"""Pack the Hugging Face Space folder on the server: fp16 state dict + the 15 Pakistani official test tiles + app files. Run ON THE SERVER: cd ~/pakflood/code && .venv/bin/python pakflood/pack_demo.py --out ~/pakflood/demo Then copy ~/pakflood/demo to the laptop and push it to the Space. """ import argparse import json import shutil from pathlib import Path import numpy as np import pandas as pd import rasterio import torch CKPT = "workdir/pakflood/lora_r32_all/v1/checkpoints/best_state_dict.ckpt" PER_TILE = Path.home() / "pakflood/preds/lora_r32_all_emsr838/per_tile.csv" SPACE_SRC = Path("pakflood/space") # app.py, requirements.txt, README.md def main(): ap = argparse.ArgumentParser() ap.add_argument("--out", default=str(Path.home() / "pakflood/demo")) ap.add_argument("--root", default="data/geoid-flood") a = ap.parse_args() out, root = Path(a.out), Path(a.root) (out / "assets").mkdir(parents=True, exist_ok=True) (out / "samples").mkdir(exist_ok=True) # 1. state dict, 'model.' prefix stripped, float16 sd = torch.load(CKPT, map_location="cpu", weights_only=False) sd = sd.get("state_dict", sd) sd = {k[len("model."):] if k.startswith("model.") else k: (v.half() if v.is_floating_point() else v) for k, v in sd.items()} torch.save(sd, out / "assets" / "pakflood_lora_r32_state_dict_fp16.pt") print("state dict:", len(sd), "tensors,", sum(v.numel() for v in sd.values()) / 1e6, "M params") # 2. the 15 official Pakistani test tiles (pre, post, label), deflate-compressed copies df = pd.read_csv(root / "data_tiles_pk_test.csv", usecols=["tile_id", "event_id", "modality", "image_time", "label_id", "split"]) df = df[(df.modality == "s1grd") & (df.split == "test") & (df.event_id == "EMSR838")].drop_duplicates(["label_id", "image_time", "tile_id"]) iou = {} if PER_TILE.exists(): pt = pd.read_csv(PER_TILE) for _, r in pt.iterrows(): tp, fp, fn = r.cm_22, r.cm_02 + r.cm_12, r.cm_20 + r.cm_21 iou[r.label_id] = float(tp / max(tp + fp + fn, 1)) samples = {} for lid, g in df.groupby("label_id"): pre, post = g[g.image_time == "pre"].tile_id.iloc[0], g[g.image_time == "post"].tile_id.iloc[0] aoi = lid.replace("_label", "").rsplit("-", 1)[0] short = lid.replace("_label", "") d = out / "samples" / short d.mkdir(exist_ok=True) for name, src in [("pre", root / aoi / "s1grd" / f"{pre}.tif"), ("post", root / aoi / "s1grd" / f"{post}.tif"), ("label", root / aoi / "label" / f"{lid}.tif")]: with rasterio.open(src) as r: data, prof = r.read(), r.profile if name != "label": data = data[:2].astype(np.float32) # VV, VH only prof.update(count=2, dtype="float32", compress="deflate", predictor=3, tiled=True, blockxsize=256, blockysize=256) else: prof.update(compress="deflate", tiled=True, blockxsize=256, blockysize=256) with rasterio.open(d / f"{name}.tif", "w", **prof) as w: w.write(data) with rasterio.open(d / "label.tif") as r: lab = r.read(1) px = abs(r.res[0] * r.res[1]) / 1e6 samples[short] = {"pre": f"samples/{short}/pre.tif", "post": f"samples/{short}/post.tif", "label": f"samples/{short}/label.tif", "flood_km2_label": float((lab == 2).sum() * px), "iou_r32": iou.get(lid)} print("packed", short, f"{samples[short]['flood_km2_label']:.1f} km2 flood, r32 IoU", iou.get(lid)) (out / "samples.json").write_text(json.dumps(dict(sorted(samples.items(), key=lambda kv: -kv[1]["flood_km2_label"])), indent=1)) # 3. app files for f in ["app.py", "requirements.txt", "README.md", ".gitattributes"]: if (SPACE_SRC / f).exists(): shutil.copy(SPACE_SRC / f, out / f) total = sum(p.stat().st_size for p in out.rglob("*") if p.is_file()) / 1e6 print(f"demo folder ready: {out} ({total:.0f} MB)") if __name__ == "__main__": main()