from pathlib import Path import glob import json import csv import tqdm import hashlib from convert_utils.base_convertor import BaseConvertor from time_utils.time_utils import format_time, time_to_seconds, seconds_to_time class PrepareBenchmarkTemplate(BaseConvertor): VIDEO_ROOT = "/home/dyvm6xra/dyvm6xrauser36/dyvm6xrauser36/huankguan2/Project/OmniEval/benchmarks/streamingbench/StreamingBench/videos" def read_data(self, data_file: str) -> list: with open(data_file, mode="r", encoding="utf-8") as f: reader = csv.DictReader(f) data = list(reader) return data def get_video_uuid(self, item) -> str: video_uuid = hashlib.sha256( self.get_video_path(item).encode("utf-8") ).hexdigest()[:8] return video_uuid def get_video_path(self, item) -> str: task_name, sample_name, sample_id, _ = item["question_id"].split("_") video_path = ( f"{self.VIDEO_ROOT}/{task_name}/{sample_name}_{sample_id}/video.mp4" ) return video_path def get_stream_addr(self, item) -> str: return f"/home/dyvm6xra/dyvm6xrauser36/dyvm6xrauser36/huankguan2/Project/OmniEval/benchmarks/streamingbench/stream_chunks/{self.get_video_uuid(item)}" def build_sqa(self, item) -> list: """event_id: int, timestamp: str""" question_t = format_time(item["time_stamp"]) response_t = question_t event_ts = [question_t, response_t] def auto_fill_label(options): options = eval(options) for i in range(len(options)): p = str(chr(ord('A') + i)) + ". " if not options[i].startswith(p): options[i] = p + options[i] return str(options) sqa = [] for i, ts in enumerate(self.deduplicate_and_sorting(event_ts)): event = { "event_id": i + 1, "timestamp": ts, "question_id": item["question_id"], } if ts == question_t: event = event | { "task_type": item["task_type"], "question": item["question"], "options": auto_fill_label(item["options"]), } if ts == response_t: event = event | {"response": item["answer"]} sqa.append(event) return sqa inputs = [ ( "contextual_understanding.jsonl", ["StreamingBench/annotations/Contextual_Understanding.csv"], ), ( "omni_source_understanding.jsonl", ["StreamingBench/annotations/Omni_Source_Understanding.csv"], ), ( "real_time_visual_understanding.jsonl", ["StreamingBench/annotations/Real_Time_Visual_Understanding.csv"], ), ] for outfile, samples in inputs: converter = PrepareBenchmarkTemplate(outfile=outfile, start_id=1) converter.clear() for x in tqdm.tqdm(samples): converter.convert(x) print("done:", outfile)