tokenizer-arena / compression_util.py
xu-song's picture
update
7c73423
raw
history blame
13.1 kB
"""
## more statistics
code:
math:
digit:
whitespace:
top_oov: most frequent oov chars
ranking: thumb_up thumb_down
"""
import json
import os
import sys
from difflib import SequenceMatcher
import pandas as pd
from datasets import load_dataset
from utils.log_util import logger
from vocab import tokenizer_factory, TokenizerConfig
from typing import List, Optional, Union, Literal
CURRENT_DIR = os.path.dirname(os.path.abspath(__file__))
common_units = ["g_bytes/b_tokens", "b_tokens/g_bytes", "t_bytes/t_tokens", "t_tokens/t_bytes", "n_chars/n_tokens", ]
common_corpuses = sorted(["cc100/en", "cc100/zh-Hans", "cc100/es", "cc100/fr", "cc100/de", "cc100/ko",
"cc100/fa", "cc100/ar", "cc100/ja"])
VALID_CODES_CC100 = [
"am", "ar", "as", "az", "be", "bg", "bn", "bn_rom", "br", "bs", "ca", "cs", "cy", "da", "de",
"el", "en", "eo", "es", "et", "eu", "fa", "ff", "fi", "fr", "fy", "ga", "gd", "gl", "gn", "gu",
"ha", "he", "hi", "hi_rom", "hr", "ht", "hu", "hy", "id", "ig", "is", "it", "ja", "jv", "ka",
"kk", "km", "kn", "ko", "ku", "ky", "la", "lg", "li", "ln", "lo", "lt", "lv", "mg", "mk", "ml",
"mn", "mr", "ms", "my", "my_zaw", "ne", "nl", "no", "ns", "om", "or", "pa", "pl", "ps", "pt",
"qu", "rm", "ro", "ru", "sa", "si", "sc", "sd", "sk", "sl", "so", "sq", "sr", "ss", "su", "sv",
"sw", "ta", "ta_rom", "te", "te_rom", "th", "tl", "tn", "tr", "ug", "uk", "ur", "ur_rom", "uz",
"vi", "wo", "xh", "yi", "yo", "zh-Hans", "zh-Hant", "zu",
]
# code: https://huggingface.co/datasets/codeparrot/github-code-clean python java c sql html
# math:
def get_n_bytes_of_string(string_text):
n_bytes = len(string_text.encode("utf-8"))
return n_bytes
def unit_convertor(stat, unit):
n_tokens = stat["_n_tokens"]
n_chars = stat["_n_chars"]
n_bytes = stat["_n_bytes"]
if n_tokens is None:
return None
n_tokens_in_billion = n_tokens / (1000 * 1000 * 1000)
n_tokens_in_trillion = n_tokens / (1000 * 1000 * 1000 * 1000)
n_bytes_in_mb = n_bytes / (1024 * 1024)
n_bytes_in_gb = n_bytes_in_mb / 1024
n_bytes_in_tb = n_bytes_in_gb / 1024
# n_chars_in_billion = n_chars / (1000 * 1000 * 1000)
if unit == "n_tokens/n_bytes":
value = n_tokens / n_bytes
elif unit in ["char/token", "chars_per_token"]: # 重要:平均一个token包含多少个字符。
value = n_chars / n_tokens
elif unit in ["token/char", "tokens_per_char"]: # 一个中文汉字需要几个token?
value = n_tokens / n_chars
elif unit == "g_bytes/b_tokens":
value = n_bytes_in_gb / n_tokens_in_billion
elif unit == "b_tokens/g_bytes":
value = n_tokens_in_billion / n_bytes_in_gb
elif unit == "t_bytes/t_tokens": # 重要:
value = n_bytes_in_tb / n_tokens_in_trillion
elif unit == "t_tokens/t_bytes":
value = n_tokens_in_trillion / n_bytes_in_tb
else:
raise "measure not support"
return round(value, 3)
def _merge_stats_by_corpus(stats_by_corpus, oov_threshold=0.3):
"""
"""
all_stats = list(stats_by_corpus.values())
assert len(set([stats["tokenizer"] for stats in all_stats])) == 1
lossless = all(stat['lossless'] for stat in all_stats)
is_support = all(stat['oov_ratio'] < oov_threshold for stat in all_stats)
merged_stats = {
"tokenizer": all_stats[0]["tokenizer"],
"organization": all_stats[0]["organization"],
"vocab_size": all_stats[0]["vocab_size"],
"_n_bytes": 0,
"_n_tokens": 0 if is_support else None,
"_n_chars": 0,
"_n_oov_chars": 0,
"lossless": True,
}
for stats in all_stats:
merged_stats["_n_bytes"] += stats["_n_bytes"]
merged_stats["_n_chars"] += stats["_n_chars"]
if is_support: # The number of tokens cannot be accurately counted, when there are too many UNKs.
merged_stats["_n_tokens"] += stats["_n_tokens"]
merged_stats["_n_oov_chars"] += stats["_n_oov_chars"]
merged_stats["lossless"] &= stats['lossless']
merged_stats.update({
"oov_ratio": float("%.4g" % (stats["_n_oov_chars"] / stats["_n_chars"])),
"lossless": lossless
})
return merged_stats
def to_dataframe(stats, units=None):
if units is None:
units = common_units
elif not isinstance(units, list):
units = [units]
table = []
for stat in stats.values():
columns = {k: v for k, v in stat.items() if not k.startswith("_")}
for unit in units:
if unit not in stat:
columns[unit] = unit_convertor(stat, unit)
else:
logger.error(f"unit {unit} not support")
table.append(columns)
df = pd.DataFrame(table)
return df
cache = {}
def tokenize_corpus(
tokenizer_name: str,
corpuses: List[str],
cache_dir: str = "stats"
) -> dict:
"""
:param tokenizer_name:
:param corpuses:
:param cache_dir:
:return:
"""
def _assert_oov(tokenizer, oov_candidate):
tokenizer.encode()
def _char_based_oov(src_text, decoded_text, tokenizer):
oov_charset = [] # keep the order in src_text
decoded_charset = set(decoded_text)
for char in dict.fromkeys(src_text):
if char not in decoded_charset \
and char != tokenizer.decode(tokenizer.encode(char, add_special_tokens=False)):
oov_charset.append(char)
n_oov_chars = sum([1 for char in src_text if char in oov_charset])
return n_oov_chars, oov_charset
def _diff_path(src_text, decoded_text):
s = SequenceMatcher(a=src_text, b=decoded_text)
changes = []
for tag, i1, i2, j1, j2 in s.get_opcodes():
if tag != "equal":
changes.append('{:7} text[{}:{}] --> decoded_text[{}:{}] {!r:>8} --> {!r}'.format(
tag, i1, i2, j1, j2, src_text[i1:i2], decoded_text[j1:j2]))
return changes
def _tokenize(tokenizer, datasets, detail_path=None):
"""
:param tokenizer:
:param datasets:
:param detail_path:
:return:
"""
n_bytes = 0
n_tokens = 0
n_chars = 0
n_oov_chars = 0
diff_details = []
oov_charset = set()
unk_token_id = None
if hasattr(tokenizer, "unk_token"):
unk_token_id = tokenizer.unk_token_id
for dataset in datasets:
for item in dataset:
text = item["text"]
n_bytes += get_n_bytes_of_string(text)
n_chars += len(text)
ids = tokenizer.encode(text, add_special_tokens=False)
# detect oov
decoded_text = tokenizer.decode(ids)
decoded_text_without_unk = tokenizer.decode([token_id for token_id in ids if token_id != unk_token_id])
if decoded_text != text:
_n_oov_chars, _oov_charset = _char_based_oov(text, decoded_text_without_unk, tokenizer)
diffs = _diff_path(text, decoded_text)
diff_details.append(
{
"text": text,
"decoded_text": decoded_text,
"diff": diffs,
"n_oov_chars": _n_oov_chars,
'oov_ratio': _n_oov_chars / len(text),
'oov_charset': json.dumps(_oov_charset, ensure_ascii=False),
}
)
n_oov_chars += _n_oov_chars
oov_charset.update(_oov_charset)
n_tokens += len(ids)
stat = {
"_n_bytes": n_bytes,
"_n_tokens": n_tokens,
"_n_chars": n_chars,
"_n_oov_chars": n_oov_chars,
"oov_ratio": n_oov_chars / n_chars,
'_oov_charset': json.dumps(list(oov_charset), ensure_ascii=False),
"lossless": len(diff_details) == 0
}
if detail_path and diff_details:
logger.info(f"saving tokenization detail to '{detail_path}'")
with open(detail_path, "w", encoding="utf-8") as f:
f.write(json.dumps(diff_details, ensure_ascii=False, indent=2))
# print(f"{tokenizer_config.name_or_path}, {infer_tokenizer_type(tokenizer_config)}\n"
# f"lossless: false; unk_token: {get_unk(tokenizer_config)},"
# f" unk_ratio: {unk_count / len(encoding):.4f}; oov: []")
# for diff_detail in diff_details:
# # print(f"text[{i}] = {str(bytes(text[i:], 'utf-8'))}\n"
# # f"decoding[{i}] = {str(bytes(decoding[i:], 'utf-8'))}")
# f.write(f"text= {json.dumps(text[i:], ensure_ascii=False)}, \n"
# f"decoding[{i}] = {json.dumps(decoding[i:], ensure_ascii=False)}")
return stat
# load from cache
cache_id = f"{tokenizer_name} @ {'.'.join(corpuses)}"
cache_path = os.path.join(cache_dir, "compression_rate.json")
if not cache and os.path.exists(cache_path):
with open(cache_path, "r", encoding="utf-8") as f_tmp:
cache.update(json.load(f_tmp))
if cache_id in cache:
# logger.info(f"loading {cache_id} from in-memory cache")
return cache[cache_id]
# tokenize corpus
tokenizer = tokenizer_factory.get_tokenizer(tokenizer_name)
datasets = [load_dataset("eson/cc100-samples", corpus.replace("cc100/", ""), split="train") for corpus in corpuses]
stat = {
"tokenizer": tokenizer_factory.get_name_with_hyperlink(tokenizer_name),
"organization": tokenizer_factory.get_tokenizer_config(tokenizer_name).org,
"vocab_size": len(tokenizer),
}
tokenize_detail_dir = os.path.join(cache_dir, "compression_rate")
os.makedirs(tokenize_detail_dir, exist_ok=True)
tokenize_detail_path = os.path.join(tokenize_detail_dir, cache_id.replace("/", ".") + ".diff.json")
stat.update(_tokenize(tokenizer, datasets, detail_path=tokenize_detail_path))
# add basic info
# save to cache
len_before = len(cache)
cache[cache_id] = stat
len_after = len(cache)
logger.info(f"saving '{cache_id}' to memory and file cache '{cache_path}': {len_before}->{len_after}")
with open(cache_path, "w", encoding="utf-8") as f_tmp:
json.dump(cache, f_tmp, ensure_ascii=False, indent=2)
return stat
def get_compression_leaderboard(
corpuses: List[str] = ['cc100/en'],
unit: str = "b_tokens/g_bytes",
tokenizer_filter: Optional[str] = None,
return_type: Optional[Literal["dict", "dataframe"]] = "dataframe"
) -> Union[pd.DataFrame, dict]:
"""
"""
logger.info(f"corpuses: {corpuses}; unit: {unit}; tokenizer_filter: {tokenizer_filter}")
stats = {}
if tokenizer_filter is not None:
tokenizer_names = [tokenizer_name for tokenizer_name in tokenizer_factory.all_tokenizer_names
if tokenizer_filter.lower() in tokenizer_name.lower()]
else:
tokenizer_names = tokenizer_factory.all_tokenizer_names
for tokenizer_name in tokenizer_names:
stats_by_corpus = {}
for corpus in corpuses:
stats_by_corpus[corpus] = tokenize_corpus(tokenizer_name, [corpus])
stats[tokenizer_name] = _merge_stats_by_corpus(stats_by_corpus)
if return_type == "dataframe":
token_number_unit, file_size_unit = unit.split("/")
reverse_unit = f"{file_size_unit}/{token_number_unit}"
stats = to_dataframe(stats, [unit, reverse_unit, "char/token"])
stats = stats.sort_values(["oov_ratio", unit], ascending=[True, True])
stats = stats.rename(columns={"oov_ratio": f' ⬆️oov_ratio'}).rename(columns={unit: f' ⬆️{unit}'}) # ⬇
return stats
def main():
if len(sys.argv) == 3:
tokenizer_filter = [sys.argv[1]]
corpuses = [sys.argv[2]]
else:
tokenizer_filter, corpuses = None, common_corpuses
# tokenizer_filter, corpuses = "openai", ["cc100/en", "cc100/zh-Hans"]
# tokenizer_filter, corpuses = "Qwen/Qwen1.5-14B", ["cc100/de"]
# tokenizer_filter, corpuses = "Qwen/Qwen1.5-14B", ["cc100/ja"] # oov 特别多
# tokenizer_filter, corpuses = "google-bert/bert-base-uncased", ["cc100/ja", "cc100/zh-Hans"] # oov 特别多
df = get_compression_leaderboard(corpuses, tokenizer_filter=tokenizer_filter)
# print(df.to_markdown(index=False, tablefmt='fancy_grid'))
logger.info(f"\n{df.to_markdown(index=False)}")
if __name__ == "__main__":
main()