图像标题的可查询索引
默认情况下,加载程序使用预训练的 Salesforce BLIP 图像字幕模型。
本文介绍如何使用ImageCaptionLoader教程生成一个可查询的图像标题索引。
from langchain.document_loaders import ImageCaptionLoader
准备来自维基媒体的图像 url 列表 #
list_image_urls = [
'https://upload.wikimedia.org/wikipedia/commons/thumb/5/5a/Hyla_japonica_sep01.jpg/260px-Hyla_japonica_sep01.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/7/71/Tibur%C3%B3n_azul_%28Prionace_glauca%29%2C_canal_Fayal-Pico%2C_islas_Azores%2C_Portugal%2C_2020-07-27%2C_DD_14.jpg/270px-Tibur%C3%B3n_azul_%28Prionace_glauca%29%2C_canal_Fayal-Pico%2C_islas_Azores%2C_Portugal%2C_2020-07-27%2C_DD_14.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/2/21/Thure_de_Thulstrup_-_Battle_of_Shiloh.jpg/251px-Thure_de_Thulstrup_-_Battle_of_Shiloh.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/2/21/Passion_fruits_-_whole_and_halved.jpg/270px-Passion_fruits_-_whole_and_halved.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/5/5e/Messier83_-_Heic1403a.jpg/277px-Messier83_-_Heic1403a.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/b/b6/2022-01-22_Men%27s_World_Cup_at_2021-22_St._Moritz%E2%80%93Celerina_Luge_World_Cup_and_European_Championships_by_Sandro_Halank%E2%80%93257.jpg/288px-2022-01-22_Men%27s_World_Cup_at_2021-22_St._Moritz%E2%80%93Celerina_Luge_World_Cup_and_European_Championships_by_Sandro_Halank%E2%80%93257.jpg',
'https://upload.wikimedia.org/wikipedia/commons/thumb/9/99/Wiesen_Pippau_%28Crepis_biennis%29-20220624-RM-123950.jpg/224px-Wiesen_Pippau_%28Crepis_biennis%29-20220624-RM-123950.jpg',
]
创建加载器 Create the loader #
loader = ImageCaptionLoader(path_images=list_image_urls)
list_docs = loader.load()
list_docs
/Users/saitosean/dev/langchain/.venv/lib/python3.10/site-packages/transformers/generation/utils.py:1313: UserWarning: Using `max_length`'s default (20) to control the generation length. This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we recommend using `max_new_tokens` to control the maximum length of the generation.
warnings.warn()
[Document(page_content='an image of a frog on a flower [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/5/5a/Hyla_japonica_sep01.jpg/260px-Hyla_japonica_sep01.jpg'}),
Document(page_content='an image of a shark swimming in the ocean [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/7/71/Tibur%C3%B3n_azul_%28Prionace_glauca%29%2C_canal_Fayal-Pico%2C_islas_Azores%2C_Portugal%2C_2020-07-27%2C_DD_14.jpg/270px-Tibur%C3%B3n_azul_%28Prionace_glauca%29%2C_canal_Fayal-Pico%2C_islas_Azores%2C_Portugal%2C_2020-07-27%2C_DD_14.jpg'}),
Document(page_content='an image of a painting of a battle scene [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/2/21/Thure_de_Thulstrup_-_Battle_of_Shiloh.jpg/251px-Thure_de_Thulstrup_-_Battle_of_Shiloh.jpg'}),
Document(page_content='an image of a passion fruit and a half cut passion [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/2/21/Passion_fruits_-_whole_and_halved.jpg/270px-Passion_fruits_-_whole_and_halved.jpg'}),
Document(page_content='an image of the spiral galaxy [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/5/5e/Messier83_-_Heic1403a.jpg/277px-Messier83_-_Heic1403a.jpg'}),
Document(page_content='an image of a man on skis in the snow [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/b/b6/2022-01-22_Men%27s_World_Cup_at_2021-22_St._Moritz%E2%80%93Celerina_Luge_World_Cup_and_European_Championships_by_Sandro_Halank%E2%80%93257.jpg/288px-2022-01-22_Men%27s_World_Cup_at_2021-22_St._Moritz%E2%80%93Celerina_Luge_World_Cup_and_European_Championships_by_Sandro_Halank%E2%80%93257.jpg'}),
Document(page_content='an image of a flower in the dark [SEP]', metadata={'image_path': 'https://upload.wikimedia.org/wikipedia/commons/thumb/9/99/Wiesen_Pippau_%28Crepis_biennis%29-20220624-RM-123950.jpg/224px-Wiesen_Pippau_%28Crepis_biennis%29-20220624-RM-123950.jpg'})]
from PIL import Image
import requests
Image.open(requests.get(list_image_urls[0], stream=True).raw).convert('RGB')
创建索引 Create the index #
from langchain.indexes import VectorstoreIndexCreator
index = VectorstoreIndexCreator().from_loaders([loader])
/Users/saitosean/dev/langchain/.venv/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install
from .autonotebook import tqdm as notebook_tqdm
/Users/saitosean/dev/langchain/.venv/lib/python3.10/site-packages/transformers/generation/utils.py:1313: UserWarning: Using `max_length`'s default (20) to control the generation length. This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we recommend using `max_new_tokens` to control the maximum length of the generation.
warnings.warn()
Using embedded DuckDB without persistence: data will be transient
查询 Query #
query = "What's the painting about?"
index.query(query)
' The painting is about a battle scene.'
query = "What kind of images are there?"
index.query(query)
' There are images of a spiral galaxy, a painting of a battle scene, a flower in the dark, and a frog on a flower.'