流水线任务与评估

使用 Hugging Face 的多模态模型

James Chapman

Curriculum Manager, DataCamp

流水线 vs. 模型组件

当前做法

from transformers import BlipProcessor, BlipForConditionalGeneration
checkpoint = "Salesforce/blip-image-captioning-base"
processor = BlipProcessor.from_pretrained(checkpoint)
model = BlipForConditionalGeneration.from_pretrained(checkpoint)

Pipelines（流水线）

from transformers import pipeline
pipe = pipeline("image-to-text", model=checkpoint)

示例对比

直接调用预处理器与模型

inputs = processor(images=image, 
                   return_tensors="pt")
gen = model.generate(**inputs)
processor.decode(gen[0])

流水线

pipe(image)

[{'generated_text': 
'a man wearing a black shirt'}]

datapoint5

查找模型与任务

通过 API 为流水线查找模型：

from huggingface_hub import HfApi
model = list(api.list_models(task="text-to-image", limit=5))
pipe = pipeline("text-to-image", model[0].id)

Hugging Face 上模型主页的截图

向模型传递选项

底层为 MusicgenForConditionalGeneration

pipe = pipeline(task="text-to-audio", 
                model="facebook/musicgen-small", framework="pt")

generate_kwargs = {"temperature": 0.8, "max_new_tokens": 20}

outputs = pipe("Classic rock riff", generate_kwargs=generate_kwargs)

temperature（0-1）：控制随机性与创造性
max_new_tokens：限制生成的 token 数量

评估流水线性能

Accuracy（准确率）：正确分类的总体占比
Precision（精确率）：预测为该类时有多常正确
Recall（召回率）：实际该类被正确识别的比例
F1 分数：综合精确率与召回率

from evaluate import evaluator

task_evaluator = evaluator("image-classification")

metrics_dict = {
    "precision": "precision",
    "recall": "recall",
    "f1": "f1",
}

label_map = pipe.model.config.label2id

评估流水线性能

eval_results = task_evaluator.compute(
  model_or_pipeline=pipe,

  data=dataset,

  metric=evaluate.combine(metrics_dict),

  label_mapping=label_map)

print(eval_results)

{'precision': 0.999001923076923, 
'recall': 0.999, 
'f1': 0.9989999609405906, ...}

pipe = pipeline(task="image-classification",
model="ideepankarsharma2003/AI_ImageClassi
fication_MidjourneyV6_SDXL"
)
dataset = load_dataset("ideepankarsharma2003/
Midjourney_v6_Classification_small_shuffled")

来自数据集中一张狗在水下的 AI 生成图像示例

Vamos praticar!

使用 Hugging Face 的多模态模型