japanese clip

日本紙

該存儲庫包括Rinna Co.，Ltd。的日本剪輯（對比語言圖像預訓練）變體的代碼。

目錄
消息
預驗證的模型
用法
引用
執照

消息

2022年7月

v0.2.0發行了！

• Clip和Cloob車型都升級了！現在， rinna/japanese-cloob-vit-b-16達到54.64。
• 為零彈出的成像網分類，發布了我們的日語提示模板和示例代碼（請參見scripts/example.py ）。這些模板根據Openai 80模板清洗了日語。
• 改變了引用

預驗證的模型

模型名稱	top1*	top5*
rinna/japesentle-clloob-vit-b-16	54.64	72.86
rinna/japesent-clip-vit-b-16	50.69	72.35
Sonoisa/clip-vit-B-32-JEAMENES-V1	38.88	60.71
多語言夾	14.36	27.28

*零擊成像網驗證設置TOP-K精度。

用法

1. 安裝軟件包

$ pip install git+https://github.com/rinnakk/japanese-clip.git

2. 跑步

 from PIL import Image
import torch
import japanese_clip as ja_clip

device = "cuda" if torch . cuda . is_available () else "cpu"
# ja_clip.available_models()
# ['rinna/japanese-clip-vit-b-16', 'rinna/japanese-cloob-vit-b-16']
# If you want v0.1.0 models, set `revision='v0.1.0'`
model , preprocess = ja_clip . load ( "rinna/japanese-clip-vit-b-16" , cache_dir = "/tmp/japanese_clip" , device = device )
tokenizer = ja_clip . load_tokenizer ()

image = preprocess ( Image . open ( "./data/dog.jpeg" )). unsqueeze ( 0 ). to ( device )
encodings = ja_clip . tokenize (
    texts = [ "犬" , "猫" , "象" ],
    max_seq_len = 77 ,
    device = device ,
    tokenizer = tokenizer , # this is optional. if you don't pass, load tokenizer each time
)

with torch . no_grad ():
    image_features = model . get_image_features ( image )
    text_features = model . get_text_features ( ** encodings )
    
    text_probs = ( 100.0 * image_features @ text_features . T ). softmax ( dim = - 1 )

print ( "Label probs:" , text_probs )  # prints: [[1.0, 0.0, 0.0]] 

引用

引用這個存儲庫：

@inproceedings{japanese-clip,
  author = {シーン 誠, 趙 天雨, 沢田 慶},
  title = {日本語における言語画像事前学習モデルの構築と公開},
  booktitle= {The 25th Meeting on Image Recognition and Understanding},
  year = 2022,
  month = July,
}

執照

Apache 2.0許可證