chatbot.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
from transformers import pipeline

class ChatBotManager():

    def __init__(self):
        self.available_model_types = ['SmolLM2-1.7B (non-conversational)', \
                                      'CUSsupport-chat-t5-small (conversational)']
        self.model = None
        self.selected_model_type = None

    def get_available_model_types(self):
        return self.available_model_types

    def __create_chatbot__(self, model_type, device='cpu'):
        if (model_type==self.available_model_types[0]):
            return SmolLM2_1_7B_ChatBot(device)
        elif (model_type==self.available_model_types[1]):
            return CUSsupport_chat_t5_small(device)
        elif (model_type==self.available_model_types[2]):
            return Distilgpt2_tiny_conversational(device)

    def obtain_answer(self, message, history, model_type):
        if (self.selected_model_type!=model_type):
            self.selected_model_type = model_type

            self.model = self.__create_chatbot__(self.selected_model_type)

        return self.model.get_answer(message, history)

# Abstract class.
class BaseChatBot():

    def __init__(self, device='cpu'):
        self.device = device

    def __get_input_ids__(self, input_text, device='cpu'):
        input_ids = self.tokenizer(input_text, return_tensors="pt")
        input_ids = input_ids.to(device)

        return input_ids

    def get_answer(self, message, history):
        output_message = None
        if (len(message.strip())>0):
            input_ids = self.__get_input_ids__(message, self.device)

            new_ids_aux = self.model.generate(**input_ids, repetition_penalty=1.5, max_new_tokens=1500)
            new_text_aux = self.tokenizer.decode(new_ids_aux[0])

            output_message = {'role': 'assistant', 'content': new_text_aux}

        return output_message

class SmolLM2_1_7B_ChatBot(BaseChatBot):

    def __init__(self, device='cpu'):
        super().__init__(device)
        model_type = "HuggingFaceTB/SmolLM2-1.7B"
        self.tokenizer = AutoTokenizer.from_pretrained(model_type, cache_dir='./cache')
        self.model = AutoModelForCausalLM.from_pretrained(model_type, cache_dir='./cache')
        self.model.to(device)

class CUSsupport_chat_t5_small(BaseChatBot):

    def __init__(self, device='cpu'):
        super().__init__(device)
        self.t2t_pipeline = pipeline("text2text-generation", \
                                     model="mrSoul7766/CUSsupport-chat-t5-small",
                                     device=torch.device(self.device))

    def get_answer(self, message, history):
        output_message = None
        if (len(message.strip())>0):
            new_text_aux = self.t2t_pipeline(f"answer: {message}", max_length=512)[0]['generated_text']
            output_message = {'role': 'assistant', 'content': new_text_aux}

        return output_message