Instructions to use msherry/pakflood-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use msherry/pakflood-lora with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
File size: 3,250 Bytes
17285c6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | """Bootstrap 95% confidence intervals for flood IoU from per-tile confusion matrices.
Usage:
.venv/bin/python pakflood/bootstrap.py <name>=<per_tile.csv> [<name>=<per_tile.csv> ...] [--unit tile|event] [--split test] [--n 2000]
Each per_tile.csv comes from pakflood/predict_tiles.py (same tiles for every model). Resamples units (tiles or
events/AOIs) with replacement, recomputes flood IoU from the summed confusion matrices, and reports the 2.5-97.5
percentile interval per model and for every pairwise difference (paired bootstrap: same resample for all models).
"""
import argparse
import numpy as np
import pandas as pd
def iou_from(df: pd.DataFrame, cls: int = 2) -> float:
cm = np.array([[df[f"cm_{i}{j}"].sum() for j in range(3)] for i in range(3)], dtype=np.float64)
tp = cm[cls, cls]
fp = cm[:, cls].sum() - tp
fn = cm[cls, :].sum() - tp
return tp / max(tp + fp + fn, 1)
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("models", nargs="+", help="name=path/to/per_tile.csv")
ap.add_argument("--unit", default="event", choices=["tile", "event"])
ap.add_argument("--split", default=None)
ap.add_argument("--n", type=int, default=2000)
ap.add_argument("--seed", type=int, default=0)
a = ap.parse_args()
tables = {}
for m in a.models:
name, path = m.split("=", 1)
d = pd.read_csv(path)
if a.split:
d = d[d.split == a.split]
d = d.sort_values("label_id").reset_index(drop=True)
d["unit"] = d.label_id if a.unit == "tile" else d.event.str.extract(r"(EMSR\d+)")[0]
tables[name] = d
names = list(tables)
ref = tables[names[0]]
for n in names[1:]:
assert list(tables[n].label_id) == list(ref.label_id), f"{n}: tile set differs from {names[0]}"
units = sorted(ref.unit.unique())
print(f"{len(ref)} tiles, {len(units)} {a.unit}s, {a.n} resamples")
rng = np.random.default_rng(a.seed)
groups = {n: {u: g for u, g in t.groupby("unit")} for n, t in tables.items()}
cols = [f"cm_{i}{j}" for i in range(3) for j in range(3)]
sums = {n: {u: g[cols].sum().values.astype(np.float64) for u, g in groups[n].items()} for n in names}
def iou_vec(cmv: np.ndarray) -> float:
cm = cmv.reshape(3, 3)
tp = cm[2, 2]
return tp / max(cm[:, 2].sum() + cm[2, :].sum() - tp, 1)
boots = {n: np.empty(a.n) for n in names}
for b in range(a.n):
pick = rng.choice(units, size=len(units), replace=True)
for n in names:
boots[n][b] = iou_vec(sum(sums[n][u] for u in pick))
print("\nflood IoU with 95% bootstrap CI")
for n in names:
lo, hi = np.percentile(boots[n], [2.5, 97.5])
print(f" {n:24s} {iou_from(tables[n]):.4f} [{lo:.4f}, {hi:.4f}]")
if len(names) > 1:
print("\npairwise differences (row minus column), 95% CI; interval excluding 0 = significant")
for i, n1 in enumerate(names):
for n2 in names[i + 1 :]:
d = boots[n1] - boots[n2]
lo, hi = np.percentile(d, [2.5, 97.5])
print(f" {n1} - {n2}: {iou_from(tables[n1]) - iou_from(tables[n2]):+.4f} [{lo:+.4f}, {hi:+.4f}]")
if __name__ == "__main__":
main()
|