零样本视频分类

使用 Hugging Face 的多模态模型

James Chapman

Curriculum Manager, DataCamp

零样本视频分类

业务场景：

竞争对手的广告带来更多销量
他们的广告更积极吗？

¹ https://repository.duke.edu/dc/adviews/dmbb37204

CLAP

Contrastive Language Audio Pretraining
类似于训练 CLIP 的方法
- 独立的文本编码器与音频编码器
63.3 万 对音频 + 描述
对比式预训练对齐音频与文本嵌入

CLAP 示意图：匹配嵌入

¹ https://arxiv.org/html/2211.06687v4

方法：多模态 ZSL

使用 CLAP 的"零样本学习"

晚期融合方法放大图第 1 部分：CLAP

¹ https://arxiv.org/html/2310.02298v3

方法：多模态 ZSL

使用 CLAP 的"零样本学习"

+ 使用 CLIP 的零样本学习

通用文本类别
合并 CLIP 与 CLAP 的概率

晚期融合方法放大图第 2 部分：CLIP+CLAP

视频与音频

使用 moviepy 分离 .mp4 的音频与视频流：

from moviepy.editor import VideoFileClip
from moviepy.video.io.ffmpeg_tools import ffmpeg_extract_subclip

ffmpeg_extract_subclip("advert.mp4", 0, 5, "advert_5s.mp4")


video = VideoFileClip("advert_5s.mp4")

audio = video.audio

audio.write_audiofile("advert_5s.mp3")

准备音频与视频

from decord import VideoReader
from PIL import Image
video_reader = VideoReader(video_path)
video = video_reader.get_batch(range(20)).asnumpy()
video = video[:, :, :, ::-1]
video = [Image.fromarray(frame) for frame in video]


from datasets import Dataset, Audio
audio_dataset = Dataset.from_dict({"audio": [audio_path]}).cast_column("audio", Audio())
audio_sample = audio_dataset[0]["audio"]["array"]

视频预测

emotions = ["joy", "fear", "anger", "sadness", "disgust", "surprise", "neutral"]


image_class = pipeline(model="openai/clip-vit-large-patch14", 
                       task="zero-shot-image-classification")


predictions = image_classifier(video, candidate_labels=emotions)
scores = [
    {l['label']: l['score'] for l in prediction}
    for prediction in predictions
]


avg_image_scores = {emotion: sum([s[emotion] for s in scores])/len(scores) for emotion in emotions}
print(f"Average scores: {avg_image_scores}")

Average scores: {'joy': 0.10063326267991216, 'fear': 0.0868348691612482, ...}

音频预测与融合

audio_class = pipeline(model="laion/clap-htsat-unfused", 
                       task="zero-shot-audio-classification")


audio_scores = audio_class(audio_sample, candidate_labels=emotions)

audio_scores = {l['label']: l['score'] for l in audio_scores}

multimodal_scores = {emotion: (avg_image_scores[emotion] + audio_scores[emotion])/2
                     for emotion in emotions}
print(f"Multimodal scores: {multimodal_scores}")

Multimodal scores: {'joy': 0.3109628591220826, 'fear': 0.09013736313208938, 
'anger': 0.011454355076421053, 'sadness': 0.06018101833760738, 'disgust': 0.07207315033301712, 
'surprise': 0.252118631079793, 'neutral': 0.2030726027674973}

¡Vamos a practicar!

使用 Hugging Face 的多模态模型