpy as np
import torch
from diffusers import DiffusionPipeline


def preprocess_naip_ssl(npy_path: Path) -> torch.Tensor:
    # Copied from dataset material:
    #   rearrange (n_embed, 1024) -> (1024, h, h), normalize per-feature.
    feat = np.load(npy_path).astype(np.float32)  # (n_embed, 1024) or (1024,)
    if feat.ndim == 1:
        feat = feat[:, None]
    mean = feat.mean(axis=0, keepdims=True)
    std = feat.std(axis=0, keepdims=True)
    feat = (feat - mean) / (std + 1e-8)
    h = int(np.sqrt(feat.shape[0]))
    feat = feat.reshape(h, h, feat.shape[1]).transpose(2, 0, 1)  # (1024, h, h)
    return torch.from_numpy(feat).float()


def main() -> None:
    repo = Path(__file__).resolve().parent
    demo_dir = repo / "demo_images"
    demo_data = repo / "demo_data"
    demo_dir.mkdir(exist_ok=True)

    # Use repo-local demo assets only (3x sample -> magnification label 2).
    src_img = demo_dir / "input.jpeg"
    src_feat = demo_data / "0_ssl_feat.npy"
    if not src_img.exists():
        raise FileNotFoundError(f"Missing demo input image: {src_img}")
    if not src_feat.exists():
        raise FileNotFoundError(f"Missing demo SSL feature: {src_feat}")

    ssl_feat = preprocess_naip_ssl(src_feat).unsqueeze(0).to("cuda")  # (1, 1024, h, h)
    magnification = torch.tensor([2], device="cuda", dtype=torch.long)

    pipe = DiffusionPipeline.from_pretrained(
        str(repo),
        custom_pipeline=str(repo / "pipeline_zoomldm.py"),
        trust_remote_code=True,
        local_files_only=True,
    ).to("cuda")

    out = pipe(
        ssl_features=ssl_feat,
        magnification=magnification,
        num_inference_steps=50,
        guidance_scale=2.0,
        generator=torch.Generator(device="cuda").manual_seed(42),
    )
    out.images[0].save(demo_dir / "output.jpeg")
    print(f"Saved {demo_dir / 'input.jpeg'}")
    print(f"Saved {demo_dir / 'output.jpeg'}")


if __name__ == "__main__":
    main()