| from __future__ import annotations |
| import argparse,csv,json |
| from collections import Counter |
| from pathlib import Path |
| TASKS={1:"Image Description",2:"Basic Relational Reasoning",3:"Comparative Reasoning",4:"Compound Reasoning",5:"Extrapolative/Counterfactual Reasoning"};FIELDS=["dataset_version","difficulty_score","total_cube_count","max_height","color_cluster_count","has_ambiguous_visual_floater","vertical_axis","seed"] |
| def main(): |
| p=argparse.ArgumentParser();p.add_argument("annotations",type=Path);a=p.parse_args();rows=[];images=0;bad=[] |
| for line in a.annotations.read_text(encoding="utf-8").splitlines(): |
| if not line:continue |
| r=json.loads(line);images+=1 |
| if len(r["questions"])!=5:bad.append(r["id"]) |
| meta=json.dumps({k:r[k] for k in FIELDS},sort_keys=True,separators=(",",":")) |
| for q in r["questions"]:rows.append({"question_id":q["question_id"],"task":TASKS[q["difficulty_level"]],"image":Path(r["image_path"]).name,"prompt":q["question_text"],"groundtruth":str(q["ground_truth"]),"answer_format":json.dumps(q["answer_format"]) if isinstance(q["answer_format"],dict) else str(q["answer_format"]),"metadata":meta}) |
| out=a.annotations.parent;cols=["question_id","task","image","prompt","groundtruth","answer_format","metadata"] |
| with (out/"dataset_final.csv").open("w",encoding="utf-8-sig",newline="") as f:w=csv.DictWriter(f,fieldnames=cols,lineterminator="\n");w.writeheader();w.writerows(rows) |
| with (out/"dataset_final.jsonl").open("w",encoding="utf-8",newline="\n") as f: |
| for r in rows:f.write(json.dumps(r,separators=(",",":"))+"\n") |
| with (out/"question_set.csv").open("w",encoding="utf-8-sig",newline="") as f:w=csv.DictWriter(f,fieldnames=["question_id","task","image","prompt"],lineterminator="\n");w.writeheader();w.writerows({k:r[k] for k in w.fieldnames} for r in rows) |
| with (out/"answer_key.csv").open("w",encoding="utf-8-sig",newline="") as f:w=csv.DictWriter(f,fieldnames=["question_id","task","image","prompt","groundtruth","answer_format"],lineterminator="\n");w.writeheader();w.writerows({k:r[k] for k in w.fieldnames} for r in rows) |
| counts=Counter(r["task"] for r in rows);print(f"Images: {images}\nRows: {len(rows)}\nExpected: {5*images}\nImages without 5 questions: {len(bad)}");[print(f"{x}: {counts[x]}") for x in TASKS.values()];assert len(rows)==5*images and not bad;print("Sanity check: PASS") |
| if __name__=="__main__":main() |
|
|
