# coding=utf-8
# author: xusong
# time: 2022/8/23 16:06
"""
## TODO:
- i18 国际化 https://blog.csdn.net/qq_26212731/article/details/78457198 request.header中也有language
- iter_vocab 的 warmup
- 开关
- add_special_token 开关
- theme 开关 light/dark
- token_id/tokens/bytes 开关
- 中文字词统计,是否要包括 _ G 等字符
- 评测
- OOV评测
- 通过 javascript 添加 hover_text
- 英文 utf-8编码
- 词典支持下载,借用image下载的标签,
- baichuan的单字数量怎么两万多个?
- qwen: ValueError: Unclosed image token
- 路径修改为全path meta-llama/Llama-2-13b-hf
plots
table
## related demo
- [](http://text-processing.com/demo/tokenize/)
- [gpt-tokenizer](https://gpt-tokenizer.dev/)
- [llama-tokenizer-js](https://belladoreai.github.io/llama-tokenizer-js/example-demo/build/)
- [](https://huggingface.co/spaces/Xenova/the-tokenizer-playground)
## 可视化
[ The, 2, QUICK, Brown, Foxes, jumped, over, the, lazy, dog's, bone ]
"""
import gradio as gr
from vocab import all_tokenizers
from util import *
from examples import example_fn, example_types
get_window_url_params = """
function(url_params) {
const params = new URLSearchParams(window.location.search);
url_params = JSON.stringify(Object.fromEntries(params));
return url_params;
}
"""
with gr.Blocks(css="css/style.css", title="Tokenizer Arena") as demo:
gr.HTML("""
Tokenizer Arena ⚔️
""")
# links: https://www.coderstool.com/utf8-encoding-decoding
# 功能:输入文本,进行分词
# 分词器:常见的分词器有集中,
# 背景:方便分词、看词粒度、对比
with gr.Row():
gr.Markdown("## Input Text")
dropdown_examples = gr.Dropdown(
example_types,
type="index",
show_label=False,
container=False,
scale=0,
elem_classes="example-style"
)
user_input = gr.Textbox(
# value=default_user_input,
label="Input Text",
lines=5,
show_label=False,
)
gr.Markdown("## Tokenization")
# compress rate setting
with gr.Accordion("Compress Rate Setting", open=True):
gr.Markdown("Please select corpus and unit of compress rate, get more details at [github](https://github.com/xu-song/tokenizer-arena/). ")
with gr.Row():
compress_rate_corpus = gr.CheckboxGroup(
["cc100-en", "cc100-zh-Hans", "cc100-es"], # , "code"
value=["cc100-en", "cc100-zh-Hans"],
label="corpus",
# info=""
)
compress_rate_unit = gr.Radio(
["b_tokens/g_bytes", "g_bytes/b_tokens", "t_tokens/t_bytes", "t_bytes/t_tokens", "n_chars/n_tokens"],
value="b_tokens/g_bytes",
label="unit",
)
# TODO: Token Setting
# with gr.Accordion("Token Filter Setting", open=False):
# gr.Markdown(
# "Get total number of tokens which contain the following character)")
# gr.Radio(
# ["zh-Hans", "", "number", "space"],
# value="zh",
# )
with gr.Row():
with gr.Column(scale=6):
with gr.Group():
tokenizer_type_1 = gr.Dropdown(
all_tokenizers,
label="Tokenizer 1",
)
with gr.Group():
"""
"""
with gr.Row():
stats_vocab_size_1 = gr.TextArea(
label="Vocab Size",
lines=1,
elem_classes="statistics"
)
stats_zh_token_size_1 = gr.TextArea(
label="ZH char/word",
lines=1,
elem_classes="statistics",
visible=False
)
stats_compress_rate_1 = gr.TextArea(
label="Compress Rate",
lines=1,
elem_classes="statistics"
)
stats_overlap_token_size_1 = gr.TextArea(
# value=default_stats_overlap_token_size,
label="Overlap Tokens",
lines=1,
elem_classes="statistics"
)
# stats_3 = gr.TextArea(
# label="Compress Rate",
# lines=1,
# elem_classes="statistics"
# )
# https://www.onlinewebfonts.com/icon/418591
gr.Image("images/VS.svg", scale=1, show_label=False,
show_download_button=False, container=False,
show_share_button=False)
with gr.Column(scale=6):
with gr.Group():
tokenizer_type_2 = gr.Dropdown(
all_tokenizers,
label="Tokenizer 2",
)
with gr.Group():
with gr.Row():
stats_vocab_size_2 = gr.TextArea(
label="VocabSize",
lines=1,
elem_classes="statistics"
)
stats_zh_token_size_2 = gr.TextArea(
label="ZH char/word", # 中文字/词
lines=1,
elem_classes="statistics",
visible=False
)
stats_compress_rate_2 = gr.TextArea(
label="Compress Rate",
lines=1,
elem_classes="statistics"
)
stats_filtered_token_2 = gr.TextArea(
label="filtered tokens",
lines=1,
elem_classes="statistics",
visible=False
)
stats_overlap_token_size_2 = gr.TextArea(
label="Overlap Tokens",
lines=1,
elem_classes="statistics"
)
# TODO: 图 表 压缩率
with gr.Row():
# dynamic change label
with gr.Column():
output_text_1 = gr.Highlightedtext(
show_legend=True,
elem_classes="space-show"
)
with gr.Column():
output_text_2 = gr.Highlightedtext(
show_legend=True,
elem_classes="space-show"
)
with gr.Row():
output_table_1 = gr.Dataframe()
output_table_2 = gr.Dataframe()
# setting
# compress_rate_unit.change(compress_rate_unit_change, [compress_rate_unit],
# [stats_compress_rate_1, stats_compress_rate_2])
tokenizer_type_1.change(tokenize, [user_input, tokenizer_type_1],
[output_text_1, output_table_1])
tokenizer_type_1.change(basic_count, [tokenizer_type_1], [stats_vocab_size_1, stats_zh_token_size_1])
tokenizer_type_1.change(get_overlap_token_size, [tokenizer_type_1, tokenizer_type_2],
[stats_overlap_token_size_1, stats_overlap_token_size_2])
tokenizer_type_1.change(get_compress_rate, [tokenizer_type_1, compress_rate_corpus, compress_rate_unit],
[stats_compress_rate_1])
# TODO: every=3
user_input.change(tokenize_pair,
[user_input, tokenizer_type_1, tokenizer_type_2],
[output_text_1, output_table_1, output_text_2, output_table_2]) # , pass_request=1
tokenizer_type_2.change(tokenize, [user_input, tokenizer_type_2],
[output_text_2, output_table_2])
tokenizer_type_2.change(basic_count, [tokenizer_type_2], [stats_vocab_size_2, stats_zh_token_size_2])
tokenizer_type_2.change(get_overlap_token_size, [tokenizer_type_1, tokenizer_type_2],
[stats_overlap_token_size_1, stats_overlap_token_size_2])
tokenizer_type_2.change(get_compress_rate, [tokenizer_type_2, compress_rate_corpus, compress_rate_unit],
[stats_compress_rate_2])
compress_rate_unit.change(get_compress_rate, [tokenizer_type_1, compress_rate_corpus, compress_rate_unit],
[stats_compress_rate_1])
compress_rate_unit.change(get_compress_rate, [tokenizer_type_2, compress_rate_corpus, compress_rate_unit],
[stats_compress_rate_2])
dropdown_examples.change(
example_fn,
dropdown_examples,
[user_input, tokenizer_type_1, tokenizer_type_2]
)
demo.load(js=open("js/onload.js", "r", encoding="utf-8").read())
demo.load(
fn=on_load,
inputs=[user_input], # 这里只需要传个空object即可。
outputs=[user_input, tokenizer_type_1, tokenizer_type_2],
js=get_window_url_params
)
if __name__ == "__main__":
# demo.queue(max_size=20).launch()
demo.launch()
# demo.launch(share=True)