# stealt & modified from https://github.com/zyddnys/manga-image-translator/blob/main/manga_translator/translators/chatgpt.py import re import time from typing import List, Dict, Union import yaml import traceback import inspect import openai from .base import BaseTranslator, register_translator OPENAPI_V1_API = int(openai.__version__.split('.')[0]) >= 1 class InvalidNumTranslations(Exception): pass @register_translator('ChatGPT') class GPTTranslator(BaseTranslator): concate_text = False cht_require_convert = True params: Dict = { 'api key': '', 'model': { 'type': 'selector', 'options': [ 'gpt-4o', 'gpt-4-turbo', 'gpt3', 'gpt35-turbo', 'gpt4', ], 'value': 'gpt-4o' }, 'override model': '', 'prompt template': { 'type': 'editor', 'value': 'Please help me to translate the following text from a manga to {to_lang} (if it\'s already in {to_lang} or looks like gibberish you have to output it as it is instead):\n', }, 'chat system template': { 'type': 'editor', 'value': 'You are a professional translation engine, please translate the text into a colloquial, elegant and fluent content, without referencing machine translations. You must only translate the text content, never interpret it. If there\'s any issue in the text, output the text as is.\nTranslate to {to_lang}.', }, 'chat sample': { 'type': 'editor', 'value': '''日本語-简体中文: source: - 二人のちゅーを 目撃した ぼっちちゃん - ふたりさん - 大好きなお友達には あいさつ代わりに ちゅーするんだって - アイス あげた - 喜多ちゃんとは どどど どういった ご関係なのでしようか... - テレビで見た！ target: - 小孤独目击了两人的接吻 - 二里酱 - 我听说人们会把亲吻作为与喜爱的朋友打招呼的方式 - 我给了她冰激凌 - 喜多酱和你是怎么样的关系啊... - 我在电视上看到的！''' }, 'invalid repeat count': 2, 'max requests per minute': 20, 'delay': 0.3, 'max tokens': 4096, 'temperature': 0.5, 'top p': 1., # 'return prompt': False, 'retry attempts': 5, 'retry timeout': 15, '3rd party api url': '', 'frequency penalty': 0.0, 'presence penalty': 0.0, 'low vram mode': { 'value': False, 'description': 'check it if you\'re running it locally on a single device and encountered a crash due to vram OOM', 'type': 'checkbox', } } def _setup_translator(self): self.lang_map['简体中文'] = 'Simplified Chinese' self.lang_map['繁體中文'] = 'Traditional Chinese' self.lang_map['日本語'] = 'Japanese' self.lang_map['English'] = 'English' self.lang_map['한국어'] = 'Korean' self.lang_map['Tiếng Việt'] = 'Vietnamese' self.lang_map['čeština'] = 'Czech' self.lang_map['Français'] = 'French' self.lang_map['Deutsch'] = 'German' self.lang_map['magyar nyelv'] = 'Hungarian' self.lang_map['Italiano'] = 'Italian' self.lang_map['Polski'] = 'Polish' self.lang_map['Português'] = 'Portuguese' self.lang_map['limba română'] = 'Romanian' self.lang_map['русский язык'] = 'Russian' self.lang_map['Español'] = 'Spanish' self.lang_map['Türk dili'] = 'Turkish' self.lang_map['украї́нська мо́ва'] = 'Ukrainian' self.lang_map['Thai'] = 'Thai' self.lang_map['Arabic'] = 'Arabic' self.lang_map['Malayalam'] = 'Malayalam' self.lang_map['Tamil'] = 'Tamil' self.lang_map['Hindi'] = 'Hindi' self.token_count = 0 self.token_count_last = 0 @property def model(self) -> str: return self.params['model']['value'] @property def temperature(self) -> float: return self.params['temperature'] @property def max_tokens(self) -> int: return self.params['max tokens'] @property def top_p(self) -> float: return self.params['top p'] @property def retry_attempts(self) -> int: return self.params['retry attempts'] @property def retry_timeout(self) -> int: return self.params['retry timeout'] @property def chat_system_template(self) -> str: to_lang = self.lang_map[self.lang_target] return self.params['chat system template']['value'].format(to_lang=to_lang) @property def chat_sample(self): if self.model == 'gpt3': return None samples = self.params['chat sample']['value'] try: samples = yaml.load(self.params['chat sample']['value'], Loader=yaml.FullLoader) except: self.logger.error(f'failed to load parse sample: {samples}') samples = {} src_tgt = self.lang_source + '-' + self.lang_target if src_tgt in samples: src_list = samples[src_tgt]['source'] tgt_list = samples[src_tgt]['target'] src_queries = '' tgt_queries = '' for i, (src, tgt) in enumerate(zip(src_list, tgt_list)): src_queries += f'\n<|{i+1}|>{src}' tgt_queries += f'\n<|{i+1}|>{tgt}' src_queries = src_queries.lstrip() tgt_queries = tgt_queries.lstrip() return [src_queries, tgt_queries] else: return None def _assemble_prompts(self, queries: List[str], from_lang: str = None, to_lang: str = None, max_tokens = None): if from_lang is None: from_lang = self.lang_map[self.lang_source] if to_lang is None: to_lang = self.lang_map[self.lang_target] prompt = '' if max_tokens is None: max_tokens = self.max_tokens # return_prompt = self.params['return prompt'] prompt_template = self.params['prompt template']['value'].format(to_lang=to_lang).rstrip() prompt += prompt_template i_offset = 0 num_src = 0 for i, query in enumerate(queries): prompt += f'\n<|{i+1-i_offset}|>{query}' num_src += 1 # If prompt is growing too large and theres still a lot of text left # split off the rest of the queries into new prompts. # 1 token = ~4 characters according to https://platform.openai.com/tokenizer # TODO: potentially add summarizations from special requests as context information if max_tokens * 2 and len(''.join(queries[i+1:])) > max_tokens: # if return_prompt: # prompt += '\n<|1|>' yield prompt.lstrip(), num_src prompt = prompt_template # Restart counting at 1 i_offset = i + 1 num_src = 0 # if return_prompt: # prompt += '\n<|1|>' yield prompt.lstrip(), num_src def _format_prompt_log(self, to_lang: str, prompt: str) -> str: chat_sample = self.chat_sample if self.model != 'gpt3' and chat_sample is not None: return '\n'.join([ 'System:', self.chat_system_template, 'User:', chat_sample[0], 'Assistant:', chat_sample[1], 'User:', prompt, ]) else: return '\n'.join([ 'System:', self.chat_system_template, 'User:', prompt, ]) def _translate(self, src_list: List[str]) -> List[str]: translations = [] # self.logger.debug(f'Temperature: {self.temperature}, TopP: {self.top_p}') from_lang = self.lang_map[self.lang_source] to_lang = self.lang_map[self.lang_target] queries = src_list # return_prompt = self.params['return prompt'] chat_sample = self.chat_sample for prompt, num_src in self._assemble_prompts(queries, from_lang, to_lang): retry_attempt = 0 while True: try: response = self._request_translation(prompt, chat_sample) new_translations = re.split(r'<\|\d+\|>', response)[-num_src:] if len(new_translations) != num_src: # https://github.com/dmMaze/BallonsTranslator/issues/379 _tr2 = re.sub(r'<\|\d+\|>', '', response) _tr2 = _tr2.split('\n') if len(_tr2) == num_src: new_translations = _tr2 else: raise InvalidNumTranslations break except InvalidNumTranslations: retry_attempt += 1 message = f'number of translations does not match to source:\nprompt:\n {prompt}\ntranslations:\n {new_translations}\nopenai response:\n {response}' if retry_attempt >= self.retry_attempts: self.logger.error(message) new_translations = [''] * num_src break self.logger.warning(message + '\n' + f'Restarting request. Attempt: {retry_attempt}') except Exception as e: retry_attempt += 1 if retry_attempt >= self.retry_attempts: new_translations = [''] * num_src break self.logger.warning(f'Translation failed due to {e}. Attempt: {retry_attempt}, sleep for {self.retry_timeout} secs...') self.logger.error(f'Request traceback: ', traceback.format_exc()) time.sleep(self.retry_timeout) # time.sleep(self.retry_timeout) # if return_prompt: # new_translations = new_translations[:-1] # if chat_sample is not None: # new_translations = new_translations[1:] translations.extend([t.strip() for t in new_translations]) if self.token_count_last: self.logger.info(f'Used {self.token_count_last} tokens (Total: {self.token_count})') return translations def _request_translation_gpt3(self, prompt: str) -> str: if OPENAPI_V1_API: openai_completions_create = openai.completions.create else: openai_completions_create = openai.Completion.create response = openai_completions_create( model='text-davinci-003', prompt=prompt, max_tokens=self.max_tokens // 2, # Assuming that half of the tokens are used for the query temperature=self.temperature, top_p=self.top_p, frequency_penalty=float(self.params['frequency penalty']), presence_penalty=float(self.params['presence penalty']) ) if OPENAPI_V1_API: if response.usage is not None: self.token_count += response.usage.total_tokens self.token_count_last = response.usage.total_tokens else: self.token_count += response.usage['total_tokens'] self.token_count_last = response.usage['total_tokens'] return response.choices[0].text def _request_translation_with_chat_sample(self, prompt: str, model: str, chat_sample: List) -> str: messages = [ {'role': 'system', 'content': self.chat_system_template}, {'role': 'user', 'content': prompt}, ] if chat_sample is not None: messages.insert(1, {'role': 'user', 'content': chat_sample[0]}) messages.insert(2, {'role': 'assistant', 'content': chat_sample[1]}) func_args = { 'model': model, 'messages': messages, 'temperature': self.temperature, 'top_p': self.top_p, } max_tokens = self.max_tokens // 2 # Assuming that half of the tokens are used for the query func_parameters = inspect.signature(openai.chat.completions.create).parameters if 'max_completion_tokens' in func_parameters: func_args['max_completion_tokens'] = max_tokens else: func_args['max_tokens'] = max_tokens if 'presence_penalty' in func_parameters: func_args['presence_penalty'] = self.params['presence penalty'] func_args['frequency_penalty'] = self.params['frequency penalty'] if OPENAPI_V1_API: openai_chatcompletions_create = openai.chat.completions.create else: openai_chatcompletions_create = openai.ChatCompletion.create response = openai_chatcompletions_create(**func_args) if OPENAPI_V1_API: if response.usage is not None: self.token_count += response.usage.total_tokens self.token_count_last = response.usage.total_tokens else: self.token_count += response.usage['total_tokens'] self.token_count_last = response.usage['total_tokens'] for choice in response.choices: if OPENAPI_V1_API: return choice.message.content else: if 'text' in choice: return choice.text # If no response with text is found, return the first response's content (which may be empty) return response.choices[0].message.content @property def api_url(self): url = self.params['3rd party api url'].strip() if not url: return None # 对于小于v1.0.0版本的openai包，末尾的斜杠会导致请求失败，因此弹出警告 if url.endswith('v1/'): if not OPENAPI_V1_API: self.logger.warning(f"The OpenAI package version you are using is outdated. Please remove the trailing slash after 'v1' in the URL: {url}") # 检查是否包含"/v1" if '/v1' not in url: self.logger.warning(f"API URL does not contain '/v1': {url}, please ensure it's the correct URL.") return url def _request_translation(self, prompt, chat_sample: List): self.logger.debug(f'chatgpt prompt: \n {prompt}' ) openai.api_key = self.params['api key'].strip() base_url = self.api_url if OPENAPI_V1_API: openai.base_url = base_url else: if base_url is None: base_url = 'https://api.openai.com/v1' openai.api_base = base_url override_model = self.params['override model'].strip() if override_model != '': model: str = override_model else: model:str = self.model if model == 'gpt3': return self._request_translation_gpt3(prompt) elif model == 'gpt35-turbo': model = 'gpt-3.5-turbo' elif model == 'gpt4': model = 'gpt-4' return self._request_translation_with_chat_sample(prompt, model, chat_sample)