pdate configs: remove density splits, add region shift, 90:10 train/val, shared test sets

README.md

@@ -1,125 +1,69 @@
 ---
 configs:
-- config_name: biome_India_Karnataka_train_DRY__ood_WET
-  data_files:
-  - split: train
-    path: "data/configs/biome_India_Karnataka_train_DRY__ood_WET/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_India_Karnataka_train_DRY__ood_WET/id_test-*.parquet"
-  - split: ood_test
-    path: "data/configs/biome_India_Karnataka_train_DRY__ood_WET/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_India_Karnataka_train_DRY__ood_WET/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_India_Karnataka_train_DRY__ood_WET/ood_diff_density-*.parquet"
-- config_name: biome_India_Karnataka_train_WET__ood_DRY
+- config_name: intl_train_IN__ood_US
   data_files:
   - split: train
-    path: "data/configs/biome_India_Karnataka_train_WET__ood_DRY/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_India_Karnataka_train_WET__ood_DRY/id_test-*.parquet"
+    path: "data/configs/intl_train_IN__ood_US/train-*.parquet"
+  - split: val
+    path: "data/configs/intl_train_IN__ood_US/val-*.parquet"
   - split: ood_test
-    path: "data/configs/biome_India_Karnataka_train_WET__ood_DRY/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_India_Karnataka_train_WET__ood_DRY/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_India_Karnataka_train_WET__ood_DRY/ood_diff_density-*.parquet"
-- config_name: biome_India_Rajasthan_train_DRY__ood_WET
+    path: "data/configs/intl_train_IN__ood_US/ood_test-*.parquet"
+- config_name: intl_train_US__ood_IN
   data_files:
   - split: train
-    path: "data/configs/biome_India_Rajasthan_train_DRY__ood_WET/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_India_Rajasthan_train_DRY__ood_WET/id_test-*.parquet"
+    path: "data/configs/intl_train_US__ood_IN/train-*.parquet"
+  - split: val
+    path: "data/configs/intl_train_US__ood_IN/val-*.parquet"
   - split: ood_test
-    path: "data/configs/biome_India_Rajasthan_train_DRY__ood_WET/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_India_Rajasthan_train_DRY__ood_WET/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_India_Rajasthan_train_DRY__ood_WET/ood_diff_density-*.parquet"
-- config_name: biome_India_Rajasthan_train_WET__ood_DRY
+    path: "data/configs/intl_train_US__ood_IN/ood_test-*.parquet"
+- config_name: biome_Rajasthan_train_WET__ood_DRY
   data_files:
   - split: train
-    path: "data/configs/biome_India_Rajasthan_train_WET__ood_DRY/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_India_Rajasthan_train_WET__ood_DRY/id_test-*.parquet"
+    path: "data/configs/biome_Rajasthan_train_WET__ood_DRY/train-*.parquet"
+  - split: val
+    path: "data/configs/biome_Rajasthan_train_WET__ood_DRY/val-*.parquet"
   - split: ood_test
-    path: "data/configs/biome_India_Rajasthan_train_WET__ood_DRY/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_India_Rajasthan_train_WET__ood_DRY/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_India_Rajasthan_train_WET__ood_DRY/ood_diff_density-*.parquet"
-- config_name: biome_US_California_train_DRY__ood_WET
+    path: "data/configs/biome_Rajasthan_train_WET__ood_DRY/ood_test-*.parquet"
+- config_name: biome_Rajasthan_train_DRY__ood_WET
   data_files:
   - split: train
-    path: "data/configs/biome_US_California_train_DRY__ood_WET/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_US_California_train_DRY__ood_WET/id_test-*.parquet"
+    path: "data/configs/biome_Rajasthan_train_DRY__ood_WET/train-*.parquet"
+  - split: val
+    path: "data/configs/biome_Rajasthan_train_DRY__ood_WET/val-*.parquet"
   - split: ood_test
-    path: "data/configs/biome_US_California_train_DRY__ood_WET/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_US_California_train_DRY__ood_WET/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_US_California_train_DRY__ood_WET/ood_diff_density-*.parquet"
-- config_name: biome_US_California_train_WET__ood_DRY
+    path: "data/configs/biome_Rajasthan_train_DRY__ood_WET/ood_test-*.parquet"
+- config_name: biome_Karnataka_train_WET__ood_DRY
   data_files:
   - split: train
-    path: "data/configs/biome_US_California_train_WET__ood_DRY/train-*.parquet"
-  - split: id_test
-    path: "data/configs/biome_US_California_train_WET__ood_DRY/id_test-*.parquet"
+    path: "data/configs/biome_Karnataka_train_WET__ood_DRY/train-*.parquet"
+  - split: val
+    path: "data/configs/biome_Karnataka_train_WET__ood_DRY/val-*.parquet"
   - split: ood_test
-    path: "data/configs/biome_US_California_train_WET__ood_DRY/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/biome_US_California_train_WET__ood_DRY/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/biome_US_California_train_WET__ood_DRY/ood_diff_density-*.parquet"
-- config_name: in_state_train_Karnataka__ood_Rajasthan
+    path: "data/configs/biome_Karnataka_train_WET__ood_DRY/ood_test-*.parquet"
+- config_name: biome_Karnataka_train_DRY__ood_WET
   data_files:
   - split: train
-    path: "data/configs/in_state_train_Karnataka__ood_Rajasthan/train-*.parquet"
-  - split: id_test
-    path: "data/configs/in_state_train_Karnataka__ood_Rajasthan/id_test-*.parquet"
+    path: "data/configs/biome_Karnataka_train_DRY__ood_WET/train-*.parquet"
+  - split: val
+    path: "data/configs/biome_Karnataka_train_DRY__ood_WET/val-*.parquet"
   - split: ood_test
-    path: "data/configs/in_state_train_Karnataka__ood_Rajasthan/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/in_state_train_Karnataka__ood_Rajasthan/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/in_state_train_Karnataka__ood_Rajasthan/ood_diff_density-*.parquet"
-- config_name: in_state_train_Rajasthan__ood_Karnataka
+    path: "data/configs/biome_Karnataka_train_DRY__ood_WET/ood_test-*.parquet"
+- config_name: region_train_North__ood_South
   data_files:
   - split: train
-    path: "data/configs/in_state_train_Rajasthan__ood_Karnataka/train-*.parquet"
-  - split: id_test
-    path: "data/configs/in_state_train_Rajasthan__ood_Karnataka/id_test-*.parquet"
+    path: "data/configs/region_train_North__ood_South/train-*.parquet"
+  - split: val
+    path: "data/configs/region_train_North__ood_South/val-*.parquet"
   - split: ood_test
-    path: "data/configs/in_state_train_Rajasthan__ood_Karnataka/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/in_state_train_Rajasthan__ood_Karnataka/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/in_state_train_Rajasthan__ood_Karnataka/ood_diff_density-*.parquet"
-- config_name: intl_train_IN__ood_US
+    path: "data/configs/region_train_North__ood_South/ood_test-*.parquet"
+- config_name: region_train_South__ood_North
   data_files:
   - split: train
-    path: "data/configs/intl_train_IN__ood_US/train-*.parquet"
-  - split: id_test
-    path: "data/configs/intl_train_IN__ood_US/id_test-*.parquet"
+    path: "data/configs/region_train_South__ood_North/train-*.parquet"
+  - split: val
+    path: "data/configs/region_train_South__ood_North/val-*.parquet"
   - split: ood_test
-    path: "data/configs/intl_train_IN__ood_US/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/intl_train_IN__ood_US/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/intl_train_IN__ood_US/ood_diff_density-*.parquet"
-- config_name: intl_train_US__ood_IN
-  data_files:
-  - split: train
-    path: "data/configs/intl_train_US__ood_IN/train-*.parquet"
-  - split: id_test
-    path: "data/configs/intl_train_US__ood_IN/id_test-*.parquet"
-  - split: ood_test
-    path: "data/configs/intl_train_US__ood_IN/ood_test-*.parquet"
-  - split: ood_same_density
-    path: "data/configs/intl_train_US__ood_IN/ood_same_density-*.parquet"
-  - split: ood_diff_density
-    path: "data/configs/intl_train_US__ood_IN/ood_diff_density-*.parquet"
+    path: "data/configs/region_train_South__ood_North/ood_test-*.parquet"
 ---
 
 # Tree Distribution Shift — Satellite Tree Detection (COCO + HF Datasets)
@@ -127,21 +71,45 @@ configs:
 This is a dataset containing ~30K COCO tree crown annotated satellite image tiles of 400x400 px dimensions. These annotations come from all states in India and California in the United States. 
 This dataset is organized as **configs** (distribution shift benchmarks). Each config provides **three splits**:
 
-- `train`
-- `id_test` (same distribution as train, held-out)
-- `ood_test` (different distribution)
+- `train` — 90% of the in-distribution pool
+- `val` — 10% held-out from the in-distribution pool (same images appear as `ood_test` in the paired config)
+- `ood_test` — held-out portion from the out-of-distribution pool
 
 ## What is a config?
 
-A **config** fully defines a benchmark setting (e.g., country shift, state shift, biome shift).  
+A **config** fully defines a benchmark setting (e.g., country shift, biome shift, region shift).  
 When you select a config, you automatically get:
 
 - `train` split
-- `id_test` split
+- `val` split
 - `ood_test` split
 
 No filtering or custom split logic is required.
 
+## Shared test sets
+
+Configs come in pairs. For every pair, the test portion of each pool is **constant** regardless of whether it is used as `val` (in-distribution) or `ood_test` (out-of-distribution). For example:
+
+| Config | train | val | ood_test |
+|--------|-------|-----|----------|
+| `intl_train_IN__ood_US` | 90% of India | 10% of India | 10% of US |
+| `intl_train_US__ood_IN` | 90% of US | 10% of US | 10% of India |
+
+The "10% of India" set is identical in both rows, and so is the "10% of US" set.
+
+## Available configs
+
+| # | Config name | Shift type | Train on | OOD from |
+|---|-------------|-----------|----------|----------|
+| 1 | `intl_train_IN__ood_US` | Country | India | US |
+| 2 | `intl_train_US__ood_IN` | Country | US | India |
+| 3 | `biome_Rajasthan_train_WET__ood_DRY` | Biome | Rajasthan WET | Rajasthan DRY |
+| 4 | `biome_Rajasthan_train_DRY__ood_WET` | Biome | Rajasthan DRY | Rajasthan WET |
+| 5 | `biome_Karnataka_train_WET__ood_DRY` | Biome | Karnataka WET | Karnataka DRY |
+| 6 | `biome_Karnataka_train_DRY__ood_WET` | Biome | Karnataka DRY | Karnataka WET |
+| 7 | `region_train_North__ood_South` | Region | North India | South India |
+| 8 | `region_train_South__ood_North` | Region | South India | North India |
+
 ---
 
 # Export to COCO (recommended for training)
@@ -167,30 +135,30 @@ https://huggingface.co/datasets/aadityabuilds/tree-distribution-shift
 
 Example config:
 ```
-biome_India_Karnataka_train_DRY__ood_WET
+intl_train_IN__ood_US
 ```
 
 ## Step 4 — Export to COCO (one command)
 ```bash
 python tools/export_coco.py \
   --repo aadityabuilds/tree-distribution-shift \
-  --config biome_India_Karnataka_train_DRY__ood_WET \
+  --config intl_train_IN__ood_US \
   --out ./coco_out
 ```
 
 ## Output structure
 ```
 coco_out/
-└── biome_India_Karnataka_train_DRY__ood_WET/
+└── intl_train_IN__ood_US/
     ├── train/
     │   ├── images/
     │   │   └── *.tiff
     │   └── annotations/
     │       └── instances_train.json
-    ├── id_test/
+    ├── val/
     │   ├── images/
     │   └── annotations/
-    │       └── instances_id_test.json
+    │       └── instances_val.json
     └── ood_test/
         ├── images/
         └── annotations/
@@ -208,21 +176,21 @@ Use this when you want programmatic access (e.g., notebooks, statistics, custom
 pip install datasets
 ```
 
-## Load a config (gets train / id_test / ood_test)
+## Load a config (gets train / val / ood_test)
 ```python
 from datasets import load_dataset
 
 ds = load_dataset(
     "aadityabuilds/tree-distribution-shift",
-    "biome_India_Karnataka_train_DRY__ood_WET"
+    "intl_train_IN__ood_US"
 )
 
-train   = ds["train"]
-id_test = ds["id_test"]
-ood_test= ds["ood_test"]
+train    = ds["train"]
+val      = ds["val"]
+ood_test = ds["ood_test"]
 
 print(ds)
-print(len(train), len(id_test), len(ood_test))
+print(len(train), len(val), len(ood_test))
 print(train[0].keys())
 ```
 
@@ -233,12 +201,4 @@ Each row includes:
 - `image_bytes` (raw image bytes)
 - `coco_annotations` (COCO annotations for that image)
 - `coco_categories` (COCO categories)
-- metadata fields (country, state, zone, biome, etc.)
-
-
-## Density-aware OOD splits
-Some configs also include:
-- ood_same_density
-- ood_diff_density
-
-These partition the OOD split by whether  matches the IID train distribution.
+- metadata fields (country, state, zone, region, biome, etc.)

tools/build_master_parquet.py

@@ -69,6 +69,7 @@ def main():
             "country": r.country,
             "state": r.state,
             "zone": str(r.zone),
+            "region": str(getattr(r, "region", "")),
             "width": int(img_info.get("width", -1)),
             "height": int(img_info.get("height", -1)),
             "image_bytes": b,

tools/make_configs_from_metadata.py

@@ -36,7 +36,7 @@ def write_rows_to_parquet(rows_iter, out_dir: Path, split: str, rows_per_shard:
 def fetch_rows_by_filenames(master: ds.Dataset, filenames: Set[str]):
     # For 27K rows, scanning + filtering in Python is acceptable and robust.
     scanner = master.scanner(columns=[
-        "image_id","filename","country","state","zone","width","height",
+        "image_id","filename","country","state","zone","region","width","height",
         "image_bytes","coco_annotations","coco_categories"
     ])
     for batch in scanner.to_batches():
@@ -53,7 +53,7 @@ def main():
     ap.add_argument("--master_dir", required=True, help="hf_repo/data/master")
     ap.add_argument("--out_configs_dir", required=True, help="hf_repo/data/configs")
     ap.add_argument("--mode", choices=["state_shift", "zone_shift"], required=True)
-    ap.add_argument("--train_ratio", type=float, default=0.8)
+    ap.add_argument("--train_ratio", type=float, default=0.9)
     ap.add_argument("--seed", type=int, default=42)
     ap.add_argument("--rows_per_shard", type=int, default=256)
     ap.add_argument("--country", default=None, help="optional: only generate configs for this country (e.g. US)")
@@ -84,7 +84,7 @@ def main():
             if len(id_files) < args.min_images_per_region:
                 continue
 
-            train_files, id_test_files = split_list(id_files, args.train_ratio, args.seed)
+            train_files, val_files = split_list(id_files, args.train_ratio, args.seed)
 
             for (c_ood, s_ood) in keys:
                 if c_ood != c_id or s_ood == s_id:
@@ -94,7 +94,7 @@ def main():
                     continue
 
                 cfg = f"state_{c_id}_{s_id}__ood_state_{c_ood}_{s_ood}"
-                configs.append((cfg, train_files, id_test_files, set(ood_files)))
+                configs.append((cfg, train_files, val_files, set(ood_files)))
                 if len(configs) >= args.max_configs:
                     break
             if len(configs) >= args.max_configs:
@@ -110,7 +110,7 @@ def main():
             if len(id_files) < args.min_images_per_region:
                 continue
 
-            train_files, id_test_files = split_list(id_files, args.train_ratio, args.seed)
+            train_files, val_files = split_list(id_files, args.train_ratio, args.seed)
 
             for z_ood in zones:
                 if z_ood == z_id:
@@ -120,18 +120,18 @@ def main():
                     continue
 
                 cfg = f"zone_{z_id}__ood_zone_{z_ood}"
-                configs.append((cfg, train_files, id_test_files, set(ood_files)))
+                configs.append((cfg, train_files, val_files, set(ood_files)))
                 if len(configs) >= args.max_configs:
                     break
             if len(configs) >= args.max_configs:
                 break
 
-    for cfg_name, train_files, id_test_files, ood_files in configs:
+    for cfg_name, train_files, val_files, ood_files in configs:
         cfg_dir = out_configs / cfg_name
         cfg_dir.mkdir(parents=True, exist_ok=True)
 
         write_rows_to_parquet(fetch_rows_by_filenames(master, train_files), cfg_dir, "train", args.rows_per_shard)
-        write_rows_to_parquet(fetch_rows_by_filenames(master, id_test_files), cfg_dir, "id_test", args.rows_per_shard)
+        write_rows_to_parquet(fetch_rows_by_filenames(master, val_files), cfg_dir, "val", args.rows_per_shard)
         write_rows_to_parquet(fetch_rows_by_filenames(master, ood_files), cfg_dir, "ood_test", args.rows_per_shard)
 
         print(f"Wrote config: {cfg_name}")

tools/make_named_configs.py

@@ -2,7 +2,6 @@ import argparse
 from pathlib import Path
 from typing import List, Set, Tuple, Dict
 import random
-from collections import Counter
 
 import pandas as pd
 import pyarrow.dataset as ds
@@ -10,29 +9,16 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 
 
-def split_list(items: List[str], train_ratio: float, seed: int) -> Tuple[Set[str], Set[str]]:
+def split_pool(items: List[str], train_ratio: float, seed: int) -> Tuple[List[str], List[str]]:
+    """Deterministically split a pool into (train, test) portions."""
     rng = random.Random(seed)
     items = list(items)
     rng.shuffle(items)
     n_train = int(round(train_ratio * len(items)))
-    return set(items[:n_train]), set(items[n_train:])
+    return items[:n_train], items[n_train:]
 
 
-def dominant_density(meta: pd.DataFrame, filenames: Set[str]) -> str:
-    sub = meta[meta["filename"].isin(filenames)]
-    counts = Counter(sub["density_category"].astype(str).str.upper().str.strip().tolist())
-    return "HIGH" if counts.get("HIGH", 0) >= counts.get("LOW", 0) else "LOW"
-
-
-def partition_ood_by_density(meta: pd.DataFrame, ood_files: List[str], iid_dom_density: str) -> Tuple[Set[str], Set[str]]:
-    ood_df = meta[meta["filename"].isin(ood_files)].copy()
-    ood_df["density_category"] = ood_df["density_category"].astype(str).str.upper().str.strip()
-    same = set(ood_df[ood_df["density_category"] == iid_dom_density]["filename"].tolist())
-    diff = set(ood_df[ood_df["density_category"] != iid_dom_density]["filename"].tolist())
-    return same, diff
-
-
-def ensure_min(name: str, files: Set[str], min_count: int):
+def ensure_min(name: str, files, min_count: int):
     if len(files) < min_count:
         raise RuntimeError(f"{name}: only {len(files)} files (<{min_count}).")
 
@@ -44,8 +30,8 @@ def build_light_index(master_dir: str, scan_batch_size: int) -> Dict[str, Dict]:
     """
     master = ds.dataset(master_dir, format="parquet")
     cols = [
-        "image_id","filename","country","state","zone","width","height",
-        "coco_annotations","coco_categories"
+        "image_id", "filename", "country", "state", "zone", "region",
+        "width", "height", "coco_annotations", "coco_categories",
     ]
     scanner = master.scanner(columns=cols, batch_size=scan_batch_size)
 
@@ -75,38 +61,30 @@ def write_rows_to_parquet(rows_iter, out_dir: Path, split: str, rows_per_shard:
         pq.write_table(table, out_dir / f"{split}-{shard:05d}.parquet", compression="zstd")
 
 
-def rows_from_files(filenames: Set[str], light_index: Dict[str, Dict], images_root: Path):
-    """
-    Materialize full rows by reading image bytes directly from disk.
-    """
+def rows_from_files(filenames, light_index: Dict[str, Dict], images_root: Path):
+    """Materialize full rows by reading image bytes directly from disk."""
     for fn in filenames:
         base = light_index.get(fn)
         if base is None:
             continue
         img_path = images_root / fn
-        # If your images are nested, adjust this join logic accordingly.
-        img_bytes = img_path.read_bytes()
-
-        yield {
-            **base,
-            "image_bytes": img_bytes,
-        }
+        yield {**base, "image_bytes": img_path.read_bytes()}
 
 
 def main():
     ap = argparse.ArgumentParser()
-    ap.add_argument("--src_root", required=True, help="root containing metadata.csv and world_images/")
-    ap.add_argument("--master_dir", required=True, help="hf_repo/data/master (parquet without needing image_bytes scan)")
-    ap.add_argument("--out_configs_dir", required=True, help="hf_repo/data/configs")
-
-    ap.add_argument("--train_ratio", type=float, default=0.8)
+    ap.add_argument("--src_root", required=True,
+                    help="root containing metadata.csv and world_images/")
+    ap.add_argument("--master_dir", required=True,
+                    help="hf_repo/data/master (parquet shards)")
+    ap.add_argument("--out_configs_dir", required=True,
+                    help="hf_repo/data/configs")
+    ap.add_argument("--train_ratio", type=float, default=0.9)
     ap.add_argument("--seed", type=int, default=42)
     ap.add_argument("--rows_per_shard", type=int, default=16)
     ap.add_argument("--scan_batch_size", type=int, default=32)
-
-    ap.add_argument("--min_train_pool", type=int, default=200)
-    ap.add_argument("--min_ood_pool", type=int, default=200)
-    ap.add_argument("--min_ood_bucket", type=int, default=50)
+    ap.add_argument("--min_pool", type=int, default=200,
+                    help="minimum images required in a pool to create a config")
     args = ap.parse_args()
 
     src_root = Path(args.src_root)
@@ -120,20 +98,14 @@ def main():
 
     meta = pd.read_csv(meta_path)
 
-    required = {"filename", "country", "state", "zone", "biome", "density_category"}
+    required = {"filename", "country", "state", "zone", "biome", "region"}
     missing = required - set(meta.columns)
     if missing:
         raise RuntimeError(f"metadata.csv missing required columns: {sorted(missing)}")
 
     meta["biome"] = meta["biome"].astype(str).str.upper().str.strip()
-    meta["density_category"] = meta["density_category"].astype(str).str.upper().str.strip()
+    meta["region"] = meta["region"].astype(str).str.strip()
 
-    allowed = {"LOW", "HIGH"}
-    bad = set(meta["density_category"].unique()) - allowed
-    if bad:
-        raise RuntimeError(f"density_category contains unexpected values: {sorted(bad)} (expected only LOW/HIGH)")
-
-    # Build lightweight index (no image_bytes)
     print("Building lightweight master index (no image_bytes)...")
     light_index = build_light_index(args.master_dir, args.scan_batch_size)
     print(f"Indexed {len(light_index)} rows from master.")
@@ -141,65 +113,101 @@ def main():
     out_root = Path(args.out_configs_dir)
     out_root.mkdir(parents=True, exist_ok=True)
 
-    def materialize(cfg_name: str, iid_pool: List[str], ood_pool: List[str]):
-        iid_pool_set = set(iid_pool)
-        ood_pool_set = set(ood_pool)
-
-        ensure_min(cfg_name + ":iid_pool", iid_pool_set, args.min_train_pool)
-        ensure_min(cfg_name + ":ood_pool", ood_pool_set, args.min_ood_pool)
-
-        train_files, id_test_files = split_list(iid_pool, args.train_ratio, args.seed)
-        iid_dom = dominant_density(meta, train_files)
-        ood_same, ood_diff = partition_ood_by_density(meta, ood_pool, iid_dom)
-
-        ensure_min(cfg_name + ":ood_same_density", ood_same, args.min_ood_bucket)
-        ensure_min(cfg_name + ":ood_diff_density", ood_diff, args.min_ood_bucket)
-
-        cfg_dir = out_root / cfg_name
-        cfg_dir.mkdir(parents=True, exist_ok=True)
-
-        write_rows_to_parquet(rows_from_files(train_files, light_index, images_root), cfg_dir, "train", args.rows_per_shard)
-        write_rows_to_parquet(rows_from_files(id_test_files, light_index, images_root), cfg_dir, "id_test", args.rows_per_shard)
-
-        write_rows_to_parquet(rows_from_files(ood_pool_set, light_index, images_root), cfg_dir, "ood_test", args.rows_per_shard)
-        write_rows_to_parquet(rows_from_files(ood_same, light_index, images_root), cfg_dir, "ood_same_density", args.rows_per_shard)
-        write_rows_to_parquet(rows_from_files(ood_diff, light_index, images_root), cfg_dir, "ood_diff_density", args.rows_per_shard)
-
-        print(
-            f"Wrote config: {cfg_name} | IID={len(iid_pool_set)} OOD={len(ood_pool_set)} "
-            f"| IID_dom_density={iid_dom} | ood_same={len(ood_same)} ood_diff={len(ood_diff)}"
-        )
-
-    # ---- A) International ----
-    IN = "India"
-    US = "US"
-    files_in = meta[meta["country"] == IN]["filename"].tolist()
-    files_us = meta[meta["country"] == US]["filename"].tolist()
-    materialize("intl_train_IN__ood_US", iid_pool=files_in, ood_pool=files_us)
-    materialize("intl_train_US__ood_IN", iid_pool=files_us, ood_pool=files_in)
-
-    # ---- B) India State Shift ----
-    RAJ = "Rajasthan"
-    KAR = "Karnataka"
-    raj = meta[(meta["country"] == IN) & (meta["state"] == RAJ)]["filename"].tolist()
-    kar = meta[(meta["country"] == IN) & (meta["state"] == KAR)]["filename"].tolist()
-    materialize("in_state_train_Rajasthan__ood_Karnataka", iid_pool=raj, ood_pool=kar)
-    materialize("in_state_train_Karnataka__ood_Rajasthan", iid_pool=kar, ood_pool=raj)
-
-    # ---- C) Biome within (country,state) ----
-    for (country, state), df in meta.groupby(["country", "state"]):
-        wet = df[df["biome"] == "WET"]["filename"].tolist()
-        dry = df[df["biome"] == "DRY"]["filename"].tolist()
-        if len(wet) < args.min_train_pool or len(dry) < args.min_ood_pool:
-            continue
-
-        safe_country = str(country).replace(" ", "_")
-        safe_state = str(state).replace(" ", "_")
-
-        materialize(f"biome_{safe_country}_{safe_state}_train_WET__ood_DRY", iid_pool=wet, ood_pool=dry)
-        materialize(f"biome_{safe_country}_{safe_state}_train_DRY__ood_WET", iid_pool=dry, ood_pool=wet)
-
-    print(f"Done. Density-aware configs written under: {out_root}")
+    # Pre-split every pool deterministically so that the test portion is
+    # identical regardless of whether the pool is used as ID or OOD.
+    pool_splits: Dict[str, Tuple[List[str], List[str]]] = {}
+
+    def get_pool_split(pool_key: str, filenames: List[str]) -> Tuple[List[str], List[str]]:
+        if pool_key not in pool_splits:
+            pool_splits[pool_key] = split_pool(filenames, args.train_ratio, args.seed)
+        return pool_splits[pool_key]
+
+    def materialize_pair(cfg_name_a: str, cfg_name_b: str,
+                         pool_key_a: str, files_a: List[str],
+                         pool_key_b: str, files_b: List[str]):
+        """
+        Create two paired configs where the test portion of each pool is
+        constant across both configs.
+
+        Config A: train on pool A, OOD = pool B
+          train    = A's 90%
+          val      = A's 10%   (same images as ood_test in config B)
+          ood_test = B's 10%   (same images as val in config B)
+
+        Config B: symmetric.
+        """
+        train_a, test_a = get_pool_split(pool_key_a, files_a)
+        train_b, test_b = get_pool_split(pool_key_b, files_b)
+
+        ensure_min(cfg_name_a + ":train", train_a, args.min_pool)
+        ensure_min(cfg_name_a + ":val", test_a, 10)
+        ensure_min(cfg_name_a + ":ood_test", test_b, 10)
+        ensure_min(cfg_name_b + ":train", train_b, args.min_pool)
+        ensure_min(cfg_name_b + ":val", test_b, 10)
+        ensure_min(cfg_name_b + ":ood_test", test_a, 10)
+
+        for cfg_name, train_files, val_files, ood_files in [
+            (cfg_name_a, train_a, test_a, test_b),
+            (cfg_name_b, train_b, test_b, test_a),
+        ]:
+            cfg_dir = out_root / cfg_name
+            cfg_dir.mkdir(parents=True, exist_ok=True)
+
+            write_rows_to_parquet(
+                rows_from_files(train_files, light_index, images_root),
+                cfg_dir, "train", args.rows_per_shard)
+            write_rows_to_parquet(
+                rows_from_files(val_files, light_index, images_root),
+                cfg_dir, "val", args.rows_per_shard)
+            write_rows_to_parquet(
+                rows_from_files(ood_files, light_index, images_root),
+                cfg_dir, "ood_test", args.rows_per_shard)
+
+            print(f"  {cfg_name}: train={len(train_files)}  val={len(val_files)}  ood_test={len(ood_files)}")
+
+    # ---- 1) US vs India (country field) ----
+    print("\n=== Country shift: US vs India ===")
+    files_in = meta[meta["country"] == "India"]["filename"].tolist()
+    files_us = meta[meta["country"] == "US"]["filename"].tolist()
+    materialize_pair(
+        "intl_train_IN__ood_US", "intl_train_US__ood_IN",
+        "country:India", files_in,
+        "country:US", files_us,
+    )
+
+    # ---- 2) Wet vs Dry biome in Rajasthan ----
+    print("\n=== Biome shift: Rajasthan WET vs DRY ===")
+    raj = meta[(meta["country"] == "India") & (meta["state"] == "Rajasthan")]
+    raj_wet = raj[raj["biome"] == "WET"]["filename"].tolist()
+    raj_dry = raj[raj["biome"] == "DRY"]["filename"].tolist()
+    materialize_pair(
+        "biome_Rajasthan_train_WET__ood_DRY", "biome_Rajasthan_train_DRY__ood_WET",
+        "biome:Rajasthan:WET", raj_wet,
+        "biome:Rajasthan:DRY", raj_dry,
+    )
+
+    # ---- 3) Wet vs Dry biome in Karnataka ----
+    print("\n=== Biome shift: Karnataka WET vs DRY ===")
+    kar = meta[(meta["country"] == "India") & (meta["state"] == "Karnataka")]
+    kar_wet = kar[kar["biome"] == "WET"]["filename"].tolist()
+    kar_dry = kar[kar["biome"] == "DRY"]["filename"].tolist()
+    materialize_pair(
+        "biome_Karnataka_train_WET__ood_DRY", "biome_Karnataka_train_DRY__ood_WET",
+        "biome:Karnataka:WET", kar_wet,
+        "biome:Karnataka:DRY", kar_dry,
+    )
+
+    # ---- 4) North vs South India (region field) ----
+    print("\n=== Region shift: North vs South India ===")
+    files_north = meta[meta["region"] == "North"]["filename"].tolist()
+    files_south = meta[meta["region"] == "South"]["filename"].tolist()
+    materialize_pair(
+        "region_train_North__ood_South", "region_train_South__ood_North",
+        "region:North", files_north,
+        "region:South", files_south,
+    )
+
+    print(f"\nDone. All configs written under: {out_root}")
 
 
 if __name__ == "__main__":

tools/update_readme_yaml.py

@@ -8,10 +8,8 @@ CONFIGS_DIR = REPO_ROOT / "data" / "configs"
 
 SPLITS = [
     "train",
-    "id_test",
+    "val",
     "ood_test",
-    "ood_same_density",
-    "ood_diff_density",
 ]
 
 def has_files(cfg_dir: Path, split: str) -> bool:

tools/write_readme.py

@@ -22,8 +22,8 @@ def main():
         out_lines.append("  data_files:")
         out_lines.append("  - split: train")
         out_lines.append(f"    path: \"data/configs/{cfg}/train-*.parquet\"")
-        out_lines.append("  - split: id_test")
-        out_lines.append(f"    path: \"data/configs/{cfg}/id_test-*.parquet\"")
+        out_lines.append("  - split: val")
+        out_lines.append(f"    path: \"data/configs/{cfg}/val-*.parquet\"")
         out_lines.append("  - split: ood_test")
         out_lines.append(f"    path: \"data/configs/{cfg}/ood_test-*.parquet\"")
     out_lines.append("---")
@@ -37,9 +37,9 @@ def main():
     out_lines.append("This dataset is organized as **configs** (distribution shift settings).")
     out_lines.append("Each config provides 3 splits:")
     out_lines.append("")
-    out_lines.append("- `train`")
-    out_lines.append("- `id_test` (same region as train, held-out)")
-    out_lines.append("- `ood_test` (different region)")
+    out_lines.append("- `train` (90% of the in-distribution pool)")
+    out_lines.append("- `val` (10% held-out from the in-distribution pool)")
+    out_lines.append("- `ood_test` (held-out from the out-of-distribution pool)")
     out_lines.append("")
     out_lines.append("## Super simple COCO UX")
     out_lines.append("")
@@ -53,7 +53,7 @@ def main():
     out_lines.append("")
     out_lines.append(f"- `./coco_out/{example}/train/images/*.tiff`")
     out_lines.append(f"- `./coco_out/{example}/train/annotations/instances_train.json`")
-    out_lines.append(f"- `./coco_out/{example}/id_test/...`")
+    out_lines.append(f"- `./coco_out/{example}/val/...`")
     out_lines.append(f"- `./coco_out/{example}/ood_test/...`")
     out_lines.append("")
     out_lines.append("## Programmatic load (optional)")