nllb-zh-pt-14k / model.py
jamesebond's picture
Upload 5 files
c4039ba verified
raw
history blame
20.4 kB
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import torch
from modules.file import ExcelFileWriter
import os
from abc import ABC, abstractmethod
from typing import List
import re
class FilterPipeline():
def __init__(self, filter_list):
self._filter_list:List[Filter] = filter_list
def append(self, filter):
self._filter_list.append(filter)
def batch_encoder(self, inputs):
for filter in self._filter_list:
inputs = filter.encoder(inputs)
return inputs
def batch_decoder(self, inputs):
for filter in reversed(self._filter_list):
inputs = filter.decoder(inputs)
return inputs
class Filter(ABC):
def __init__(self):
self.name = 'filter'
self.code = []
@abstractmethod
def encoder(self, inputs):
pass
@abstractmethod
def decoder(self, inputs):
pass
class SpecialTokenFilter(Filter):
def __init__(self):
self.name = 'special token filter'
self.code = []
self.special_tokens = ['!', '!', '-']
def encoder(self, inputs):
filtered_inputs = []
self.code = []
for i, input_str in enumerate(inputs):
if not all(char in self.special_tokens for char in input_str):
filtered_inputs.append(input_str)
else:
self.code.append([i, input_str])
return filtered_inputs
def decoder(self, inputs):
original_inputs = inputs.copy()
for removed_indice in self.code:
original_inputs.insert(removed_indice[0], removed_indice[1])
return original_inputs
class SperSignFilter(Filter):
def __init__(self):
self.name = 's percentage sign filter'
self.code = []
def encoder(self, inputs):
encoded_inputs = []
self.code = [] # 清空 self.code
for i, input_str in enumerate(inputs):
if '%s' in input_str:
encoded_str = input_str.replace('%s', '*')
self.code.append(i) # 将包含 '%s' 的字符串的索引存储到 self.code 中
else:
encoded_str = input_str
encoded_inputs.append(encoded_str)
return encoded_inputs
def decoder(self, inputs):
decoded_inputs = inputs.copy()
for i in self.code:
decoded_inputs[i] = decoded_inputs[i].replace('*', '%s') # 使用 self.code 中的索引还原原始字符串
return decoded_inputs
class ParenSParenFilter(Filter):
def __init__(self):
self.name = 'Paren s paren filter'
self.code = []
def encoder(self, inputs):
encoded_inputs = []
self.code = [] # 清空 self.code
for i, input_str in enumerate(inputs):
if '(s)' in input_str:
encoded_str = input_str.replace('(s)', '$')
self.code.append(i) # 将包含 '(s)' 的字符串的索引存储到 self.code 中
else:
encoded_str = input_str
encoded_inputs.append(encoded_str)
return encoded_inputs
def decoder(self, inputs):
decoded_inputs = inputs.copy()
for i in self.code:
decoded_inputs[i] = decoded_inputs[i].replace('$', '(s)') # 使用 self.code 中的索引还原原始字符串
return decoded_inputs
class ChevronsFilter(Filter):
def __init__(self):
self.name = 'chevrons filter'
self.code = []
def encoder(self, inputs):
encoded_inputs = []
self.code = [] # 清空 self.code
pattern = re.compile(r'<.*?>')
for i, input_str in enumerate(inputs):
if pattern.search(input_str):
matches = pattern.findall(input_str)
encoded_str = pattern.sub('#', input_str)
self.code.append((i, matches)) # 将包含匹配模式的字符串的索引和匹配列表存储到 self.code 中
else:
encoded_str = input_str
encoded_inputs.append(encoded_str)
return encoded_inputs
def decoder(self, inputs):
decoded_inputs = inputs.copy()
for i, matches in self.code:
for match in matches:
decoded_inputs[i] = decoded_inputs[i].replace('#', match, 1) # 使用 self.code 中的匹配列表依次还原原始字符串
return decoded_inputs
class SimilarFilter(Filter):
def __init__(self):
self.name = 'similar filter'
self.code = []
def is_similar(self, str1, str2):
# 判断两个字符串是否相似(只有数字上有区别)
pattern = re.compile(r'\d+')
return pattern.sub('', str1) == pattern.sub('', str2)
def encoder(self, inputs):
encoded_inputs = []
self.code = [] # 清空 self.code
i = 0
while i < len(inputs):
encoded_inputs.append(inputs[i])
similar_strs = [inputs[i]]
j = i + 1
while j < len(inputs) and self.is_similar(inputs[i], inputs[j]):
similar_strs.append(inputs[j])
j += 1
if len(similar_strs) > 1:
self.code.append((i, similar_strs)) # 将相似字符串的起始索引和实际字符串列表存储到 self.code 中
i = j
return encoded_inputs
def decoder(self, inputs:List):
decoded_inputs = inputs
for i, similar_strs in self.code:
pattern = re.compile(r'\d+')
for j in range(len(similar_strs)):
if pattern.search(similar_strs[j]):
number = re.findall(r'\d+', similar_strs[j])[0] # 获取相似字符串的数字部分
new_str = pattern.sub(number, inputs[i]) # 将新字符串的数字部分替换为相似字符串的数字部分
else:
new_str = inputs[i] # 如果相似字符串不含数字,直接使用新字符串
if j > 0:
decoded_inputs.insert(i+j, new_str)
return decoded_inputs
class ChineseFilter:
def __init__(self, pinyin_lib_file='pinyin.txt'):
self.name = 'chinese filter'
self.code = []
self.pinyin_lib = self.load_pinyin_lib(pinyin_lib_file)
def load_pinyin_lib(self, file_path):
with open(os.path.join(script_dir,file_path), 'r', encoding='utf-8') as f:
return set(line.strip().lower() for line in f)
def is_valid_chinese(self, word):
# 判断一个单词是否符合要求:只有一个单词构成,并且首字母大写
if len(word.split()) == 1 and word[0].isupper():
# 使用pinyin_or_word函数判断是否是合法的拼音
return self.is_pinyin(word.lower())
return False
def encoder(self, inputs):
encoded_inputs = []
self.code = [] # 清空 self.code
for i, word in enumerate(inputs):
if self.is_valid_chinese(word):
self.code.append((i, word)) # 将需要过滤的中文单词的索引和拼音存储到 self.code 中
else:
encoded_inputs.append(word)
return encoded_inputs
def decoder(self, inputs):
decoded_inputs = inputs.copy()
for i, word in self.code:
decoded_inputs.insert(i, word) # 根据索引将过滤的中文单词还原到原位置
return decoded_inputs
def is_pinyin(self, string):
'''
judge a string is a pinyin or a english word.
pinyin_Lib comes from a txt file.
'''
string = string.lower()
stringlen = len(string)
max_len = 6
result = []
n = 0
while n < stringlen:
matched = 0
temp_result = []
for i in range(max_len, 0, -1):
s = string[0:i]
if s in self.pinyin_lib:
temp_result.append(string[:i])
matched = i
break
if i == 1 and len(temp_result) == 0:
return False
result.extend(temp_result)
string = string[matched:]
n += matched
return True
script_dir = os.path.dirname(os.path.abspath(__file__))
parent_dir = os.path.dirname(os.path.dirname(os.path.dirname(script_dir)))
class Model():
def __init__(self, modelname, selected_lora_model, selected_gpu):
def get_gpu_index(gpu_info, target_gpu_name):
"""
从 GPU 信息中获取目标 GPU 的索引
Args:
gpu_info (list): 包含 GPU 名称的列表
target_gpu_name (str): 目标 GPU 的名称
Returns:
int: 目标 GPU 的索引,如果未找到则返回 -1
"""
for i, name in enumerate(gpu_info):
if target_gpu_name.lower() in name.lower():
return i
return -1
if selected_gpu != "cpu":
gpu_count = torch.cuda.device_count()
gpu_info = [torch.cuda.get_device_name(i) for i in range(gpu_count)]
selected_gpu_index = get_gpu_index(gpu_info, selected_gpu)
self.device_name = f"cuda:{selected_gpu_index}"
else:
self.device_name = "cpu"
print("device_name", self.device_name)
self.model = AutoModelForSeq2SeqLM.from_pretrained(modelname).to(self.device_name)
self.tokenizer = AutoTokenizer.from_pretrained(modelname)
# self.translator = pipeline('translation', model=self.original_model, tokenizer=self.tokenizer, src_lang=original_language, tgt_lang=target_language, device=device)
def generate(self, inputs, original_language, target_languages, max_batch_size):
filter_list = [SpecialTokenFilter(), SperSignFilter(), ParenSParenFilter(), ChevronsFilter(), SimilarFilter(), ChineseFilter()]
filter_pipeline = FilterPipeline(filter_list)
def language_mapping(original_language):
d = {
"Achinese (Arabic script)": "ace_Arab",
"Achinese (Latin script)": "ace_Latn",
"Mesopotamian Arabic": "acm_Arab",
"Ta'izzi-Adeni Arabic": "acq_Arab",
"Tunisian Arabic": "aeb_Arab",
"Afrikaans": "afr_Latn",
"South Levantine Arabic": "ajp_Arab",
"Akan": "aka_Latn",
"Amharic": "amh_Ethi",
"North Levantine Arabic": "apc_Arab",
"Standard Arabic": "arb_Arab",
"Najdi Arabic": "ars_Arab",
"Moroccan Arabic": "ary_Arab",
"Egyptian Arabic": "arz_Arab",
"Assamese": "asm_Beng",
"Asturian": "ast_Latn",
"Awadhi": "awa_Deva",
"Central Aymara": "ayr_Latn",
"South Azerbaijani": "azb_Arab",
"North Azerbaijani": "azj_Latn",
"Bashkir": "bak_Cyrl",
"Bambara": "bam_Latn",
"Balinese": "ban_Latn",
"Belarusian": "bel_Cyrl",
"Bemba": "bem_Latn",
"Bengali": "ben_Beng",
"Bhojpuri": "bho_Deva",
"Banjar (Arabic script)": "bjn_Arab",
"Banjar (Latin script)": "bjn_Latn",
"Tibetan": "bod_Tibt",
"Bosnian": "bos_Latn",
"Buginese": "bug_Latn",
"Bulgarian": "bul_Cyrl",
"Catalan": "cat_Latn",
"Cebuano": "ceb_Latn",
"Czech": "ces_Latn",
"Chokwe": "cjk_Latn",
"Central Kurdish": "ckb_Arab",
"Crimean Tatar": "crh_Latn",
"Welsh": "cym_Latn",
"Danish": "dan_Latn",
"German": "deu_Latn",
"Dinka": "dik_Latn",
"Jula": "dyu_Latn",
"Dzongkha": "dzo_Tibt",
"Greek": "ell_Grek",
"English": "eng_Latn",
"Esperanto": "epo_Latn",
"Estonian": "est_Latn",
"Basque": "eus_Latn",
"Ewe": "ewe_Latn",
"Faroese": "fao_Latn",
"Persian": "pes_Arab",
"Fijian": "fij_Latn",
"Finnish": "fin_Latn",
"Fon": "fon_Latn",
"French": "fra_Latn",
"Friulian": "fur_Latn",
"Nigerian Fulfulde": "fuv_Latn",
"Scottish Gaelic": "gla_Latn",
"Irish": "gle_Latn",
"Galician": "glg_Latn",
"Guarani": "grn_Latn",
"Gujarati": "guj_Gujr",
"Haitian Creole": "hat_Latn",
"Hausa": "hau_Latn",
"Hebrew": "heb_Hebr",
"Hindi": "hin_Deva",
"Chhattisgarhi": "hne_Deva",
"Croatian": "hrv_Latn",
"Hungarian": "hun_Latn",
"Armenian": "hye_Armn",
"Igbo": "ibo_Latn",
"Iloko": "ilo_Latn",
"Indonesian": "ind_Latn",
"Icelandic": "isl_Latn",
"Italian": "ita_Latn",
"Javanese": "jav_Latn",
"Japanese": "jpn_Jpan",
"Kabyle": "kab_Latn",
"Kachin": "kac_Latn",
"Arabic": "ar_AR",
"Chinese": "zho_Hans",
"Spanish": "spa_Latn",
"Dutch": "nld_Latn",
"Kazakh": "kaz_Cyrl",
"Korean": "kor_Hang",
"Lithuanian": "lit_Latn",
"Malayalam": "mal_Mlym",
"Marathi": "mar_Deva",
"Nepali": "ne_NP",
"Polish": "pol_Latn",
"Portuguese": "por_Latn",
"Russian": "rus_Cyrl",
"Sinhala": "sin_Sinh",
"Tamil": "tam_Taml",
"Turkish": "tur_Latn",
"Ukrainian": "ukr_Cyrl",
"Urdu": "urd_Arab",
"Vietnamese": "vie_Latn",
"Thai":"tha_Thai",
"Khmer":"khm_Khmr"
}
return d[original_language]
def process_gpu_translate_result(temp_outputs):
outputs = []
for temp_output in temp_outputs:
length = len(temp_output[0]["generated_translation"])
for i in range(length):
temp = []
for trans in temp_output:
temp.append({
"target_language": trans["target_language"],
"generated_translation": trans['generated_translation'][i],
})
outputs.append(temp)
excel_writer = ExcelFileWriter()
excel_writer.write_text(os.path.join(parent_dir,r"temp/empty.xlsx"), outputs, 'A', 1, len(outputs))
self.tokenizer.src_lang = language_mapping(original_language)
if self.device_name == "cpu":
# Tokenize input
input_ids = self.tokenizer(inputs, return_tensors="pt", padding=True, max_length=128).to(self.device_name)
output = []
for target_language in target_languages:
# Get language code for the target language
target_lang_code = self.tokenizer.lang_code_to_id[language_mapping(target_language)]
# Generate translation
generated_tokens = self.model.generate(
**input_ids,
forced_bos_token_id=target_lang_code,
max_length=128
)
generated_translation = self.tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# Append result to output
output.append({
"target_language": target_language,
"generated_translation": generated_translation,
})
outputs = []
length = len(output[0]["generated_translation"])
for i in range(length):
temp = []
for trans in output:
temp.append({
"target_language": trans["target_language"],
"generated_translation": trans['generated_translation'][i],
})
outputs.append(temp)
return outputs
else:
# 最大批量大小 = 可用 GPU 内存字节数 / 4 / (张量大小 + 可训练参数)
# max_batch_size = 10
# Ensure batch size is within model limits:
print("length of inputs: ",len(inputs))
batch_size = min(len(inputs), int(max_batch_size))
batches = [inputs[i:i + batch_size] for i in range(0, len(inputs), batch_size)]
print("length of batches size: ", len(batches))
temp_outputs = []
processed_num = 0
for index, batch in enumerate(batches):
# Tokenize input
print(">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>")
print(len(batch))
print(batch)
batch = filter_pipeline.batch_encoder(batch)
print(batch)
temp = []
if len(batch) > 0:
input_ids = self.tokenizer(batch, return_tensors="pt", padding=True).to(self.device_name)
for target_language in target_languages:
target_lang_code = self.tokenizer.lang_code_to_id[language_mapping(target_language)]
generated_tokens = self.model.generate(
**input_ids,
forced_bos_token_id=target_lang_code,
)
generated_translation = self.tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
print(generated_translation)
generated_translation = filter_pipeline.batch_decoder(generated_translation)
print(generated_translation)
print(len(generated_translation))
# Append result to output
temp.append({
"target_language": target_language,
"generated_translation": generated_translation,
})
input_ids.to('cpu')
del input_ids
else:
for target_language in target_languages:
generated_translation = filter_pipeline.batch_decoder(batch)
print(generated_translation)
print(len(generated_translation))
# Append result to output
temp.append({
"target_language": target_language,
"generated_translation": generated_translation,
})
temp_outputs.append(temp)
processed_num += len(batch)
if (index + 1) * max_batch_size // 1000 - index * max_batch_size // 1000 == 1:
print("Already processed number: ", len(temp_outputs))
process_gpu_translate_result(temp_outputs)
outputs = []
for temp_output in temp_outputs:
length = len(temp_output[0]["generated_translation"])
for i in range(length):
temp = []
for trans in temp_output:
temp.append({
"target_language": trans["target_language"],
"generated_translation": trans['generated_translation'][i],
})
outputs.append(temp)
return outputs
for filter in self._filter_list:
inputs = filter.encoder(inputs)
return inputs
def batch_decoder(self, inputs):
for filter in reversed(self._filter_list):
inputs = filter.decoder(inputs)
return inputs