MatchLab's picture
Upload folder using huggingface_hub
6f89716 verified
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import json
import math
import argparse
import random
from pathlib import Path
from typing import Any, Dict, List, Tuple, DefaultDict, Optional
from collections import defaultdict
# ---------------- Mapping helpers ----------------
def load_answer_maps(path: str) -> Tuple[Dict[str, int], Dict[int, str]]:
"""
Loads the 2-entry list:
[ {text->id}, { "0": "text", "1": "text", ... } ]
Returns:
text2id (normalized text -> id), id2text (int id -> canonical text)
"""
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
if not isinstance(data, list) or len(data) < 2:
raise ValueError("answer_dict file must be a list of two dicts: [text2id, id2text].")
raw_text2id, raw_id2text = data[0], data[1]
if not isinstance(raw_text2id, dict) or not isinstance(raw_id2text, dict):
raise ValueError("Both elements in answer_dict must be dicts.")
# Normalize text keys to a stable form for lookup
def norm_text(s: str) -> str:
return " ".join(s.strip().lower().split())
text2id = {norm_text(k): int(v) for k, v in raw_text2id.items()}
id2text = {int(k): str(v) for k, v in raw_id2text.items()}
return text2id, id2text
def normalize_answer_text(s: str) -> str:
"""Canonicalize spacing/case for consistent lookup; keep original wording style simple."""
return " ".join(s.strip().lower().split())
# ---------------- Data helpers ----------------
def pick_situation(d: Dict[str, Any]) -> str:
return d.get("situation_text") or d.get("situation_multimodal") or ""
def get_location(d: Dict[str, Any]) -> List[float]:
md = d.get("meta_data") or {}
sp = md.get("start_point")
loc = sp if isinstance(sp, (list, tuple)) and len(sp) >= 3 else d.get("location", [0, 0, 0])
return [float(loc[0]), float(loc[1]), float(loc[2])]
def yaw_to_vec3(yaw: float) -> List[float]:
return [math.cos(yaw), math.sin(yaw), 0.0]
def quat_to_yaw(q: List[float]) -> float:
if len(q) != 4:
return 0.0
# try (x,y,z,w)
x, y, z, w = map(float, q)
try:
siny_cosp = 2.0 * (w * z + x * y)
cosy_cosp = 1.0 - 2.0 * (y * y + z * z)
return math.atan2(siny_cosp, cosy_cosp)
except Exception:
pass
# fallback (w,x,y,z)
w, x, y, z = map(float, q)
try:
siny_cosp = 2.0 * (w * z + x * y)
cosy_cosp = 1.0 - 2.0 * (y * y + z * z)
return math.atan2(siny_cosp, cosy_cosp)
except Exception:
return 0.0
def get_orientation_vec(d: Dict[str, Any]) -> List[float]:
md = d.get("meta_data") or {}
if isinstance(md.get("start_ori"), (int, float)):
return yaw_to_vec3(float(md["start_ori"]))
q = d.get("orientation")
if isinstance(q, (list, tuple)) and len(q) == 4:
return yaw_to_vec3(quat_to_yaw(list(map(float, q))))
if isinstance(q, (list, tuple)) and len(q) >= 3:
return [float(q[0]), float(q[1]), float(q[2])]
return [1.0, 0.0, 0.0]
# ---------------- Answer extraction (eight_direction) ----------------
def eight_direction_answer(item: Dict[str, Any], text2id: Dict[str, int], id2text: Dict[int, str]) -> Optional[str]:
"""
Returns a SINGLE canonical text answer for eight_direction (e.g., 'turn left'),
resolving from either [code, 'text'] or just code or text, using id<->text maps.
"""
act = item.get("action") or item.get("meta_data", {}).get("action_type") or {}
v = act.get("eight_direction")
# Case A: list/tuple like [1, "turn left"] or [1]
if isinstance(v, (list, tuple)) and len(v) > 0:
# Prefer explicit string label when present
if len(v) >= 2 and isinstance(v[1], str) and v[1].strip():
return id2text.get(text2id.get(normalize_answer_text(v[1]), -999), v[1].strip())
# Else map integer code to text
if isinstance(v[0], int):
return id2text.get(v[0], str(v[0]))
# Case B: plain int code
if isinstance(v, int):
return id2text.get(v, str(v))
# Case C: plain string
if isinstance(v, str) and v.strip():
norm = normalize_answer_text(v)
# If mappable, return canonical text from id2text; else return original cleaned text
if norm in text2id:
return id2text.get(text2id[norm], v.strip())
return v.strip()
return None
# ---------------- Conversion ----------------
def convert_to_flat(nested: Dict[str, Any], text2id: Dict[str, int], id2text: Dict[int, str]) -> List[Dict[str, Any]]:
"""
Convert nested JSON to flat records, keeping ONLY samples with eight_direction,
and normalizing the answer to the canonical text form.
"""
out = []
for scan_id, steps in nested.items():
if not isinstance(steps, dict):
continue
for idx_key, item in steps.items():
if not isinstance(item, dict):
continue
answer_text = eight_direction_answer(item, text2id, id2text)
if not answer_text:
continue # strictly keep eight_direction examples
situation = pick_situation(item)
location = get_location(item)
orientation = get_orientation_vec(item)
try:
index = int(idx_key)
except Exception:
index = len(out)
rec = {
"question": "What action should I take next step?",
"situation_text": situation,
"answers": [answer_text], # normalized canonical text (e.g., "turn left")
"scan_id": item.get("scan_id") or scan_id,
"location": location, # [x, y, z]
"interaction": item.get("interaction"),
"orientation": orientation, # [dx, dy, 0]
"type": "navigation",
"index": index,
"question_id": f"{item.get('scan_id') or scan_id}_{index}",
}
out.append(rec)
return out
# ---------------- Annotations ----------------
def records_to_annotations(records: List[Dict[str, Any]], text2id: Dict[str, int]) -> Dict[str, Any]:
"""
Map flat records into 'annotations' format.
- Sets answer_id using the provided text2id mapping.
- If text not found in mapping, answer_id = -1.
"""
ann_list: List[Dict[str, Any]] = []
for r in records:
dx, dy, _ = r["orientation"]
x, y, z = r["location"]
answers = r.get("answers") or []
answers_obj = []
for a in answers:
norm = normalize_answer_text(str(a))
ans_id = text2id.get(norm, -1)
answers_obj.append(
{
"answer": str(a),
"answer_confidence": "yes",
"answer_id": ans_id
}
)
ann = {
"scan_id": r["scan_id"],
"question_type": r.get("type", "navigation"),
"answer_type": "other",
"question_id": r["question_id"],
"answers": answers_obj,
"rotation": {"_x": dx, "_y": dy, "_z": 0.0, "_w": 0.0},
"position": {"x": x, "y": y, "z": z},
}
ann_list.append(ann)
return {"annotations": ann_list}
# ---------------- Split by scene ----------------
def group_by_scene(records: List[Dict[str, Any]]) -> DefaultDict[str, List[Dict[str, Any]]]:
buckets: DefaultDict[str, List[Dict[str, Any]]] = defaultdict(list)
for r in records:
buckets[r["scan_id"]].append(r)
return buckets
def split_scenes(scene_ids: List[str], val_ratio: float, test_ratio: float, seed: int) -> Tuple[List[str], List[str], List[str]]:
rnd = random.Random(seed)
ids = list(scene_ids)
rnd.shuffle(ids)
n = len(ids)
n_val = int(round(n * val_ratio))
n_test = int(round(n * test_ratio))
n_val = min(n_val, n)
n_test = min(n_test, max(0, n - n_val))
train = ids[n_val + n_test :]
val = ids[:n_val]
test = ids[n_val:n_val + n_test]
return train, val, test
def flatten_from_ids(buckets: Dict[str, List[Dict[str, Any]]], ids: List[str]) -> List[Dict[str, Any]]:
out: List[Dict[str, Any]] = []
for sid in ids:
out.extend(buckets.get(sid, []))
return out
# ---------------- CLI ----------------
def main():
parser = argparse.ArgumentParser(description="Convert & split dataset (8:1:1 by scene) and write annotation files with answer_id from answer_dict.")
parser.add_argument("--input", type=str, default="msnn/msnn_scannet.json", help="Path to input JSON file")
parser.add_argument("--output_dir", type=str, default=None, help="Directory to write outputs (default: input's directory)")
parser.add_argument("--answer_dict", type=str, default="msnn/answer_dict.json", help="Path to the answer mapping JSON (list of two dicts)")
parser.add_argument("--seed", type=int, default=42, help="Random seed for scene split")
parser.add_argument("--val_ratio", type=float, default=0.1, help="Validation ratio (by scene)")
parser.add_argument("--test_ratio", type=float, default=0.1, help="Test ratio (by scene)")
args = parser.parse_args()
# Load mappings
text2id, id2text = load_answer_maps(args.answer_dict)
# Load nested data
with open(args.input, "r", encoding="utf-8") as f:
nested = json.load(f)
# Convert and group
flat = convert_to_flat(nested, text2id, id2text)
buckets = group_by_scene(flat)
scene_ids = sorted(buckets.keys())
train_ids, val_ids, test_ids = split_scenes(scene_ids, args.val_ratio, args.test_ratio, args.seed)
train_recs = flatten_from_ids(buckets, train_ids)
val_recs = flatten_from_ids(buckets, val_ids)
test_recs = flatten_from_ids(buckets, test_ids)
# Output dir
out_dir = Path(args.output_dir) if args.output_dir else Path(args.input).parent
out_dir.mkdir(parents=True, exist_ok=True)
# Dataset filenames
p_train = out_dir / "msnn_train_eight_direction.json"
p_val = out_dir / "msnn_val_eight_direction.json"
p_test = out_dir / "msnn_test_eight_direction.json"
# Annotation filenames
pa_train = out_dir / "msnn_train_eight_direction_annotations.json"
pa_val = out_dir / "msnn_val_eight_direction_annotations.json"
pa_test = out_dir / "msnn_test_eight_direction_annotations.json"
# Write datasets
with open(p_train, "w", encoding="utf-8") as f: json.dump(train_recs, f, ensure_ascii=False, indent=4)
with open(p_val, "w", encoding="utf-8") as f: json.dump(val_recs, f, ensure_ascii=False, indent=4)
with open(p_test, "w", encoding="utf-8") as f: json.dump(test_recs, f, ensure_ascii=False, indent=4)
# Build & write annotations (with answer_id from mapping)
with open(pa_train, "w", encoding="utf-8") as f: json.dump(records_to_annotations(train_recs, text2id), f, ensure_ascii=False, indent=4)
with open(pa_val, "w", encoding="utf-8") as f: json.dump(records_to_annotations(val_recs, text2id), f, ensure_ascii=False, indent=4)
with open(pa_test, "w", encoding="utf-8") as f: json.dump(records_to_annotations(test_recs, text2id), f, ensure_ascii=False, indent=4)
print(f"Scenes: total={len(scene_ids)} | train={len(train_ids)} | val={len(val_ids)} | test={len(test_ids)}")
print(f"Samples: train={len(train_recs)} | val={len(val_recs)} | test={len(test_recs)}")
print("Wrote:")
print(f" {p_train}")
print(f" {p_val}")
print(f" {p_test}")
print(f" {pa_train}")
print(f" {pa_val}")
print(f" {pa_test}")
if __name__ == "__main__":
main()