import os # os.environ["CUDA_VISIBLE_DEVICES"] = "0" # 指定要使用的GPU设备编号 from transformers import pipeline import argparse import openai import tiktoken import torch from scipy.spatial.distance import cosine from transformers import AutoModel, AutoTokenizer from argparse import Namespace from langchain.chat_models import ChatOpenAI import gradio as gr import random import time from langchain.prompts.chat import ( ChatPromptTemplate, SystemMessagePromptTemplate, AIMessagePromptTemplate, HumanMessagePromptTemplate, ) from langchain.schema import ( AIMessage, HumanMessage, SystemMessage ) from text import Text def download_models(): # Import our models. The package will take care of downloading the models automatically model_args = Namespace(do_mlm=None, pooler_type="cls", temp=0.05, mlp_only_train=False, init_embeddings_model=None) model = AutoModel.from_pretrained("silk-road/luotuo-bert", trust_remote_code=True, model_args=model_args) return model openai.api_key = os.environ.get('OPENAI_API_KEY') folder_name = "Suzumiya" current_directory = os.getcwd() new_directory = os.path.join(current_directory, folder_name) pkl_path = './pkl/texts.pkl' text_image_pkl_path='./pkl/text_image.pkl' dict_path = "characters/haruhi/text_image_dict.txt" dict_text_pkl_path = './pkl/dict_text.pkl' image_path = "characters/haruhi/images" model = download_models() text = Text("characters/haruhi/texts", text_image_pkl_path=text_image_pkl_path, dict_text_pkl_path=dict_text_pkl_path, model=model, num_steps=50, pkl_path=pkl_path, dict_path=dict_path, image_path=image_path) if not os.path.exists(new_directory): os.makedirs(new_directory) print(f"文件夹 '{folder_name}' 创建成功！") else: print(f"文件夹 '{folder_name}' 已经存在。") enc = tiktoken.get_encoding("cl100k_base") class Run: def __init__(self, **params): """ * 命令行参数的接入 * 台词folder,记录台词 * system prompt存成txt文件，支持切换 * 支持设定max_len_story 和max_len_history * 支持设定save_path * 实现一个colab脚本，可以clone转换后的项目并运行，方便其他用户体验 """ self.folder = params['folder'] # self.system_prompt = params['system_prompt'] with open(params['system_prompt'], 'r') as f: self.system_prompt = f.read() self.max_len_story = params['max_len_story'] self.max_len_history = params['max_len_history'] self.save_path = params['save_path'] self.titles, self.title_to_text = self.read_prompt_data() self.embeddings, self.embed_to_title = self.title_text_embedding(self.titles, self.title_to_text) # self.embeddings, self.embed_to_title = [], [] # 一个封装 OpenAI 接口的函数，参数为 Prompt，返回对应结果 def get_completion_from_messages(self, messages, model="gpt-3.5-turbo", temperature=0): response = openai.ChatCompletion.create( model=model, messages=messages, temperature=temperature, # 控制模型输出的随机程度 ) # print(str(response.choices[0].message)) return response.choices[0].message["content"] def read_prompt_data(self): """ read prompt-data for in-context-learning """ titles = [] title_to_text = {} for file in os.listdir(self.folder): if file.endswith('.txt'): title_name = file[:-4] titles.append(title_name) with open(os.path.join(self.folder, file), 'r') as f: title_to_text[title_name] = f.read() return titles, title_to_text def get_embedding(self, text): tokenizer = AutoTokenizer.from_pretrained("silk-road/luotuo-bert") model = download_models() if len(text) > 512: text = text[:512] texts = [text] # Tokenize the text inputs = tokenizer(texts, padding=True, truncation=False, return_tensors="pt") # Extract the embeddings # Get the embeddings with torch.no_grad(): embeddings = model(**inputs, output_hidden_states=True, return_dict=True, sent_emb=True).pooler_output return embeddings[0] def title_text_embedding(self, titles, title_to_text): """titles-text-embeddings""" embeddings = [] embed_to_title = [] for title in titles: text = title_to_text[title] # divide text with \n\n divided_texts = text.split('\n\n') for divided_text in divided_texts: embed = self.get_embedding(divided_text) embeddings.append(embed) embed_to_title.append(title) return embeddings, embed_to_title def get_cosine_similarity(self, embed1, embed2): return torch.nn.functional.cosine_similarity(embed1, embed2, dim=0) def retrieve_title(self, query_embed, embeddings, embed_to_title, k): # compute cosine similarity between query_embed and embeddings cosine_similarities = [] for embed in embeddings: cosine_similarities.append(self.get_cosine_similarity(query_embed, embed)) # sort cosine similarity sorted_cosine_similarities = sorted(cosine_similarities, reverse=True) top_k_index = [] top_k_title = [] for i in range(len(sorted_cosine_similarities)): current_title = embed_to_title[cosine_similarities.index(sorted_cosine_similarities[i])] if current_title not in top_k_title: top_k_title.append(current_title) top_k_index.append(cosine_similarities.index(sorted_cosine_similarities[i])) if len(top_k_title) == k: break return top_k_title def organize_story_with_maxlen(self, selected_sample): maxlen = self.max_len_story # title_to_text, _ = self.read_prompt_data() story = "凉宫春日的经典桥段如下:\n" count = 0 final_selected = [] print(selected_sample) for sample_topic in selected_sample: # find sample_answer in dictionary sample_story = self.title_to_text[sample_topic] sample_len = len(enc.encode(sample_story)) # print(sample_topic, ' ' , sample_len) if sample_len + count > maxlen: break story += sample_story story += '\n' count += sample_len final_selected.append(sample_topic) return story, final_selected def organize_message(self, story, history_chat, history_response, new_query): messages = [{'role': 'system', 'content': self.system_prompt}, {'role': 'user', 'content': story}] n = len(history_chat) if n != len(history_response): print('warning, unmatched history_char length, clean and start new chat') # clean all history_chat = [] history_response = [] n = 0 for i in range(n): messages.append({'role': 'user', 'content': history_chat[i]}) messages.append({'role': 'user', 'content': history_response[i]}) messages.append({'role': 'user', 'content': new_query}) return messages def keep_tail(self, history_chat, history_response): max_len = self.max_len_history n = len(history_chat) if n == 0: return [], [] if n != len(history_response): print('warning, unmatched history_char length, clean and start new chat') return [], [] token_len = [] for i in range(n): chat_len = len(enc.encode(history_chat[i])) res_len = len(enc.encode(history_response[i])) token_len.append(chat_len + res_len) keep_k = 1 count = token_len[n - 1] for i in range(1, n): count += token_len[n - 1 - i] if count > max_len: break keep_k += 1 return history_chat[-keep_k:], history_response[-keep_k:] def organize_message_langchain(self, story, history_chat, history_response, new_query): # messages = [{'role':'system', 'content':SYSTEM_PROMPT}, {'role':'user', 'content':story}] messages = [ SystemMessage(content=self.system_prompt), HumanMessage(content=story) ] n = len(history_chat) if n != len(history_response): print('warning, unmatched history_char length, clean and start new chat') # clean all history_chat = [] history_response = [] n = 0 for i in range(n): messages.append(HumanMessage(content=history_chat[i])) messages.append(AIMessage(content=history_response[i])) # messages.append( {'role':'user', 'content':new_query }) messages.append(HumanMessage(content=new_query)) return messages def get_response(self, user_message, chat_history_tuple): history_chat = [] history_response = [] if len(chat_history_tuple) > 0: for cha, res in chat_history_tuple: history_chat.append(cha) history_response.append(res) history_chat, history_response = self.keep_tail(history_chat, history_response) print('history done') new_query = user_message query_embed = self.get_embedding(new_query) # print("1") # embeddings, embed_to_title = self.title_text_embedding(self.titles, self.title_to_text) print("2") selected_sample = self.retrieve_title(query_embed, self.embeddings, self.embed_to_title, 7) print("3") story, selected_sample = self.organize_story_with_maxlen(selected_sample) ## TODO: visualize seletected sample later print('当前辅助sample:', selected_sample) messages = self.organize_message_langchain(story, history_chat, history_response, new_query) chat = ChatOpenAI(temperature=0) return_msg = chat(messages) response = return_msg.content return response def save_response(self, chat_history_tuple): with open(f"{self.save_path}/conversation_{time.time()}.txt", "w") as file: for cha, res in chat_history_tuple: file.write(cha) file.write("\n---\n") file.write(res) file.write("\n---\n") def create_gradio(self): # from google.colab import drive # drive.mount(drive_path) with gr.Blocks() as demo: gr.Markdown( """ ## Chat凉宫春日 ChatHaruhi 项目地址 [https://github.com/LC1332/Chat-Haruhi-Suzumiya](https:// github.com/LC1332/Chat-Haruhi-Suzumiya) 骆驼项目地址 [https://github.com/LC1332/Luotuo-Chinese-LLM](https:// github.com/LC1332/Luotuo-Chinese-LLM) 此版本为图文版本，完整功能（+语音）的demo见项目 角色名建议输入 阿虚 或者影视剧中有的人物。或者也可以是新学生或者老师。 """ ) image_input = gr.Textbox(visible=False) # japanese_input = gr.Textbox(visible=False) with gr.Row(): chatbot = gr.Chatbot() image_output = gr.Image() role_name = gr.Textbox(label="角色名", placeholde="输入角色名") msg = gr.Textbox(label="输入") with gr.Row(): clear = gr.Button("Clear") sub = gr.Button("Submit") image_button = gr.Button("给我一个图") # japanese_output = gr.Textbox(interactive=False) def respond(role_name, user_message, chat_history): input_message = role_name + ':「' + user_message + '」' bot_message = self.get_response(input_message, chat_history) chat_history.append((input_message, bot_message)) self.save_response(chat_history) # time.sleep(1) # jp_text = pipe(f'<-zh2ja-> {bot_message}')[0]['translation_text'] return "" , chat_history, bot_message clear.click(lambda: None, None, chatbot, queue=False) msg.submit(respond, [role_name, msg, chatbot], [msg, chatbot, image_input]) sub.click(fn=respond, inputs=[role_name, msg, chatbot], outputs=[msg, chatbot, image_input]) # with gr.Tab("text_to_text"): # text_input = gr.Textbox() # text_output = gr.Textbox() # text_button = gr.Button('begin') # text_button.click(text.text_to_text, inputs=text_input, outputs=text_output) # with gr.Tab("text_to_iamge"): # with gr.Row(): # image_input = gr.Textbox() # image_output = gr.Image() # image_button = gr.Button("给我一个图") image_button.click(text.text_to_image, inputs=image_input, outputs=image_output) demo.launch(debug=True) if __name__ == '__main__': parser = argparse.ArgumentParser(description="-----[Chat凉宫春日]-----") parser.add_argument("--folder", default="characters/haruhi/texts", help="text folder") parser.add_argument("--system_prompt", default="characters/haruhi/system_prompt.txt", help="store system_prompt") parser.add_argument("--max_len_story", default=1500, type=int) parser.add_argument("--max_len_history", default=1200, type=int) # parser.add_argument("--save_path", default="/content/drive/MyDrive/GPTData/Haruhi-Lulu/") parser.add_argument("--save_path", default=os.getcwd()+"/Suzumiya") options = parser.parse_args() params = { "folder": options.folder, "system_prompt": options.system_prompt, "max_len_story": options.max_len_story, "max_len_history": options.max_len_history, "save_path": options.save_path } # pipe = pipeline(model="engmatic-earth/mt5-zh-ja-en-trimmed-fine-tuned-v1", device=0,max_length=120) run = Run(**params) run.create_gradio() # history_chat = [] # history_response = [] # chat_timer = 5 # new_query = '鲁鲁:你好我是新同学鲁鲁' # query_embed = run.get_embedding(new_query) # titles, title_to_text = run.read_prompt_data() # embeddings, embed_to_title = run.title_text_embedding(titles, title_to_text) # selected_sample = run.retrieve_title(query_embed, embeddings, embed_to_title, 7) # print('限制长度之前:', selected_sample) # story, selected_sample = run.organize_story_with_maxlen(selected_sample) # print('当前辅助sample:', selected_sample) # messages = run.organize_message(story, history_chat, history_response, new_query) # response = run.get_completion_from_messages(messages) # print(response) # history_chat.append(new_query) # history_response.append(response) # history_chat, history_response = run.keep_tail(history_chat, history_response) # print(history_chat, history_response)