|
|
|
|
|
from pathlib import Path |
|
|
import sys |
|
|
sys.path.append(str(Path.cwd() / "annotation")) |
|
|
|
|
|
import json |
|
|
import os |
|
|
from typing import List, Union, Dict, Any, Callable, Optional |
|
|
from concurrent.futures import ThreadPoolExecutor, as_completed |
|
|
from datatypes import VideoAnnotation, Metadata |
|
|
from utils import get_optimal_workers, extract_label, convert_to_linux_path |
|
|
|
|
|
|
|
|
def annotate_video( |
|
|
file_path: str, |
|
|
label: str, |
|
|
video_filter: Callable[[str, Any], bool] = lambda path: True, |
|
|
**kwargs |
|
|
) -> VideoAnnotation: |
|
|
if not video_filter(file_path, **kwargs): |
|
|
return None |
|
|
|
|
|
json_content: VideoAnnotation = { |
|
|
'video': convert_to_linux_path(file_path), |
|
|
'label': label, |
|
|
'conversations': [ |
|
|
{ |
|
|
'from': 'human', |
|
|
'value': '<image>\nThis video is a Youtube video on one of many categories such as Education, Film & Animation, Comedy, Entertainment, Music, Howto & Style, and People & Blogs, etc. The engagement rate defined for each such video is based on the number of potential likes and dislikes only when published on Youtube. The higher number of likes and lower number of dislikes, the more engaged the video is. The final prediction label is either 0 (not engaged), 1 (neutral), or 2 (engaged). Please predict one of the three labels for this video, based on its contents only.' |
|
|
}, |
|
|
{ |
|
|
'from': 'gpt', |
|
|
'value': f'The engagement label of the video is {label}.' |
|
|
} |
|
|
] |
|
|
} |
|
|
return json_content |
|
|
|
|
|
|
|
|
|
|
|
def dump_json( |
|
|
metadata: Metadata, |
|
|
video_filter: Callable[[str, Any], bool] = lambda path: True, |
|
|
**kwargs |
|
|
) -> List[VideoAnnotation]: |
|
|
print(f'Annotating {len(metadata)} videos...') |
|
|
json_contents: List[VideoAnnotation] = [] |
|
|
|
|
|
with ThreadPoolExecutor(max_workers=get_optimal_workers()) as executor: |
|
|
futures = [] |
|
|
for (file_path, label) in metadata: |
|
|
futures.append(executor.submit(annotate_video, file_path, label, video_filter=video_filter, **kwargs)) |
|
|
|
|
|
for future in as_completed(futures): |
|
|
result = future.result() |
|
|
if result: |
|
|
json_contents.append(result) |
|
|
|
|
|
return json_contents |
